main.pdf
Abstract
Introduction
Transition-metal complexes are at the heart of homogeneous catalysis, metalloenzymes and molecular magnetism, and much of their chemistry is decided by the near-degeneracy of their d d d orbitals. Partially filled d d d shells, metal–ligand covalency and several low-lying spin states give rise to strong static correlation, which single-reference methods and approximate density functionals describe unreliably, most visibly in spin-state energetics [3 , 4 ] [1, 2]. Multiconfigurational methods such as complete-active-space configuration interaction (CASCI) and self-consistent field (CASSCF) [5 , 6 , 7 ] , combined with a perturbative treatment of dynamic correlation such as second-order N N N -electron valence state perturbation theory (NEVPT2) [8 ] , address this problem, but their accuracy and cost are both fixed by the choice of active space. Exact diagonalization is routinely possible for about 20 orbitals, and the largest nonrelativistic calculations, with about 10 12 10^{12} 1 0 12 determinants, reach 23 to 24 orbitals [9 , 10 ] . High-accuracy approximate solvers such as the density-matrix renormalization group (DMRG) [11 , 12 ] , full configuration interaction quantum Monte Carlo (FCIQMC) [13 , 14 , 15 ] and selected configuration interaction [16 , 17 ] extend the reach to several tens of orbitals at a cost that still grows steeply with the size and the entanglement of the space.
The choice of active space also determines the resources of a quantum computation. Quantum computers have been proposed as solvers for active-space Hamiltonians [18 , 19 , 20 ] , and in the Jordan–Wigner mapping every spatial orbital requires two qubits [21 ] . Near-term algorithms such as sample-based quantum diagonalization (SQD) [22 , 23 ] now treat active spaces of tens of qubits on current processors, so the active space is the quantity that decides whether a problem fits the hardware at all.
SQD is now being applied to chemically relevant problems. It has been used to follow carbene formation from diazirines in active spaces of up to 30 orbitals [24 ] , extended to variable molecular geometries and reaction paths with circuits of up to 124 qubits [25 ] , and applied to open-shell 3 d 3d 3 d complexes: for [ C o ( H 2 O ) 5 ( C O 2 ) ] 2 + / 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{2+/3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 2 + /3 + , one of the application systems of this work, SQD on an IBM Heron processor reproduced coupled-cluster and heat-bath configuration-interaction energies in active spaces of up to 50 qubits [26 ] . The active space also governs the preparation of the initial state. Classifying electronic structures by their orbital entanglement, Mörchen et al. showed that the hardest cases for state preparation are those in which no small set of determinants dominates the wavefunction, whatever orbital basis is used [27 ] , and generative models, refined by reinforcement learning, now construct ground-state preparation circuits for molecular electronic-structure problems [28 ] . In all of these studies the active space is an input, fixed beforehand by classical means (for the cobalt complex by projection onto preselected atomic orbitals), and the quantum result can be no better than that choice.
Active spaces are still most often chosen by chemical intuition [29 ] . Automated schemes select orbitals from projections onto predefined atomic orbitals in the atomic valence active space (AVAS) method [30 ] , from natural-orbital occupations of an approximate correlated calculation [31 ] , by machine-learning classification of candidate active spaces [32 ] , or from the orbital entanglement of an approximate DMRG wavefunction, quantified by single-orbital entropies and mutual information [33 , 34 , 35 , 36 ] . Automated selection has since become an active field in its own right [37 ] : multi-step protocols such as the Active Space Finder construct orbitals and select active spaces from approximate correlated calculations and have been benchmarked for electronic excitation energies [38 ] , and tensor-network approaches locate strongly correlated regions from a matrix-product-state representation of the purities of all orbital partitions, without high-order density matrices [39 ] . Entanglement-based selection is the most direct criterion, because it asks the correlated wavefunction which orbitals are correlated, but it requires a preliminary DMRG calculation over a large orbital space. Projection-based schemes are inexpensive but rely on a predefined list of atomic orbitals and do not distinguish spin states of the same complex.
Machine learning, now widely used to explore chemical space [40 , 41 , 42 ] , offers a way to obtain entanglement information at mean-field cost. Golub et al. [43 ] showed that DMRG single-orbital entropies of transition-metal systems can be learned from integral-based orbital descriptors. Subsequent work has combined such a predictor with reinforcement learning of the selection threshold, rewarded by the agreement of NEVPT2 with DMRG energies [44 ] , and has built graph networks on orbital-correlation features that predict correlation energies and spin-state gaps of iron(II) complexes [45 ] . Agents driven by large language models now choose and refine the active spaces of CASSCF calculations [46 ] and select electron-structure methods from correlation diagnostics [47 ] . Most of these methods, however, have been trained or tested on main-group molecules, small transition-metal species or closed-shell states, and open-shell systems remain underrepresented in the data sets used to develop machine-learned electronic-structure methods [48 ] .
A model that predicts orbital entanglement for open-shell transition-metal complexes has to capture three features. First, the active space depends on the spin state: the high- and low-spin states of the same complex share their nuclear framework but need different active spaces, so the representation must resolve the spin state at the level of the orbitals. Second, orbital entanglement is collective: whether an orbital is important depends on the other orbitals with which it is coupled, so the model should see the frontier orbitals together rather than each one in isolation. Third, single-orbital entropies alone do not say which orbitals are correlated with each other , information that is needed to keep correlated pairs together in the active space. Beyond these features, the usefulness of any such model rests on its transfer to chemistry beyond its training data, which requires a benchmark whose difficulty is defined independently of the model.
Here we introduce SpinAware, a model that predicts DMRG single-orbital entropies and orbital-pair mutual information of transition-metal complexes from the unrestricted Hartree–Fock (UHF) solution of each spin state, and we use it to build active spaces for classical and quantum multireference calculations. The work makes four contributions. (i) A spin-resolved representation of every orbital by 60 descriptors computed from the UHF solution, most of them matrix elements of the electronic Hamiltonian (Sec. II A). (ii) A transformer that predicts the entropy of each orbital in the context of all other frontier orbitals, together with a pairwise network for the mutual information (Sec. II B). (iii) An out-of-distribution benchmark of experimentally characterized complexes whose difficulty is assigned from composition and connectivity alone (Sec. II D). (iv) A complete workflow from the predictions to an active space and on to a quantum processor: a candidate pool from predicted entropies and mutual information, a natural-orbital occupation number (NOON) scan that selects the final active space, a test of the resulting spin-state gaps against coupled-cluster theory, and SQD on IBM quantum hardware (§ 2.0.2 ). SpinAware reproduces the DMRG entropy profiles of 11 spin states with a mean absolute error (MAE) of 0.027, and the candidate pool contains 99% of the DMRG-active orbitals while keeping on average 16 of the 40 frontier orbitals. For a nickel and a cobalt complex, SQD with samples from IBM processors reproduces the exact spin-state gaps of the selected spaces to within 0.1 kcal mol− 1 ^{-1} − 1 for up to 32 qubits.
Methods
SpinAware takes a single UHF calculation for each spin state of a complex as input and returns, for a window of 20 occupied and 20 virtual orbitals around the frontier, the single-orbital entropy of every orbital and the mutual information of every orbital pair. Section II A describes how each orbital is represented by 60 descriptors computed from the UHF solution. Section II B describes the two networks: a transformer that predicts the entropy of each orbital in the context of all other window orbitals (Sec. II B 2), and a pairwise network that predicts the mutual information (Sec. II B 3). Section II C summarizes the training data, the DMRG reference calculations and the training, and Section II D introduces the out-of-distribution benchmark on which the models are evaluated in Section III.
Methods
The single-orbital entropies s i ( 1 ) s_i^{(1)} s i ( 1 ) and the mutual information I i j I_{ij} I ij that decide whether an orbital belongs in the active space require a correlated calculation over a large orbital space, which is precisely the cost that automated active-space selection seeks to avoid. We therefore describe every orbital only through quantities available from a single UHF calculation, and we choose them so that each one is either a matrix element of the electronic Hamiltonian or a direct consequence of one: the one-electron energy of an orbital, the cost of doubly occupying it, its composition in terms of metal and ligand functions, its spin polarization, and its two-electron coupling to the other orbitals. Our representation builds on the integral-based orbital encoding introduced for machine-learned active-space selection by Golub et al. [43 ] and extends it by spin polarization, a continuous measure of metal–ligand mixing, the full pattern of exchange coupling, and the environment of the orbital window together with the molecular charge and spin state. This yields 60 descriptors per orbital, collected in Table I and illustrated in Figure 1 and Figure 2 . After defining the reference wavefunction and the exchange matrix (Sec. II A 1), we introduce the descriptors in six groups, one per subsection (Secs. II A 2 to II A 7).
Figure 1. Spin and composition descriptors. a,b, Spin-occupation difference Δ n α β \Delta n_{\alpha\beta} Δ n α β for the d d d orbitals of a d 6 d^6 d 6 ion in the high-spin (quintet, a a a ) and low-spin (singlet, b b b ) states; the ones mark the four unpaired electrons of the quintet, whereas every occupied d d d orbital of the singlet is paired. c,d, Binary shell flags (c c c ) assign the same code to a metal-centered and a ligand-dominated orbital; continuous shell weights (d d d ) separate them.
Figure 2. Exchange descriptors. a–c, Rows of the exchange matrix and their statistics for three archetypal orbitals: a metal d d d orbital in a correlated manifold (a), coupled with comparable strength to several partners; the bonding orbital of an isolated pair (b), coupled strongly to a single partner; and an inert ligand orbital (c), weakly coupled to all orbitals. The leading couplings of a and b are equal; the mean μ i \mu_i μ i , spread s i s_i s i and partner count n i n_i n i (Eq. (7) ) separate them. Values are schematic, and the partner counts use κ = 0.01 E h \kappa= 0.01 E_{\mathrm{h}} κ = 0.01 E h for illustration. d, Boundary exchange: exchange integrals between a window orbital and the three orbitals immediately below and above the 40-orbital window reveal correlated partners outside the window.
Reference wavefunction and the exchange matrix
For a complex of total charge Q Q Q and spin multiplicity M = 2 S + 1 M = 2S + 1 M = 2 S + 1 , the UHF equations are solved in the def2-SVP basis [49 ] at the target multiplicity. This gives spin orbitals
ϕ i σ ( r ) = ∑ μ c μ i σ χ μ ( r ) , σ ∈ { α , β } , \phi_i^\sigma(\mathbf{r}) = \sum_\mu c_{\mu i}^\sigma\chi_\mu(\mathbf{r}), \quad\sigma\in\{\alpha, \beta\}, ϕ i σ ( r ) = μ ∑ c μ i σ χ μ ( r ) , σ ∈ { α , β } ,
with energies ε i σ \varepsilon_i^\sigma ε i σ and occupations n i σ ∈ { 0 , 1 } n_i^\sigma\in\{0,1\} n i σ ∈ { 0 , 1 } . In transition-metal complexes, the static correlation that the active space has to describe originates almost entirely in the frontier region, that is, among the orbitals whose energies lie close to the gap between the highest occupied molecular orbital (HOMO) and the lowest unoccupied molecular orbital (LUMO). We therefore retain a window W \mathcal{W} W of N = 40 N = 40 N = 40 orbitals: the twenty occupied α \alpha α orbitals highest in energy, HOMO−19 to HOMO, and the twenty virtual α \alpha α orbitals lowest in energy, LUMO to LUMO+19. Dynamic correlation is spread over all orbitals and is added afterwards by perturbation theory (Sec. II C). All orbitals below the window are kept doubly occupied as a frozen core, and those above it are discarded from the DMRG calculations that provide the reference entropies. Every descriptor refers to an orbital i ∈ W i \in\mathcal{W} i ∈ W .
Beyond the mean-field level, orbitals are coupled to one another through the two-electron integrals of the Hamiltonian. Of those, we use the exchange integral, which appears in several descriptors below,
K i j = ∬ ϕ i ( r 1 ) ϕ j ( r 1 ) 1 ∣ r 1 − r 2 ∣ ϕ j ( r 2 ) ϕ i ( r 2 ) d r 1 d r 2 ≥ 0 , (1) K_{ij} = \iint\phi_i(\mathbf{r}_1)\phi_j(\mathbf{r}_1)\frac{1}{\lvert\mathbf{r}_1-\mathbf{r}_2\rvert}\phi_j(\mathbf{r}_2)\phi_i(\mathbf{r}_2)\,\mathrm{d}r_1\,\mathrm{d}r_2 \ge0, \tag*{(1)} K ij = ∬ ϕ i ( r 1 ) ϕ j ( r 1 ) ∣ r 1 − r 2 ∣ 1 ϕ j ( r 2 ) ϕ i ( r 2 ) d r 1 d r 2 ≥ 0 , (1)
which we evaluate for the α \alpha α and β \beta β coefficient sets and average, setting K i i = 0 K_{ii}=0 K ii = 0 . K i j K_{ij} K ij is the self-repulsion of the overlap density ϕ i ϕ j \phi_i\phi_j ϕ i ϕ j . It is therefore large only for orbitals that occupy the same region of space, and it vanishes for orbitals localized on distant fragments, however close their energies may be. K i j K_{ij} K ij is also the Hamiltonian matrix element that connects a reference determinant to the determinant in which an electron pair is transferred from i i i to j j j . In a two-orbital model, the weight of the doubly excited configuration, and hence the entanglement between i i i and j j j , grows with the ratio of K i j K_{ij} K ij to the energy separation of the two orbitals. K i j K_{ij} K ij thus indicates which orbitals can become correlated, and the energy gaps decide which of them are. The full N × N N \times N N × N matrix K \mathbf{K} K is used both to build descriptors and, in the model, as a structural prior on orbital–orbital interactions (Section II B).
One-electron and on-site descriptors
The first group of descriptors describes each orbital in isolation through the diagonal elements of the Hamiltonian. The one-electron integral
h i i = ∫ ϕ i ( r ) [ − 1 2 ∇ 2 − ∑ A Z A ∣ r − R A ∣ ] ϕ i ( r ) d r (2) h_{ii} = \int\phi_i(\mathbf{r}) \left[ -\frac{1}{2}\nabla^2 - \sum_A \frac{Z_A}{|\mathbf{r}-\mathbf{R}_A|} \right] \phi_i(\mathbf{r}) \,d\mathbf{r} \tag*{(2)} h ii = ∫ ϕ i ( r ) [ − 2 1 ∇ 2 − A ∑ ∣ r − R A ∣ Z A ] ϕ i ( r ) d r (2)
is strongly negative for compact orbitals held close to a highly charged nucleus, such as metal 3 d 3d 3 d functions, and less negative for diffuse ligand orbitals. The self-Coulomb integral
J i i = ∬ ∣ ϕ i ( r 1 ) ∣ 2 ∣ ϕ i ( r 2 ) ∣ 2 ∣ r 1 − r 2 ∣ d r 1 d r 2 (3) J_{ii} = \iint\frac{|\phi_i(\mathbf{r}_1)|^2 |\phi_i(\mathbf{r}_2)|^2}{|\mathbf{r}_1-\mathbf{r}_2|} \,d\mathbf{r}_1 \,d\mathbf{r}_2 \tag*{(3)} J ii = ∬ ∣ r 1 − r 2 ∣ ∣ ϕ i ( r 1 ) ∣ 2 ∣ ϕ i ( r 2 ) ∣ 2 d r 1 d r 2 (3)
is the repulsion between two electrons placed in the same orbital and plays the role of an on-site Hubbard repulsion: when it is large compared with the energy needed to move an electron to a nearby orbital, double occupancy is penalized, as in the partially filled d d d shells of transition-metal ions. The spatial extent ⟨ r 2 ⟩ i = ∑ μ ∣ c μ i ∣ 2 ⟨ χ μ ∣ r 2 ∣ χ μ ⟩ \langle r^2\rangle_i = \sum_\mu|c_{\mu i}|^2 \langle\chi_\mu|r^2|\chi_\mu\rangle ⟨ r 2 ⟩ i = ∑ μ ∣ c μ i ∣ 2 ⟨ χ μ ∣ r 2 ∣ χ μ ⟩ gives a complementary one-electron measure of the same compactness.
The canonical orbital energy ε i \varepsilon_i ε i (of the α \alpha α set) places the orbital on the energy axis. Three further labels, taken from Golub et al. [43 ] , describe the frontier position independently of molecule size. The first is the signed distance from the HOMO, k i = i − i H O M O k_i = i - i_{\mathrm{HOMO}} k i = i − i HOMO , rescaled as p i = [ 1 + exp ( − k i / 100 ) ] − 1 p_i = [1+\exp(-k_i/100)]^{-1} p i = [ 1 + exp ( − k i /100 ) ] − 1 . The second is the occupation code o i o_i o i , which takes the values 0.25, 0.50 and 0.75 for virtual, singly and doubly occupied orbitals. The third is a bonding label b i = ± 1 b_i = \pm1 b i = ± 1 , which records whether ε i \varepsilon_i ε i lies below or above the mean energy of the window.
Spin polarization
The active space of a transition-metal complex depends on its spin state. In a high-spin octahedral d 6 d^6 d 6 ion, four electrons are unpaired and all five d d d orbitals are partially occupied. In the low-spin state the t 2 g t_{2g} t 2 g set is filled, and correlation is dominated instead by the t 2 g → e g t_{2g} \to e_g t 2 g → e g near-degeneracy and by metal–ligand covalency. A d d d orbital has nearly the same shape, composition and energy whether it holds a spin pair or a single electron, so the one-electron and on-site descriptors of Section II A 2 change only modestly between the two states. Yet the distinction matters for correlation: a doubly occupied orbital has both spin channels filled and is comparatively inert, whereas a singly occupied orbital has an open spin channel into which electrons can fluctuate. A spin-blind representation would describe both states with the same input for the same nuclear framework, and no model built on it could predict that their active spaces differ. We therefore encode the spin state at the level of each orbital, through two complementary quantities (Figure 1 a,Figure 1 b).
The first is the spin-occupation difference
Δ n α β ( i ) = n i α − n i β , (4) \Delta n_{\alpha\beta}(i) = n_i^\alpha- n_i^\beta, \tag*{(4)} Δ n α β ( i ) = n i α − n i β , (4)
which takes only two values. It is 1 for an orbital that holds one unpaired α \alpha α electron, and 0 for an orbital that is either doubly occupied (n i α = n i β = 1 n_i^\alpha= n_i^\beta= 1 n i α = n i β = 1 ) or empty. For a high-spin d 6 d^6 d 6 ion it marks the four unpaired electrons, whereas in the low-spin state it is 0 throughout (Figure 1 a,Figure 1 b). For the same nuclear framework, where the shape-based descriptors barely differ, this descriptor separates the two states.
The second quantity is the spin splitting Δ ε α β ( i ) ≡ ε i α − ε i β \Delta\varepsilon_{\alpha\beta}(i) \equiv\varepsilon_i^\alpha- \varepsilon_i^\beta Δ ε α β ( i ) ≡ ε i α − ε i β between the α \alpha α and β \beta β orbital energies. In Hartree–Fock theory, and to the extent that the α \alpha α and β \beta β orbitals share the same spatial form, the Coulomb terms cancel in this difference and
Δ ε α β ( i ) ≃ − ∑ u ∈ S O M O K i u . (5) \Delta\varepsilon_{\alpha\beta}(i) \simeq- \sum_{u \in\mathrm{SOMO}} K_{iu}. \tag*{(5)} Δ ε α β ( i ) ≃ − u ∈ SOMO ∑ K i u . (5)
The splitting of an orbital is thus the total exchange it feels with the unpaired electrons: it vanishes in a closed-shell state, is largest for orbitals that share space with the singly occupied metal orbitals, and is small for ligand orbitals remote from the spin density.
The two descriptors are complementary: Δ n α β \Delta n_{\alpha\beta} Δ n α β is a binary flag, whereas Δ ε α β \Delta\varepsilon_{\alpha\beta} Δ ε α β is a continuous measure of how strongly exchange acts on each orbital. A molecular spin label ( M − 1 ) / 4 (M-1)/4 ( M − 1 ) /4 , shared by all orbitals, completes the group. Because Δ ε α β \Delta\varepsilon_{\alpha\beta} Δ ε α β and Δ n α β \Delta n_{\alpha\beta} Δ n α β are given for every orbital of the window, the model learns both that a complex is high-spin and how the α \alpha α and β \beta β orbital energies differ along the window.
In transition-metal complexes, static correlation is concentrated in orbitals with substantial metal d d d character, and the degree of metal–ligand mixing (covalency) controls how strongly a d d d -derived orbital is correlated. A simple way to encode atomic-orbital character is a set of binary flags, one for each of the twelve shell types t ∈ { 1 s , … , 5 s , 2 p , … , 5 p , 3 d , 4 d , 5 d } t \in\{1s,\ldots,5s,2p,\ldots,5p,3d,4d,5d\} t ∈ { 1 s , … , 5 s , 2 p , … , 5 p , 3 d , 4 d , 5 d } . Each flag is set to one if any basis function of that type contributes to the orbital, and we retain these twelve flags [43 ] . Binary flags, however, cannot express how much of an orbital belongs to a given shell: a nearly pure metal 3 d 3d 3 d orbital and a ligand p p p orbital with a small 3 d 3d 3 d admixture receive the same flags (Figure 1 c,Figure 1 d). Since canonical orbitals are delocalized, almost every orbital carries some small coefficient on almost every shell type, and the flags tend to be set for all orbitals alike.
We resolve this with continuous shell weights derived from normalized squared AO coefficients. The weight of shell type t t t in orbital i i i is
w i , t = ∑ μ 1 2 ( ∣ c μ i α ∣ 2 + ∣ c μ i β ∣ 2 ) ∑ μ 1 2 ( ∣ c μ i α ∣ 2 + ∣ c μ i β ∣ 2 ) , ∑ t w i , t = 1. (6) w_{i,t} = \frac{\sum_{\mu} \frac{1}{2}\left(\lvert c_{\mu i}^{\alpha}\rvert^{2}+\lvert c_{\mu i}^{\beta}\rvert^{2}\right)}{\sum_{\mu} \frac{1}{2}\left(\lvert c_{\mu i}^{\alpha}\rvert^{2}+\lvert c_{\mu i}^{\beta}\rvert^{2}\right)}, \qquad\sum_t w_{i,t}=1. \tag*{(6)} w i , t = ∑ μ 2 1 ( ∣ c μ i α ∣ 2 + ∣ c μ i β ∣ 2 ) ∑ μ 2 1 ( ∣ c μ i α ∣ 2 + ∣ c μ i β ∣ 2 ) , t ∑ w i , t = 1. (6)
The 3 d 3d 3 d weight w i , 3 d w_{i,3d} w i , 3 d measured against the ligand p p p weights is a direct, continuous measure of covalency.
Exchange connectivity of the orbital network
Static correlation is collective: an orbital becomes strongly entangled because it is nearly degenerate with, and strongly coupled to, other orbitals, not because of its own properties alone. The row { K i j } j ≠ i \{K_{ij}\}_{j\ne i} { K ij } j = i of the exchange matrix defines a weighted network over the window, and we describe the position of each orbital in this network at three levels of detail. The ten largest couplings, sorted in descending order [43 ] , identify the strongest partners.
These values alone, however, cannot tell apart physically different situations, as three archetypes illustrate (Figure 2 a–Figure 2 c). A metal d d d orbital in a partially filled shell (A) is coupled with comparable strength to the other d d d orbitals of the same center and to little else. A bonding orbital of an isolated two-center bond (B) is coupled strongly to a single partner, its antibonding counterpart, and only weakly to everything else. An orbital of a peripheral ligand (C) is weakly coupled to all orbitals. Yet A and B can have the same leading coupling, so the top of a sorted list of couplings cannot separate them.
We therefore summarize the whole exchange row by four statistics,
μ i = 1 N ∑ j K i j , s i = [ 1 N ∑ j ( K i j − μ i ) 2 ] 1 / 2 , \mu_i = \frac{1}{N}\sum_j K_{ij}, \qquad s_i = \left[\frac{1}{N}\sum_j\left(K_{ij}-\mu_i\right)^2\right]^{1/2}, μ i = N 1 j ∑ K ij , s i = [ N 1 j ∑ ( K ij − μ i ) 2 ] 1/2 ,
S i = ∑ j K i j , n i = ∑ j Θ ( K i j − κ ) , (7) S_i = \sum_j K_{ij}, \qquad n_i = \sum_j \Theta\left(K_{ij}-\kappa\right), \tag*{(7)} S i = j ∑ K ij , n i = j ∑ Θ ( K ij − κ ) , (7)
the mean, standard deviation and sum of the couplings, and the number of partners whose coupling exceeds a threshold κ \kappa κ (Θ \Theta Θ is the Heaviside step function). The sums run over all N = 40 N=40 N = 40 window orbitals with K i i = 0 K_{ii}=0 K ii = 0 , and κ = 0.05 E h \kappa=0.05 E_{\mathrm{h}} κ = 0.05 E h . S i S_i S i measures the total exchange connectivity of the orbital, and n i n_i n i the number of strong partners. We also characterize the shape of the distribution. With the normalized couplings K ~ i j = K i j / S i \widetilde{K}_{ij}=K_{ij}/S_i K ij = K ij / S i we define
C i = max j K ~ i j , P R i = ( ∑ j K ~ i j 2 ) − 1 . (8) C_i = \max_j \widetilde{K}_{ij}, \qquad\mathrm{PR}_i = \left(\sum_j \widetilde{K}_{ij}^{2}\right)^{-1}. \tag*{(8)} C i = j max K ij , PR i = ( j ∑ K ij 2 ) − 1 . (8)
The participation ratio P R i \mathrm{PR}_i PR i , familiar as a measure of the delocalization of vibrational and electronic states [50 ] , is the effective number of exchange partners of orbital i i i . It is close to one for a pair-like orbital and grows towards N − 1 N-1 N − 1 when the coupling is spread evenly. The concentration C i C_i C i is the fraction of the coupling carried by the dominant partner.
In an idealized model in which orbital A couples with strength K 0 K_0 K 0 to m m m partners, orbital B with the same strength to a single partner, and orbital C with a small strength δ ≪ K 0 \delta\ll K_0 δ ≪ K 0 to all N − 1 N-1 N − 1 orbitals (Supplementary Section S3), the leading coupling is K 0 K_0 K 0 for both A and B, whereas S i S_i S i , n i n_i n i and P R i \mathrm{PR}_i PR i are m m m times larger for A (m = 4 m=4 m = 4 for a d d d orbital coupled to the other four d d d orbitals of its center). The spectator C is separated from both by its small max j K i j \max_j K_{ij} max j K ij and S i S_i S i and by n i = 0 n_i=0 n i = 0 , although its participation ratio, N − 1 N-1 N − 1 , is the largest of the three. The shape descriptors C i C_i C i and P R i \mathrm{PR}_i PR i are therefore meaningful only together with the magnitudes S i S_i S i and max j K i j \max_j K_{ij} max j K ij , which is why both kinds of statistic are supplied to the model.
Energetic relation to the frontier and to exchange partners
Exchange coupling indicates which orbitals can interact, but whether the interaction produces strong correlation depends on their energy separation, which we describe by two energy differences. The molecular HOMO–LUMO gap Δ H L = ε L U M O − ε H O M O \Delta_{\mathrm{HL}} = \varepsilon_{\mathrm{LUMO}} - \varepsilon_{\mathrm{HOMO}} Δ HL = ε LUMO − ε HOMO is a global indicator of near-degeneracy, as small gaps generally accompany larger correlated spaces. The second descriptor links the energetic and two-electron pictures for each orbital,
g i = ∣ ε i − ε j ∗ ∣ , j ∗ = arg max j K i j , (9) g_i = \left|\varepsilon_i-\varepsilon_{j^*}\right|,\qquad j^* = \underset{j}{\operatorname{arg\,max}}\,K_{ij}, \tag*{(9)} g i = ∣ ε i − ε j ∗ ∣ , j ∗ = j arg max K ij , (9)
which is the energy separation between an orbital and its strongest exchange partner. A large K i j ∗ K_{ij^*} K i j ∗ together with a large g i g_i g i marks the bonding and antibonding partners of a chemical bond, such as the σ / σ ∗ \sigma/\sigma^* σ / σ ∗ pair of a metal–ligand bond or the d / π ∗ d/\pi^* d / π ∗ pair of back-bonding, whereas a small g i g_i g i indicates mixing within a single quasi-degenerate manifold, such as the d d d shell itself. The pair ( K i j ∗ , g i ) (K_{ij^*}, g_i) ( K i j ∗ , g i ) therefore separates these two mechanisms of static correlation.
Coupling beyond the window and global descriptors
Any finite orbital window may cut through a correlated subspace, a well-known difficulty of active-space selection [29 , 36 ] : an orbital just inside the window can be strongly coupled to one just outside it, and the resulting active space is then incomplete. A representation confined to the window implicitly assumes that all coupled partners lie inside it. This holds for compact complexes with simple monodentate ligands, but it can fail for systems with delocalized correlation, such as macrocyclic [51 ] , salen-type [52 ] or multinuclear complexes [53 , 54 , 55 , 56 , 57 ] , whose active spaces extend over the ligand framework or over several metal centers. For Fe(II) porphyrin, for example, including the full π \pi π system in the active space changes the order of the spin states [51 ] . Although the orbitals outside the window take no part in the correlated calculation, their exchange integrals with the window orbitals are fully defined (Figure 2 d). We make the edges of the window visible by evaluating Eq. (1) between each window orbital and the three orbitals immediately below (B − B_{-} B − ) and above (B + B_{+} B + ) the window. For each side we record max k ∈ B ± K i k \max_{k\in B_{\pm}} K_{ik} max k ∈ B ± K ik and ∑ k ∈ B ± K i k \sum_{k\in B_{\pm}} K_{ik} ∑ k ∈ B ± K ik , together with the number of couplings above 0.05 E h 0.05\,E_{\mathrm{h}} 0.05 E h . When these quantities are small, orbital i i i has its partners inside W \mathcal{W} W . When they are large, the window is missing a coupled partner, and the active space suggested for that orbital should be interpreted with caution or the window enlarged.
Finally, the multiplicity M M M and the total charge Q Q Q are appended to every orbital, so that the orbital-level descriptors are read in the context of the global electron count and spin state. All reference calculations are carried out for the isolated ion, at the charge implied by the metal oxidation state and the formal ligand charges. Every descriptor is obtained from a single mean-field calculation, at a small fraction of the cost of the correlated calculation it replaces, and none depends on a predefined list of elements or ligands.
Model architecture
Physical motivation
The single-orbital entropy s i ( 1 ) s_i^{(1)} s i ( 1 ) quantifies how much orbital i i i is entangled with the rest of the electronic system. It is defined through the von Neumann entropy of the one-orbital reduced density matrix ρ i \rho_i ρ i , obtained by tracing out all degrees of freedom except orbital i i i from the DMRG wavefunction:
s i ( 1 ) = − Tr [ ρ i ln ρ i ] , (10) s_i^{(1)}=-\operatorname{Tr}[\rho_i\ln\rho_i], \tag*{(10)} s i ( 1 ) = − Tr [ ρ i ln ρ i ] , (10)
with the bounds s i ( 1 ) ∈ [ 0 , ln 4 ] s_i^{(1)} \in[0,\ln4] s i ( 1 ) ∈ [ 0 , ln 4 ] . An orbital with s i ( 1 ) ≈ 0 s_i^{(1)} \approx0 s i ( 1 ) ≈ 0 is effectively doubly occupied or empty and can safely be excluded from the active space, while an orbital with s i ( 1 ) ≳ 0.1 s_i^{(1)} \gtrsim0.1 s i ( 1 ) ≳ 0.1 carries significant entanglement and should be retained.
A fundamental property of s i ( 1 ) s_i^{(1)} s i ( 1 ) is that it is not an intrinsic property of orbital i i i alone. Rather, it measures how orbital i i i is entangled with every other orbital in the system. Consider a concrete example from the benchmark of this work. In the singlet of [ V C l 6 ] 3 − [\mathrm{VCl}_6]^{3-} [ VCl 6 ] 3 − (1), one window orbital has a DMRG entropy of s ( 1 ) = 0.099 s^{(1)} = 0.099 s ( 1 ) = 0.099 , just below the threshold τ = 0.10 \tau= 0.10 τ = 0.10 used to define the active space, and would be discarded on its own. Its mutual information with a single partner orbital (s ( 1 ) = 0.142 s^{(1)} = 0.142 s ( 1 ) = 0.142 ) is, however, 0.028, fourteen times larger than its mutual information with any other orbital (at most 0.002). The two orbitals form a correlated pair, and excluding the first while retaining its partner breaks this pair. This example illustrates a general principle: the importance of an orbital depends on the context provided by every other orbital in the molecule.
A model that maps the descriptors of each orbital to an entropy independently of the other orbitals cannot represent such pairs: the features of the partner never enter the prediction, and two orbitals that matter only together cannot be retained or discarded together. We therefore process all N N N orbitals of the window jointly, so that the predicted entropy of each orbital depends on the descriptors of every other orbital, and we guide this exchange of information by the exchange integrals K i j K_{ij} K ij (Sec. II A 1).
Model 1: Single-orbital entropy prediction
Problem statement. Given the descriptor vectors x i ∈ R 60 \mathbf{x}_i \in\mathbb{R}^{60} x i ∈ R 60 (Section II A) of the N = 40 N = 40 N = 40 window orbitals of a UHF calculation, Model 1 predicts the DMRG single-orbital entropies s i ( 1 ) s_i^{(1)} s i ( 1 ) (Eq. (10) ) of all N N N orbitals simultaneously.
Architecture overview. The model comprises three stages (Figure 3 ): (i) an embedding layer that projects each orbital’s 60 descriptors into a high-dimensional representation space; (ii) a stack of L L L transformer layers with self-attention, through which the orbitals exchange information; and (iii) a prediction head that maps each orbital’s context-enriched representation to a scalar entropy estimate.
Figure 3. Architecture of the orbital transformer for single-orbital entropy prediction. a, Input from one UHF/def2-SVP calculation on the triplet of [ V C l 6 ] 3 − [\mathrm{VCl}_6]^{3-} [ VCl 6 ] 3 − (1): structure, schematic α \alpha α and β \beta β orbital levels (V(III), d 2 d^2 d 2 ) with HOMO, LUMO and the window of 20 occupied and 20 virtual orbitals, descriptor matrix X \mathbf{X} X (40 orbitals × \times × 60 descriptors, including M M M and Q Q Q ) and exchange matrix K \mathbf{K} K . b, Network overview: the same weights act on every window orbital, through the embedding (60 → \to → 256), four transformer layers in which the orbitals exchange information by self-attention (purple lines) and the prediction head, which returns s ^ i ( l ) \hat{s}_i^{(l)} s ^ i ( l ) . c, Orbital embedding with layer normalization. d, Self-attention: after layer normalization each orbital is projected to query, key and value vectors (8 heads × \times × 32 dimensions); the scaled dot product is complemented by the normalized exchange integrals with a learnable weight α ( l ) \alpha^{(l)} α ( l ) per layer, shared by the heads and initialized at zero. e, Feed-forward network (256 → \to → 1024 → \to → 256, GELU); both sub-layers are pre-norm with residual connections. f, Prediction head shared by all 40 orbitals (256 → \to → 64 → \to → 1), applied after a final layer normalization and followed by the inverse Box–Cox transformation. g, Illustrative entropy profile with the selection thresholds τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
Stage 1: Orbital embedding. Each orbital’s descriptor vector x i ∈ R 60 \mathbf{x}_i \in\mathbb{R}^{60} x i ∈ R 60 is projected into a d d d -dimensional representation space:
h i ( 0 ) = LayerNorm ( W e m b x i + b e m b ) , W e m b ∈ R d × 60 (11) \mathbf{h}_i^{(0)} = \operatorname{LayerNorm}(\mathbf{W}_{\mathrm{emb}}\mathbf{x}_i+\mathbf{b}_{\mathrm{emb}}), \qquad\mathbf{W}_{\mathrm{emb}} \in\mathbb{R}^{d \times60} \tag*{(11)} h i ( 0 ) = LayerNorm ( W emb x i + b emb ) , W emb ∈ R d × 60 (11)
with d = 256 d = 256 d = 256 and layer normalization [58 ] . At this stage, the representations { h i ( 0 ) } i = 1 N \{\mathbf{h}_i^{(0)}\}_{i=1}^{N} { h i ( 0 ) } i = 1 N are still independent: no information has flowed between orbitals.
Stage 2: Self-attention. The core of the architecture is a stack of L = 4 L = 4 L = 4 transformer layers. Each layer l l l updates the orbital representations through two sub-steps: multi-head self-attention with an exchange-integral bias, followed by a position-wise feedforward network (FFN).
Self-attention with exchange-integral bias. We use multi-head scaled dot-product attention [59 ] with n h e a d s = 8 n_{\mathrm{heads}} = 8 n heads = 8 heads of dimension d h = d / n h e a d s = 32 d_h = d/n_{\mathrm{heads}} = 32 d h = d / n heads = 32 . In each head, the query, key and value vectors q i \mathbf{q}_i q i , k j \mathbf{k}_j k j and v j \mathbf{v}_j v j are linear projections of the layer-normalized orbital representations, and the attention logit between orbitals i i i and j j j is
a i j = q i ⋅ k j d h + α ( l ) K ~ i j , (12) a_{ij} = \frac{\mathbf{q}_i \cdot\mathbf{k}_j}{\sqrt{d_h}} + \alpha^{(l)}\widetilde{K}_{ij}, \tag*{(12)} a ij = d h q i ⋅ k j + α ( l ) K ij , (12)
where K ~ i j = K i j / K s \widetilde{K}_{ij} = K_{ij}/K_s K ij = K ij / K s is the exchange integral between orbitals i i i and j j j divided by a fixed, dataset-wide scale K s K_s K s . The scalar α ( l ) \alpha^{(l)} α ( l ) is a learnable parameter of layer l l l , shared by all heads and initialized to zero, so that training starts from standard attention. The attention weights w i j w_{ij} w ij are the softmax of a i j a_{ij} a ij over all orbitals j j j , and the outputs ∑ j w i j v j \sum_j w_{ij}\mathbf{v}_j ∑ j w ij v j of the heads are concatenated and linearly projected.
The second term supplies the exchange-coupling matrix as an additional input to the attention logits, and since α ( l ) \alpha^{(l)} α ( l ) is unconstrained, it can increase or decrease the relative logit of a pair. The exchange integral supplies pair-resolved information beyond the query–key dot product, but it is not itself an entanglement measure.
Attention and the exchange bias. Through attention, each orbital receives information from all other window orbitals already in the first layer, and the four layers update these representations nonlinearly. Each orbital thus also receives information from its correlated partners (Sec. II B 1), which a model that scores each orbital independently [43 ] cannot do. The pairwise network of Model 2 adds an explicit prediction of their mutual information. Retraining Model 1 without the exchange bias, or with the exchange matrices of each molecule randomly permuted over its orbitals, gives the same accuracy (Supplementary Table S23), so the exchange information reaches the model through the exchange descriptors of each orbital (Table 1 ).
Table 1. Orbital descriptors. Indices refer to the input vector of each window orbital; descriptors marked † are molecular quantities shared by all orbitals of a complex. The order follows the input vector of the trained model.
Feedforward network and residual connections. After the attention sub-layer, each orbital’s representation passes through the FFN:
FFN ( z ) = W 2 GELU ( W 1 z + b 1 ) + b 2 , (13) \operatorname{FFN}(z) = \mathbf{W}_2\operatorname{GELU}(\mathbf{W}_1 z+\mathbf{b}_1)+\mathbf{b}_2, \tag*{(13)} FFN ( z ) = W 2 GELU ( W 1 z + b 1 ) + b 2 , (13)
with W 1 ∈ R d f f × d \mathbf{W}_1 \in\mathbb{R}^{d_{\mathrm{ff}} \times d} W 1 ∈ R d ff × d and W 2 ∈ R d × d f f \mathbf{W}_2 \in\mathbb{R}^{d \times d_{\mathrm{ff}}} W 2 ∈ R d × d ff , where d f f = 4 d = 1024 d_{\mathrm{ff}} = 4d = 1024 d ff = 4 d = 1024 , GELU is the Gaussian error linear unit, and dropout with rate 0.1 is applied to the attention weights and inside the FFN. Each transformer layer applies pre-normalization residual connections [60 ] , where M H A i \mathrm{MHA}_i MHA i denotes the multi-head attention output for orbital i i i :
h i ′ = h i ( l − 1 ) + M H A i ( LayerNorm ( h i ( l − 1 ) ) ) , (14) \mathbf{h}_i' = \mathbf{h}_i^{(l-1)} + \mathrm{MHA}_i(\operatorname{LayerNorm}(\mathbf{h}_i^{(l-1)})), \tag*{(14)} h i ′ = h i ( l − 1 ) + MHA i ( LayerNorm ( h i ( l − 1 ) )) , (14)
h i ( l ) = h i ′ + FFN ( LayerNorm ( h i ′ ) ) . (15) \mathbf{h}_i^{(l)} = \mathbf{h}_i' + \operatorname{FFN}(\operatorname{LayerNorm}(\mathbf{h}_i')). \tag*{(15)} h i ( l ) = h i ′ + FFN ( LayerNorm ( h i ′ )) . (15)
Stage 3: Prediction head. After L = 4 L = 4 L = 4 attention layers, the final representations { h i ( L ) } i = 1 N \{\mathbf{h}_i^{(L)}\}_{i=1}^{N} { h i ( L ) } i = 1 N are passed through a shared prediction head:
y ^ i = ReLU ( w o u t T ReLU ( W h i d LayerNorm ( h i ( L ) ) + b h i d ) + b o u t ) , (16) \hat{y}_i = \operatorname{ReLU}\left(\mathbf{w}_{\mathrm{out}}^{\mathsf{T}} \operatorname{ReLU}\left(\mathbf{W}_{\mathrm{hid}} \operatorname{LayerNorm}\left(\mathbf{h}_i^{(L)}\right) + \mathbf{b}_{\mathrm{hid}}\right) + b_{\mathrm{out}}\right), \tag*{(16)} y ^ i = ReLU ( w out T ReLU ( W hid LayerNorm ( h i ( L ) ) + b hid ) + b out ) , (16)
where W h i d ∈ R 64 × 256 \mathbf{W}_{\mathrm{hid}} \in\mathbb{R}^{64 \times256} W hid ∈ R 64 × 256 and w o u t ∈ R 64 \mathbf{w}_{\mathrm{out}} \in\mathbb{R}^{64} w out ∈ R 64 , and ReLU is the rectified linear unit. The final ReLU activation enforces the physical constraint y ^ i ≥ 0 \hat{y}_i \geq0 y ^ i ≥ 0 , since s i ( 1 ) ≥ 0 s_i^{(1)} \geq0 s i ( 1 ) ≥ 0 by definition. During training the outer ReLU is replaced by the identity, and the output is clamped at zero in evaluation and inference, which gives the same function as Eq. (16) while keeping a non-zero gradient for every orbital during training.
The prediction y ^ i \hat{y}_i y ^ i is in a transformed target space. The true entropy is recovered via inverse Box–Cox transformation, as described below.
Target preprocessing. The distribution of s i ( 1 ) s_i^{(1)} s i ( 1 ) values across the training set is heavily right-skewed: the majority of orbitals have near-zero entropy, while a small fraction carry s i ( 1 ) > 0.5 s_i^{(1)} > 0.5 s i ( 1 ) > 0.5 . Direct regression on raw s i ( 1 ) s_i^{(1)} s i ( 1 ) would be dominated by the trivial near-zero predictions, with the model learning to predict “approximately zero” for all orbitals.
We apply a Box–Cox transformation [61 ] with parameter λ \lambda λ followed by min–max scaling to [ 0 , 1 ] [0, 1] [ 0 , 1 ] :
y i = MinMaxScale ( ( s i ( 1 ) + ϵ ) λ − 1 λ ) , (17) y_i = \operatorname{MinMaxScale}\left(\frac{\left(s_i^{(1)} + \epsilon\right)^{\lambda} - 1}{\lambda}\right), \tag*{(17)} y i = MinMaxScale λ ( s i ( 1 ) + ϵ ) λ − 1 , (17)
with λ = 0.1377934 \lambda= 0.1377934 λ = 0.1377934 , the same for all five models of the final ensemble, and ϵ = 10 − 8 \epsilon= 10^{-8} ϵ = 1 0 − 8 . The transformation is fitted on the training complexes and held fixed for validation, testing and benchmark predictions. The Box–Cox transform amplifies distinctions among small s i ( 1 ) s_i^{(1)} s i ( 1 ) values, where the boundary between important and unimportant orbitals lies, while compressing differences among large values that are less decision-relevant.
At inference, predicted values are mapped back to the physical entropy scale:
s ^ i ( 1 ) = max ( 0 , ( λ ⋅ y ^ i ′ + 1 ) 1 / λ − ϵ ) , (18) \hat{s}_i^{(1)} = \max\left(0, \left(\lambda\cdot\hat{y}_i' + 1\right)^{1/\lambda} - \epsilon\right), \tag*{(18)} s ^ i ( 1 ) = max ( 0 , ( λ ⋅ y ^ i ′ + 1 ) 1/ λ − ϵ ) , (18)
where y ^ i ′ \hat{y}_i' y ^ i ′ is the inverse min–max-scaled prediction.
Training objective. The model is trained by minimizing the MAE on the transformed targets over batches of molecules:
L s ( 1 ) = 1 ∣ B ∣ ∑ m ∈ B 1 N m ∑ i = 1 N m ∣ y ^ i ( m ) − y i ( m ) ∣ , (19) \mathcal{L}_{s^{(1)}} = \frac{1}{|\mathcal{B}|} \sum_{m \in\mathcal{B}} \frac{1}{N_m} \sum_{i=1}^{N_m} \left| \hat{y}_i^{(m)} - y_i^{(m)} \right|, \tag*{(19)} L s ( 1 ) = ∣ B ∣ 1 m ∈ B ∑ N m 1 i = 1 ∑ N m y ^ i ( m ) − y i ( m ) , (19)
where B \mathcal{B} B is a training batch and N m N_m N m is the number of orbitals in the window for molecule m m m . Compared with a squared error, the absolute error is less dominated by the few orbitals with very large entropy. The model contains approximately 3.19 × 10 6 3.19 \times10^{6} 3.19 × 1 0 6 trainable parameters (3 189 125 3\,189\,125 3 189 125 ). Training details are given in Sec. II C.
Motivation: beyond single-orbital entropies. Mutual information has previously been predicted by machine learning for strongly correlated transition-metal systems by Golub et al. [62 ] . Here we predict it from spin-resolved UHF descriptors and use it to complete the entropy-based selection.
The single-orbital entropy identifies orbitals that are correlated with something . However, it does not reveal which orbitals are correlated with each other . This pairwise information is carried by the mutual information
I i j = 1 2 ( s i ( 1 ) + s j ( 1 ) − s i j ( 2 ) ) , (20) I_{ij} = \frac{1}{2}\left(s_i^{(1)} + s_j^{(1)} - s_{ij}^{(2)}\right), \tag*{(20)} I ij = 2 1 ( s i ( 1 ) + s j ( 1 ) − s ij ( 2 ) ) , (20)
where s i j ( 2 ) = − Tr [ ρ i j ln ρ i j ] s_{ij}^{(2)} = -\operatorname{Tr}[\rho_{ij}\ln\rho_{ij}] s ij ( 2 ) = − Tr [ ρ ij ln ρ ij ] is the two-orbital von Neumann entropy of the two-orbital reduced density matrix ρ i j \rho_{ij} ρ ij . By subadditivity I i j ≥ 0 I_{ij} \geq0 I ij ≥ 0 , with I i j = 0 I_{ij} = 0 I ij = 0 if and only if orbitals i i i and j j j are uncorrelated, and I i j ≤ ln 4 I_{ij} \leq\ln4 I ij ≤ ln 4 . This convention, with the prefactor 1 2 \frac{1}{2} 2 1 , is used throughout (for all labels, predictions and thresholds).
The mutual information provides two capabilities that single-orbital entropy alone cannot: it identifies low-entropy partners of strongly coupled orbitals, which an entropy threshold would discard although they belong to a correlated pair, and it identifies blocks of mutually correlated orbitals whose individual entropies lie close to the threshold. Both are used to build the candidate pool of § 2.0.2 B.
Pairwise architecture. The mutual-information model is a separate network that scores every pair ( i , j ) (i,j) ( i , j ) , i < j i < j i < j , of the window. Its input combines the descriptors of the two orbitals with the exchange structure around them. Let x ~ i ∈ R 60 \tilde{\mathbf{x}}_i \in\mathbb{R}^{60} x ~ i ∈ R 60 be the standardized descriptor vector of orbital i i i and k ~ i ∈ R 40 \tilde{\mathbf{k}}_i \in\mathbb{R}^{40} k ~ i ∈ R 40 its standardized row of the exchange matrix, which records how strongly the orbital couples to every orbital of the window. To make the prediction independent of the order of the two orbitals, the pair is ordered so that a a a denotes the orbital with the larger descriptor norm and b b b the other. The pair representation is
z i j = [ x ~ a ∥ x ~ b ∥ ∣ x ~ a − x ~ b ∣ ∥ x ~ a ⊙ x ~ b ∥ k ~ a ∥ k ~ b ∥ ∣ k ~ a − k ~ b ∣ ∥ k ~ a ⊙ k ~ b ∥ ln ∣ K i j ∣ ] ∈ R 401 , (21) \mathbf{z}_{ij} = [\tilde{\mathbf{x}}_a \Vert\tilde{\mathbf{x}}_b \Vert\lvert\tilde{\mathbf{x}}_a - \tilde{\mathbf{x}}_b \rvert\Vert\tilde{\mathbf{x}}_a \odot\tilde{\mathbf{x}}_b \Vert
\tilde{\mathbf{k}}_a \Vert\tilde{\mathbf{k}}_b \Vert\lvert\tilde{\mathbf{k}}_a - \tilde{\mathbf{k}}_b \rvert\Vert\tilde{\mathbf{k}}_a \odot\tilde{\mathbf{k}}_b \Vert\ln\lvert K_{ij} \rvert] \in\mathbb{R}^{401}, \tag*{(21)} z ij = [ x ~ a ∥ x ~ b ∥ ∣ x ~ a − x ~ b ∣ ∥ x ~ a ⊙ x ~ b ∥ k ~ a ∥ k ~ b ∥ ∣ k ~ a − k ~ b ∣ ∥ k ~ a ⊙ k ~ b ∥ ln ∣ K ij ∣] ∈ R 401 , (21)
where ∥ \Vert ∥ denotes concatenation and ⊙ \odot ⊙ the element-wise product. The absolute difference and the product encode how similar the two orbitals are and along which dimensions both are large, so that the network does not have to rediscover these interactions. The exchange rows tell the network with which other orbitals each partner is coupled, and the direct exchange integral K i j K_{ij} K ij , entered on a logarithmic scale, supplies the two-electron coupling of the pair itself. On the training data, ∣ K i j ∣ \lvert K_{ij}\rvert ∣ K ij ∣ alone correlates with ln I i j \ln I_{ij} ln I ij with a coefficient of 0.56, which is why it is entered directly.
The vector z i j \mathbf{z}_{ij} z ij is projected to a hidden dimension of 384 (linear layer, layer normalization, GELU) and processed by three gated residual blocks,
u ← u + σ ( W g u ) ⊙ f ( u ) , (22) \mathbf{u} \leftarrow\mathbf{u}+\sigma(\mathbf{W}_{g}\mathbf{u})\odot f(\mathbf{u}), \tag*{(22)} u ← u + σ ( W g u ) ⊙ f ( u ) , (22)
in which f f f consists of two linear layers with layer normalization, GELU activation and dropout (rate 0.15), and the sigmoid gate σ \sigma σ lets each block decide how much it modifies the representation. A two-layer head maps the result to a scalar y ^ i j \hat{y}_{ij} y ^ ij , which predicts ln ( I i j + ϵ ) \ln(I_{ij}+\epsilon) ln ( I ij + ϵ ) with ϵ = 10 − 7 \epsilon=10^{-7} ϵ = 1 0 − 7 , and the mutual information is recovered as I ^ i j = max ( 0 , e y ^ i j − ϵ ) \hat{I}_{ij}=\max(0,e^{\hat{y}_{ij}}-\epsilon) I ^ ij = max ( 0 , e y ^ ij − ϵ ) . The network has 1.57 × 10 6 1.57\times10^{6} 1.57 × 1 0 6 trainable parameters (1 565 185).
Training objective. The distribution of I i j I_{ij} I ij is even more skewed than that of s ( 1 ) s^{(1)} s ( 1 ) : most pairs have I i j ≈ 0 I_{ij}\approx0 I ij ≈ 0 , and appreciable mutual information is concentrated in a small number of physically important pairs. We therefore regress the logarithm of I i j I_{ij} I ij and weight each pair by its mutual information,
L I = 1 ∣ B ∣ ∑ ( i j ) ∈ B w i j ℓ δ ( y ^ i j , ln ( I i j + ϵ ) ) , w i j ∝ 1 + 100 I i j , (23) \mathcal{L}_{I}=\frac{1}{\lvert\mathcal{B}\rvert}\sum_{(ij)\in\mathcal{B}}w_{ij}\ell_{\delta}\left(\hat{y}_{ij},\ln(I_{ij}+\epsilon)\right),\qquad w_{ij}\propto1+100I_{ij}, \tag*{(23)} L I = ∣ B ∣ 1 ( ij ) ∈ B ∑ w ij ℓ δ ( y ^ ij , ln ( I ij + ϵ ) ) , w ij ∝ 1 + 100 I ij , (23)
where ℓ δ \ell_{\delta} ℓ δ is the Huber loss [63 ] with δ = 2 \delta=2 δ = 2 and the weights are normalized to unit mean. The logarithmic target resolves the small values that decide whether a weakly correlated orbital is rescued, and the weights keep the rare strongly coupled pairs from being dominated by the many uncorrelated ones. The labels I i j I_{ij} I ij are obtained from the same DMRG calculations as the entropy labels, from the one- and two-orbital reduced density matrices.
Unlike Model 1, the pairwise network sees the rest of the molecule only through the exchange rows k ~ a \tilde{\mathbf{k}}_{a} k ~ a and k ~ b \tilde{\mathbf{k}}_{b} k ~ b . Sharing the transformer backbone between the two tasks, so that the pair representation is built from context-enriched orbital representations h i ( L ) \mathbf{h}_{i}^{(L)} h i ( L ) , is a natural extension that we leave to future work.
Inference
At inference time, the trained models are applied to molecules for which no DMRG calculation has been performed. A UHF single-point calculation, which takes minutes for the systems considered here, provides the orbital energies, occupations and expansion coefficients and the integrals h i i h_{ii} h ii , J i i J_{ii} J ii and K i j K_{ij} K ij . From it, the 60 descriptors of each of the N = 40 N=40 N = 40 window orbitals (Sec. II A) and the exchange matrix K \mathbf{K} K are computed in seconds. After the preprocessing described below, Model 1 returns the N N N entropies, mapped back to the physical scale by Eq. (18) , and Model 2 the mutual information of all 780 orbital pairs.
Preprocessing. For Model 1, the stored min–max transformation is applied to input columns 0–3, 30, 31, 44–49, 58 and 59 of Table I. All other columns retain their supplied representation. The exchange matrix is divided by the stored scale K s = 2.6162941 E h K_{\mathrm{s}}=2.6162941\,E_{\mathrm{h}} K s = 2.6162941 E h . These same fitted transformations are used without modification at inference.
From predictions to an active space. The predictions are not used with a single entropy threshold. A fixed threshold works when the entropy distribution is bimodal, but for many transition-metal states the profile is flat near the threshold, and the outcome of a hard cut is then decided by small, chemically insignificant deviations (§ 2.0.2 A). We therefore proceed in two steps. The predicted entropies and mutual information are first combined into a compact candidate pool designed for completeness (§ 2.0.2 B). The final active space is then selected from the pool by a NOON scan that relies only on the correlated wavefunction itself (§ 2.0.2 D).
Data and computational details
Training data. The models must learn how the orbital structure of a single UHF reference relates to orbital entanglement across the chemistry of first-row transition-metal complexes. The training set was therefore built to cover the main factors that control static correlation: the metal and its d d d -electron count, the spin state, the ligand field and the metal–ligand bond length. It comprises 3 454 spin-pure calculations on mononuclear complexes of Cr, Mn, Fe, Co and Ni (Supplementary Table S3): 2 105 constructed complexes at equilibrium geometries, 807 with metal–ligand bonds stretched by factors of 1.5 and 2.0 to sample increasingly strong static correlation, and 542 complexes taken from experimental crystal structures. Spin multiplicities range from 1 to 6 and total charges from − 9 -9 − 9 to + 3 +3 + 3 , all but nine of them between − 5 -5 − 5 and + 3 +3 + 3 . None of the benchmark complexes of § 2.0.2 D is part of the training data.
For every complex and spin state, the UHF reference and the 60 descriptors of Sec. II A were computed in the def2-SVP basis [49 ] , and the reference labels (single-orbital entropies and mutual information) were obtained from DMRG calculations [11 , 12 ] with BLOCK2 [64 ] in the same 40-orbital window, starting from the UHF occupation pattern and increasing the bond dimension over 16 sweeps from 250 and 500 to a final M = 1024 M=1024 M = 1024 with decreasing noise, with S z S_z S z symmetry and the window orbitals in their canonical (orbital-energy) order. The single-orbital entropy of orbital i i i is the entropy of DMRG site i i i with its four local states, and I i j I_{ij} I ij refers to pairs of such sites. All calculations use M S = S M_S=S M S = S (M S = 0 M_S=0 M S = 0 for singlets), the spin projection of the UHF occupation. Because the entropies depend on M S M_S M S , the labels refer to this component.
States whose UHF reference deviates from the exact ⟨ S ^ 2 ⟩ \langle\hat{S}^{2}\rangle ⟨ S ^ 2 ⟩ ... by more than 0.5, a looser filter than the criterion applied to the benchmark (§ 2.0.2 D), and DMRG calculations that had not converged were removed, so that every label describes the intended spin state. Because S z S_z S z -adapted DMRG fixes only the spin projection, the total spin of every DMRG wavefunction was verified by evaluating the exact ⟨ S ^ 2 ⟩ \langle\hat{S}^{2}\rangle ⟨ S ^ 2 ⟩ , including the overlaps between the α \alpha α and β \beta β orbitals of the UHF basis and the contribution of the occupied orbitals below the window. For states in which the S z S_z S z -adapted DMRG approached a different total spin, an energy penalty on S ^ 2 − S ( S + 1 ) \hat{S}^{2} - S(S + 1) S ^ 2 − S ( S + 1 ) was added to the Hamiltonian, which drives the DMRG to the intended spin state. In all cases the descriptors and the labels refer to the same UHF reference. The spin contamination w = [ ⟨ S ^ 2 ⟩ − S ( S + 1 ) ] / ( 2 S + 2 ) w = [\langle\hat{S}^{2}\rangle- S(S + 1)]/(2S + 2) w = [⟨ S ^ 2 ⟩ − S ( S + 1 )] / ( 2 S + 2 ) is below 0.10 for every label (below 0.05 for the penalty calculations), and calculations without a converged spin-pure state were not included. The resulting training set comprises 3 454 spin-pure states, 787 of them obtained with the spin penalty.
Spin-adapted solvers, such as SU(2) DMRG [65 , 66 ] , which is also available in block2, or FCIQMC in the graphical unitary group approach [15 , 57 , 67 , 68 ] , require the same spatial orbitals for both spins. The S z S_z S z -adapted formulation with a spin penalty was used instead, because the descriptors and labels are defined in the unrestricted orbitals of the UHF reference. Because the α \alpha α and β \beta β window orbitals span different spaces, an exact spin eigenstate generally cannot be represented in the window, so ⟨ S ^ 2 ⟩ \langle\hat{S}^{2}\rangle ⟨ S ^ 2 ⟩ can only approach S ( S + 1 ) S(S + 1) S ( S + 1 ) . Over all benchmark labels, the largest remaining deviation is w = 0.08 w = 0.08 w = 0.08 (at most 0.05 for the states of Table III ).
Table III. Prediction quality on the out-of-distribution benchmark. n D M R G n_{\mathrm{DMRG}} n DMRG and n M L n_{\mathrm{ML}} n ML are the numbers of orbitals with s 1 ≥ τ s_1 \ge\tau s 1 ≥ τ in the reference and predicted profiles; MAE is the orbital-wise mean absolute error of s 1 s_1 s 1 over the 40-orbital window. Predictions: mean of the five Model 1 runs; references: spin-checked DMRG labels (Sec. II D).
Training and validation. To test generalization to unseen molecules rather than to unseen orbitals of known molecules, the data were split by complex: all spin states, charges, stretched geometries and repeated crystal-structure determinations of a complex were placed together in the training, validation or test set. The 3 454 states belong to 1 630 distinct complexes, which were divided into 1 103 complexes for training, 125 for validation and 402 for testing, which contain 2 329, 259 and 866 states, respectively. No complex occurs in more than one set. Model 1 was trained on the Box–Cox-transformed entropies with a mean-absolute-error loss, and the checkpoint with the lowest validation error was retained. Five models were trained from different initial weights on the same split, and their mean prediction is used throughout. On the held-out test complexes the ensemble reaches an MAE of 0.025 in s ( 1 ) s^{(1)} s ( 1 ) (0.080 on the orbitals with s ( 1 ) ≥ 0.10 s^{(1)} \ge0.10 s ( 1 ) ≥ 0.10 ) and classifies orbitals as active (s ( 1 ) ≥ 0.10 s^{(1)} \ge0.10 s ( 1 ) ≥ 0.10 ) with a precision of 0.86, a recall of 0.80 and F 1 = 0.83 F_{1} = 0.83 F 1 = 0.83 , pooled over all test orbitals (single models: F 1 = 0.818 ± 0.002 F_{1} = 0.818 \pm0.002 F 1 = 0.818 ± 0.002 ).
For comparison with the model of Golub et al. [43 ] , whose code and trained models are not available, we retrained its architecture (five hidden layers of 896 units, scoring each orbital separately) with its 29 orbital features on the same data, split and initial seeds. This model reaches F 1 = 0.69 F_{1} = 0.69 F 1 = 0.69 on the test complexes (MAE 0.036) and 0.47 on the benchmark of § 2.0.2 D (mean over states, 0.54 pooled). With the 60 descriptors of this work the same network reaches 0.80 and 0.62 (0.72), and SpinAware 0.83 and 0.65 (0.77). Most of the gain therefore comes from the spin-resolved descriptors, and processing the orbitals jointly adds a smaller further improvement (Supplementary Table S25).
Model 2 was trained on the logarithm of the mutual information, on the same training, validation and test complexes as Model 1, using the 3 240 states for which mutual-information labels are available (2 174, 241 and 825 states). Five models were trained from different initial weights. On the test complexes the predicted and DMRG values of ln I i j \ln I_{ij} ln I ij correlate with a coefficient of 0.865 ± \pm ± 0.001, and pairs with I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 are identified with an area under the receiver operating characteristic curve of 0.918 (mean and standard deviation over the five models). All scalers and transformation parameters were fitted on the training data only. Architectures, optimizers and all other hyperparameters of both models are listed in Supplementary Table S4 .
Table S4: Architectures and training settings of the two models. Both models use the same split by complex; five models per task were trained from different initial weights, and their mean prediction is used. Test metrics refer to the held-out test complexes.
Table S4.
Correlated reference calculations. To judge the predicted active spaces by the quantity of chemical interest, the spin-state gap, we computed the energies of the selected spaces with CASCI, in the fixed orbitals of the lowest spin-clean UHF reference (§ 2.0.2 D). Spaces of up to 16 orbitals were solved exactly and larger spaces by DMRG-CASCI, using spin-adapted [SU(2)] DMRG in the UHF natural orbitals (UNOs), so that each state has the intended total spin by construction. Dynamical correlation outside the active space was added with strongly contracted NEVPT2 [8 ] , using reduced density matrices from DMRG where needed. Full-molecule calculations with unrestricted coupled cluster with singles, doubles and perturbative triples [UCCSD(T)] [69 ] and frozen core, together with the T 1 T_{1} T 1 diagnostic [70 ] , serve as an independent reference for the spin-state gaps. All calculations used the def2-SVP basis.
Software. Mean-field, CASCI, NEVPT2 and coupled-cluster calculations were performed with PySCF [71 , 72 ] , DMRG calculations with BLOCK2 [64 ] , and the models were implemented in PyTorch [73 ] and trained with AdamW [74 ] and cosine learning-rate annealing [75 ] . Data processing and figures used NumPy [76 ] , SciPy [77 ] , scikit-learn [78 ] and Matplotlib [79 ] . The quantum calculations used Qiskit [80 ] , ffsim [81 ] and qiskit-addon-sqd [82 ] on IBM quantum processors (§ 2.0.2 E). Crystal structures were taken from the Cambridge Structural Database [83 ] . Software versions are listed in Supplementary Table S5.
Out-of-distribution benchmark
Design principle. A model for active-space selection is useful only if it transfers beyond the chemistry that generated its training labels. We therefore assessed transferability on a benchmark of experimentally characterized complexes whose difficulty was defined from chemical composition and connectivity alone. The assignment uses neither the model nor the reference wavefunctions, so the difficulty scale is independent of the quantity being evaluated. Descriptors of electronic complexity obtained from the reference calculations (active-space size, maximal single-orbital entropy and the entropy gap at the selection threshold) were used only to interpret errors, not to assign difficulty. Throughout, “out of distribution” refers to chemistry, that is, to metals, ligands, binding modes and oxidation states absent from the training set, and not to the descriptor space, in which the benchmark orbitals lie close to the training data (§ 2.0.2 A).
Reference chemical space. The training set (Sec. II C) comprises 3 454 spin-pure calculations on 1 895 distinct mononuclear structures of Cr, Mn, Fe, Co and Ni. The 2 912 constructed complexes carry two to six monodentate ligands on a metal in oxidation state +2 or +3, taken from a fixed library of F− ^{-} − , Cl− ^{-} − , Br− ^{-} − , OH− ^{-} − , S2 − ^{2-} 2 − , H2 _2 2 O, NH3 _3 3 , NCS− ^{-} − , N3 − _3^{-} 3 − and CO. Their geometries are based on the idealized structures of the SC1MC data set [84 ] , which were optimized at the Hartree–Fock level, and on complexes derived from these templates by ligand exchange at standard metal–ligand distances. All spin states of a complex share one geometry, at which both the UHF input and the DMRG label of each state are computed. The 542 complexes taken from experimental crystal structures add organic ligands and other binding modes. No benchmark complex, or any structure sharing its crystallographic refcode, is part of the training set. The benchmark instead introduces motifs that the training geometries, checked through their metal–ligand and ligand–ligand connectivity, do not contain or contain only in isolated cases: the metals Ti and V, metal–methyl bonds, η 3 \eta^3 η 3 -allyl and η 6 \eta^6 η 6 -arene ligands, a morpholine donor, a halide bridging to aluminum, and zero-valent chromium carrying four or five carbonyl ligands (5 and 6). Chelating ligands, by contrast, are common in training and are not counted as novel. Because all training elements lie at or below Kr and are described all-electron by the def2-SVP basis, while heavier elements would require an effective core potential that changes the orbital descriptors, the benchmark was restricted to complexes of elements up to Kr (Supplementary Section S1).
A test complex can depart from this space along four independent directions. A direction counts as a departure when the corresponding motif is absent from the training geometries or occurs there only in isolated cases:
M metal identity : the metal is absent from training (here Ti and V);
L ligand chemistry : at least one ligand type is absent or rare;
T binding topology : at least one binding mode is absent or rare; chelation as such is common and does not count;
O formal oxidation state : the oxidation state lies outside { + 2 , + 3 } \{+2,+3\} { + 2 , + 3 } of the constructed complexes and its characteristic coordination (here zero-valent chromium with four or five carbonyls) is absent.
We define the novelty index N O O D ∈ { 0 , … , 4 } N_{\mathrm{OOD}} \in\{0,\ldots,4\} N OOD ∈ { 0 , … , 4 } as the number of these directions along which a complex departs from the training distribution.
Difficulty tiers. Complexes with N O O D ≤ 1 N_{\mathrm{OOD}} \le1 N OOD ≤ 1 were classed as near-distribution (Tier I), those with N O O D = 2 N_{\mathrm{OOD}} = 2 N OOD = 2 as intermediate (Tier II), and those with N O O D ≥ 3 N_{\mathrm{OOD}} \ge3 N OOD ≥ 3 as far-from-distribution (Tier III). The seven complexes discussed in the main text are listed in Table II and shown in Figure 4 . They span five metals (Ti, V, Cr, Co and Ni), d 2 d^2 d 2 to d 8 d^8 d 8 electron counts, formal oxidation states from 0 to +3, three to six coordination sites and 7 to 26 atoms. The two vanadium(III) ions 1 and 2 probe the response of the active space to ligand field strength and low-symmetry distortion at a fixed metal. Five of the seven crystal structures are salts. In every case the calculation was performed on the isolated metal-containing species, and counter-ions and lattice solvent were omitted. The source crystal structures and the omitted counter-ions are listed in Supplementary Table S1 .
Figure 4. Structures of the complexes studied in the main text. 1–7: out-of-distribution benchmark complexes, taken from experimental crystal structures (counter-ions and lattice solvent omitted). 8: application system [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]3 + ^{3+} 3 + at the equilibrium geometry of the published Co–O dissociation series (Co–O 1.97 Å), with CO2 _2 2 bound end-on through one oxygen atom. The tier of each complex (Table II ) is given below its formula. The complexes of Supplementary Section S4 are shown in Supplementary Fig. S1.
Table II. Out-of-distribution benchmark complexes: composition, novelty and coordination geometry. Novelty directions: M, metal; L, ligand; T, binding topology; O, oxidation state. Multiplicities are given as 2 S + 1 2S+1 2 S + 1 ; those marked * have UHF references outside the spin-contamination criterion and are reported separately. Sites count an η n \eta^n η n ligand as one site at its centroid (ctr). τ 4 \tau_4 τ 4 is the four-coordinate geometry index; Σ \Sigma Σ is the sum of deviations of the twelve cis angles from 90°; bite, chelate bite angle.
Table S1. Source crystal structures of the benchmark complexes. Compound names are given as deposited in the CSD. Calculations used only the metal-containing species; the counter-ions listed here were omitted.
Structural features. Coordination geometries were characterized from the experimental coordinates (Table II ), taken from the Cambridge Structural Database [83 ] . An η n \eta^n η n ligand is counted as one coordination site located at the centroid of its bound carbon atoms. Four-coordinate geometries are described by τ 4 = [ 360 ∘ − ( α + β ) ] / 141 ∘ \tau_4 = [360^\circ- (\alpha+ \beta)]/141^\circ τ 4 = [ 36 0 ∘ − ( α + β )] /14 1 ∘ , where α \alpha α and β \beta β are the two largest L–M–L angles (τ 4 = 1 \tau_4 = 1 τ 4 = 1 for a tetrahedron and 0 for a square plane) [85 ] , and octahedral distortion by Σ = ∑ i = 1 12 ∣ 90 ∘ − θ i ∣ \Sigma= \sum_{i=1}^{12} |90^\circ- \theta_i| Σ = ∑ i = 1 12 ∣9 0 ∘ − θ i ∣ over the twelve cis angles. The structures of the individual complexes are described in Supplementary Section S1.
Spin states and reference wavefunctions. For each complex we considered the spin multiplicities accessible for its d d d -electron count (Table II ). UHF references were obtained from multiple initial guesses followed by internal stability analysis, and the lowest-energy solution satisfying ∣ ⟨ S ^ 2 ⟩ − S ( S + 1 ) ∣ ≤ max [ 0.1 , 0.1 S ( S + 1 ) ] |\langle\hat{S}^2\rangle-S(S+1)| \leq\max[0.1, 0.1\,S(S+1)] ∣ ⟨ S ^ 2 ⟩ − S ( S + 1 ) ∣ ≤ max [ 0.1 , 0.1 S ( S + 1 )] was retained. States for which no such solution was found are marked in Table II and excluded from aggregate statistics. Screening for the stable solution is essential: in a control calculation on [ C o ( H 2 O ) 5 ( C O 2 ) ] 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 3 + , an unscreened quintet reference lay 149.8 mE h E_{\mathrm h} E h above the stable solution and changed the DMRG active space from three to ten orbitals. For [ C o ( C H 3 ) 4 ] − [\mathrm{Co}(\mathrm{CH}_3)_4]^{-} [ Co ( CH 3 ) 4 ] − the lowest computed state is the triplet, whose UHF reference does not meet this criterion. The retained singlet and quintet are therefore not the lowest computed states and serve as test states. For [ C r ( C O ) 5 ( H N C 4 H 8 O ) ] [\mathrm{Cr}(\mathrm{CO})_5(\mathrm{HNC}_4\mathrm{H}_8\mathrm{O})] [ Cr ( CO ) 5 ( HNC 4 H 8 O )] the triplet likewise fails the criterion. Its singlet, the expected ground state of an 18-electron Cr(0) carbonyl, lies 50 kcal mol− 1 ^{-1} − 1 below the triplet at the UHF level and 72 kcal mol− 1 ^{-1} − 1 below it in the window DMRG (each state in its own orbitals) and is included in the benchmark.
Reference labels. Single-orbital entropies s 1 s_1 s 1 and mutual information I i j I_{ij} I ij were computed with DMRG as for the training labels (Sec. II C). Orbitals with s 1 ≥ 0.10 s_1 \geq0.10 s 1 ≥ 0.10 were classed as active. The total spin of every benchmark label was verified in the same way (Sec. II C), and the closed-shell singlets were computed with the S ^ 2 \hat{S}^2 S ^ 2 penalty (⟨ S ^ 2 ⟩ ≈ 10 − 5 \langle\hat{S}^2\rangle\approx10^{-5} ⟨ S ^ 2 ⟩ ≈ 1 0 − 5 ). Because the window position and the orbital ordering change with spin multiplicity, active orbitals were compared between spin states through the molecular-orbital overlap ∣ C 1 T S C 2 ∣ |\mathbf{C}_1^\mathsf{T}\mathbf{S}\mathbf{C}_2| ∣ C 1 T S C 2 ∣ rather than by orbital index.
Results and discussion
Transferability to out-of-distribution complexes
We evaluated SpinAware on six benchmark complexes that differ from the training data in metal, ligand, binding topology or oxidation state: the vanadium(III) ions [ V C l 6 ] 3 − [\mathrm{VCl}_6]^{3-} [ VCl 6 ] 3 − (1) and [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − (2), the tetramethylcobaltate [ C o ( C H 3 ) 4 ] − [\mathrm{Co}(\mathrm{CH}_3)_4]^{-} [ Co ( CH 3 ) 4 ] − (3), the chromium(0) carbonyls [ C r ( C O ) 5 ( H N C 4 H 8 O ) ] [\mathrm{Cr}(\mathrm{CO})_5(\mathrm{HNC}_4\mathrm{H}_8\mathrm{O})] [ Cr ( CO ) 5 ( HNC 4 H 8 O )] (5) and [ C r ( C O ) 4 ( H 2 N C H 2 C O 2 ) ] − [\mathrm{Cr}(\mathrm{CO})_4(\mathrm{H}_2\mathrm{NCH}_2\mathrm{CO}_2)]^{-} [ Cr ( CO ) 4 ( H 2 NCH 2 CO 2 ) ] − (6), and the titanium, arene complex [ T i ( η 6 - C 6 H 6 ) ( A l C l 4 ) 2 ] [\mathrm{Ti}(\eta^6\text{-}\mathrm{C}_6\mathrm{H}_6)(\mathrm{AlCl}_4)_2] [ Ti ( η 6 - C 6 H 6 ) ( AlCl 4 ) 2 ] (7). The nickel allyl complex [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) and [ C o ( H 2 O ) 5 ( C O 2 ) ] 2 + / 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{2+/3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 2 + /3 + (8) are analyzed separately in § 2.0.2 C–III E, where the predicted active spaces are carried through to multireference and quantum-hardware calculations. We analyze the 11 spin states of these six complexes for which a spin-clean UHF reference exists, because the descriptors, the DMRG labels and any subsequent multireference calculation all rest on this reference. The remaining states are listed in Supplementary Section S4, together with the septet of [ C r ( C O ) 5 ( H N C 4 H 8 O ) ] [\mathrm{Cr}(\mathrm{CO})_5(\mathrm{HNC}_4\mathrm{H}_8\mathrm{O})] [ Cr ( CO ) 5 ( HNC 4 H 8 O )] , which requires an electron outside the d d d shell and is kept there as a high-spin test case (Table II ). For every state, the predicted single-orbital entropy profile is compared with the DMRG reference over the full 40-orbital window.
An orbital is classed as active when its entropy reaches a threshold τ \tau τ . Let A D M R G \mathcal{A}_{\mathrm{DMRG}} A DMRG and A M L \mathcal{A}_{\mathrm{ML}} A ML be the active sets obtained from the reference and predicted entropies. Treating A D M R G \mathcal{A}_{\mathrm{DMRG}} A DMRG as the ground truth, precision P = ∣ A M L ∩ A D M R G ∣ / ∣ A M L ∣ P = |\mathcal{A}_{\mathrm{ML}} \cap\mathcal{A}_{\mathrm{DMRG}}|/|\mathcal{A}_{\mathrm{ML}}| P = ∣ A ML ∩ A DMRG ∣/∣ A ML ∣ is the fraction of predicted active orbitals that are correct, and recall R = ∣ A M L ∩ A D M R G ∣ / ∣ A D M R G ∣ R = |\mathcal{A}_{\mathrm{ML}} \cap\mathcal{A}_{\mathrm{DMRG}}|/|\mathcal{A}_{\mathrm{DMRG}}| R = ∣ A ML ∩ A DMRG ∣/∣ A DMRG ∣ is the fraction of true active orbitals that are recovered. Their harmonic mean,
F 1 = 2 P R P + R = 2 T P 2 T P + F P + F N (24) F_1 = \frac{2PR}{P+R} = \frac{2\mathrm{TP}}{2\mathrm{TP}+\mathrm{FP}+\mathrm{FN}} \tag*{(24)} F 1 = P + R 2 P R = 2 TP + FP + FN 2 TP (24)
where TP, FP and FN count true-positive, false-positive and false-negative orbitals, rewards active spaces that are both complete and compact: an active space that includes every orbital (R = 1 R=1 R = 1 , low P P P ) or only one correct orbital (P = 1 P=1 P = 1 , low R R R ) both score poorly. When both sets are empty, F 1 = 1 F_1=1 F 1 = 1 by convention. Alongside F 1 F_1 F 1 we report the orbital-wise MAE of s 1 s_1 s 1 . The two metrics are complementary: the MAE measures how faithfully the entropy profile is reproduced, whereas F 1 F_1 F 1 measures whether the resulting active-space decision is correct.
A two-configuration model illustrates the chemical scale of the thresholds. Consider a doubly occupied orbital that loses a fraction w w w of its electron pair to a correlating partner, the elementary event of static correlation. Its single-orbital entropy is then s 1 = − ( 1 − w ) ln ( 1 − w ) − w ln w s_1=-(1-w)\ln(1-w)-w\ln w s 1 = − ( 1 − w ) ln ( 1 − w ) − w ln w , and its natural occupation is n = 2 ( 1 − w ) n=2(1-w) n = 2 ( 1 − w ) . The threshold τ = 0.10 \tau=0.10 τ = 0.10 , the value used in earlier machine-learned active-space selection [43 ] , corresponds to w ≈ 0.02 w\approx0.02 w ≈ 0.02 , that is, to an orbital that carries about 2% of a doubly excited configuration and has a natural occupation of about 1.96 (or 0.04 for the receiving virtual orbital). The permissive threshold τ = 0.07 \tau=0.07 τ = 0.07 corresponds to w ≈ 0.013 w\approx0.013 w ≈ 0.013 and occupations of about 1.97 and 0.03. It therefore lies close to the conventional natural-occupation criterion for an active orbital, occupations outside 0.02–1.98 [31 , 86 ] , which we also use in the NOON scans of § 2.0.2 D. These thresholds provide practical screening criteria, whose consequences for the selected spaces are assessed by the NOON scans and energy comparisons.
Averaged over the 11 spin states, SpinAware reproduces the DMRG entropy profiles with an MAE of 0.027 and recovers the active space with F 1 = 0.65 F_1 = 0.65 F 1 = 0.65 at $\tau = 0.10 ({xref}Table III`). Pooled over all orbitals of the 11 states, as for the test set, F 1 = 0.77 F_1 = 0.77 F 1 = 0.77 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.81 at τ = 0.07 \tau= 0.07 τ = 0.07 , and the MAE is 0.077 over the 83 DMRG-active orbitals and 0.015 over the remaining 357 orbitals. We see no sign of degradation with the novelty index: the mean F 1 F_1 F 1 is 0.53 for Tier I, 0.83 for Tier II and 0.75 for Tier III, although Tiers II and III contain only two complexes and one complex. The chromium glycinate [ C r ( C O ) 4 ( H 2 N C H 2 C O 2 ) ] − [\mathrm{Cr}(\mathrm{CO})_4(\mathrm{H}_2\mathrm{NCH}_2\mathrm{CO}_2)]^{-} [ Cr ( CO ) 4 ( H 2 NCH 2 CO 2 ) ] − , which departs from the training set in ligand and oxidation state, and the titanium arene [ T i ( η 6 - C 6 H 6 ) ( A l C l 4 ) 2 ] [\mathrm{Ti}(\eta^6\text{-}\mathrm{C}_6\mathrm{H}_6)(\mathrm{AlCl}_4)_2] [ Ti ( η 6 - C 6 H 6 ) ( AlCl 4 ) 2 ] , which departs in metal, ligand and binding topology, are among the best-predicted systems (F 1 = 0.67 F_1 = 0.67 F 1 = 0.67 –0.84 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.71–0.98 at τ = 0.07 \tau= 0.07 τ = 0.07 ), whereas the most difficult cases are the Tier I vanadium and cobalt ions. The tier-averaged MAE is likewise flat (0.028, 0.026 and 0.025 for Tiers I–III). Chemical novelty as such is therefore not what limits transfer. The model’s input is a set of orbital-resolved electronic quantities (energies, exchange couplings, metal–ligand composition and spin polarization) rather than an identity of the ligands, so what it has to transfer is how a correlated metal–ligand orbital looks electronically, not which ligands it has seen.
The comparison with the architecture of Golub et al. in Sec. II C points the same way: on this benchmark it reaches F 1 = 0.47 F_1 = 0.47 F 1 = 0.47 with its 29 orbital features and 0.62 with our 60 descriptors, against 0.65 for SpinAware, so most of the transfer comes from the spin-resolved descriptors, and processing the orbitals jointly adds a smaller improvement. This is also seen directly in the space of the model’s inputs. For every orbital we computed the distance to its nearest training orbital in the standardized 60-dimensional descriptor space and compared it with the same distance for complexes held out from training (10% of the training complexes, all their spin states and stretched geometries kept together). The benchmark orbitals lie at a median distance only 1.09 times that of the held-out complexes, and 1.8% of them fall beyond the 95th percentile of the held-out distribution, compared with 5% by construction for the held-out complexes themselves. Of the correlated orbitals (s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 ), only one lies beyond it. The distance does not follow the novelty tiers: the Tier III titanium arene [Ti(η 6 \eta^6 η 6 -C6 _6 6 H6 _6 6 )(AlCl4 _4 4 )2 _2 2 ] lies at the held-out median, whereas the Tier I complexes [VF6 _6 6 ]3 − ^{3-} 3 − and [Co(CH3 _3 3 )4 _4 4 ]− ^- − are the farthest from training. The benchmark is therefore new in its chemistry but not in the electronic patterns of its orbitals, which the descriptors map onto those of the training set. The prediction error depends only weakly on this distance (rank correlation 0.04 over all orbitals and 0.15 over the correlated ones). Only for the correlated orbitals farthest from the training data (top quarter of the distance) does the MAE rise, to 0.11 against 0.06–0.07 for the others. The residual errors trace instead to two electronic regimes, described below, that can occur in any tier.
The chromium carbonyls and the titanium arene complex contain ligands, binding modes and oxidation states that are absent from or rare in training: zero-valent chromium carrying four or five π \pi π -acidic carbonyls in [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)] and [Cr(CO)4 _4 4 (H2 _2 2 NCH2 _2 2 CO2 _2 2 )]− ^- − , an N,O-chelating amino acid in [Cr(CO)4 _4 4 (H2 _2 2 NCH2 _2 2 CO2 _2 2 )]− ^- − , and an η 6 \eta^6 η 6 -arene on titanium(II) in [Ti(η 6 \eta^6 η 6 -C6 _6 6 H6 _6 6 )(AlCl4 _4 4 )2 _2 2 ]. In the chromium(0) carbonyls, the d 6 d^6 d 6 metal donates electron density from its filled t 2 g t_{2g} t 2 g -type orbitals into the empty π ∗ \pi^* π ∗ orbitals of the carbonyls, and the carbonyls donate into the metal through σ \sigma σ bonds. The strongly correlated orbitals are therefore the metal–d–ligand bonding and antibonding combinations on either side of the frontier π \pi π gap, labeled d/p in the entropy plots. For the glycinate [Cr(CO)4 _4 4 (H2 _2 2 NCH2 _2 2 CO2 _2 2 )]− ^- − (Figure 5 ), the predicted profile follows the reference across the whole window and places its maxima on these orbitals, and at τ = 0.07 \tau= 0.07 τ = 0.07 it recovers the 26-orbital DMRG active space with F 1 = 0.98 F_1 = 0.98 F 1 = 0.98 . For the morpholine complex [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)] (Supplementary Fig. S4), SpinAware selects the same manifold with F 1 = 0.88 F_1 = 0.88 F 1 = 0.88 and 0.82 in the singlet and quintet. These active spaces are large (11–18 orbitals at τ = 0.10 \tau= 0.10 τ = 0.10 ), so the agreement cannot be attributed to a few dominant orbitals.
Figure 5. [Cr(CO)4 _4 4 (H2 _2 2 NCH2 _2 2 CO2 _2 2 )]− ^- − (6, Tier II): single-orbital entropies. An N,O-chelating amino acid on zero-valent chromium, a rare ligand and an oxidation state absent from training, is transferred accurately. Panels: singlet (a) and triplet (b) profiles; number of active orbitals (c) and F 1 F_1 F 1 (d). Upper panels: s 1 s_1 s 1 over the 40-orbital window from DMRG (blue) and SpinAware (red); dotted line, τ = 0.10 \tau= 0.10 τ = 0.10 ; dashed line, HOMO–LUMO boundary; letters, dominant orbital character. Lower panels: solid and hatched bars, τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07; dotted line, F 1 = 0.70 F_1 = 0.70 F 1 = 0.70 .
The titanium arene complex [ T i ( η 6 - C 6 H 6 ) ( A l C l 4 ) 2 ] [\mathrm{Ti}(\eta^{6}\text{-}\mathrm{C}_{6}\mathrm{H}_{6})(\mathrm{AlCl}_{4})_{2}] [ Ti ( η 6 - C 6 H 6 ) ( AlCl 4 ) 2 ] shows this transfer for a metal that is absent from training (Figure 6 and Figure 7 ). Titanium(II) is a d 2 d^{2} d 2 ion, and its two d d d electrons are partly delocalized onto the benzene ring by metal-to-arene back-donation. The shortened and lengthened C–C bonds of the coordinated ring in the crystal structure are the structural signature of this interaction. Accordingly, the dominant correlation in the singlet resides in a frontier orbital of mixed Ti d d d and arene π ∗ \pi^{*} π ∗ character (DMRG s 1 = 0.41 s_{1}=0.41 s 1 = 0.41 ). The prediction under-estimates this orbital but reproduces the correlated orbitals immediately above the gap (F 1 = 0.84 F_{1}=0.84 F 1 = 0.84 at τ = 0.10 \tau=0.10 τ = 0.10 and 0.90 0.90 0.90 at τ = 0.07 \tau=0.07 τ = 0.07 , MAE 0.029). In the triplet, all four DMRG-active orbitals are recovered, together with four further orbitals near the threshold (F 1 = 0.67 F_{1}=0.67 F 1 = 0.67 and 0.71 0.71 0.71 , MAE 0.020). The predicted mutual information reproduces both strongly coupled blocks of the reference in each spin state (Figure 7 ): the metal–arene block around the frontier gap and a second block of Ti–Cl orbitals of the tetrachloroaluminate ligands. Titanium(II) shares the d 2 d^{2} d 2 configuration of the vanadium(III) ions, which are not well predicted (see below), yet it is predicted reliably, presumably because back-donation into the arene spreads its correlation over several orbitals of moderate entropy rather than concentrating it in one strongly correlated d d d pair.
Figure 6. [ T i ( η 6 - C 6 H 6 ) ( A l C l 4 ) 2 ] [\mathrm{Ti}(\eta^{6}\text{-}\mathrm{C}_{6}\mathrm{H}_{6})(\mathrm{AlCl}_{4})_{2}] [ Ti ( η 6 - C 6 H 6 ) ( AlCl 4 ) 2 ] (7, Tier III): single-orbital entropies. An η 6 \eta^{6} η 6 -arene complex of titanium, a metal absent from training, is transferred accurately. Panels: singlet (a) and triplet (b) profiles; number of active orbitals (c) and F 1 F_{1} F 1 (d). Upper panels: s 1 s_{1} s 1 over the 40-orbital window from DMRG (blue) and SpinAware (red); dotted line, τ = 0.10 \tau=0.10 τ = 0.10 ; dashed line, HOMO–LUMO boundary; letters, dominant orbital character. Lower panels: solid and hatched bars, τ = 0.10 \tau=0.10 τ = 0.10 and 0.07; dotted line, F 1 = 0.70 F_{1}=0.70 F 1 = 0.70 .
Figure 7. [Ti(η⁶-C₆H₆)(AlCl₄)₂] (7, Tier III): mutual information. The two strongly coupled orbital blocks of the reference appear in the prediction for both spin states. Panels: singlet (left) and triplet (right); DMRG (top row) and SpinAware (bottom row). Logarithmic color scale over the 40-orbital window; red circles, rescue pairs (predicted s 1 < 0.10 s_1 < 0.10 s 1 < 0.10 and I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 with an orbital of predicted s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 ).
The residual errors arise in two electronic regimes. The first is strong static correlation . In the singlet states of the vanadium(III) ions [ V C l 6 ] 3 − [\mathrm{VCl}_{6}]^{3-} [ VCl 6 ] 3 − and [ V F 6 ] 3 − [\mathrm{VF}_{6}]^{3-} [ VF 6 ] 3 − (Figure 8 a and Supplementary Figure S2 ), DMRG assigns individual orbital entropies of about 0.6 (maximal s 1 = 0.63 s_{1}=0.63 s 1 = 0.63 and 0.66), close to the value of ln 2 ≈ 0.69 \ln2 \approx0.69 ln 2 ≈ 0.69 per orbital expected for a pair of orbitals that share two electrons equally. These large entropies identify open-shell singlet character: in an octahedral d 2 d^{2} d 2 ion the two electrons occupy two of three degenerate t 2 g t_{2g} t 2 g orbitals, and pairing them in a single orbital, as a closed-shell determinant requires, is energetically unfavorable. The singlet is therefore an equal mixture of several configurations, and a single determinant is a poor reference. SpinAware places elevated entropy on the correct V d d d orbitals but compresses its magnitude (maximal predicted s 1 ≈ 0.15 s_{1}\approx0.15 s 1 ≈ 0.15 for [ V F 6 ] 3 − [\mathrm{VF}_{6}]^{3-} [ VF 6 ] 3 − , against 0.66 from DMRG), which accounts for the largest MAEs in the set (0.044–0.047 for the vanadium singlets). Such values lie in the tail of the training distribution, where the median per-state maximal entropy is 0.22 and only 19% of the states reach a maximal entropy above 0.6, so the regression shrinks toward typical values. The error affects the magnitude of the entropy, not the position of the correlated orbitals.
Figure 8. [ V C l 6 ] 3 − \mathrm{[VCl_6]^{3-}} [ VC l 6 ] 3 − (1, Tier I): single-orbital entropies. The strongly correlated V d d d orbitals of the singlet are located, but their entropies are compressed. Panels: singlet (a) and triplet (b) profiles; number of active orbitals (c) and F 1 F_1 F 1 (d). Upper panels: s 1 s_1 s 1 over the 40-orbital window from DMRG (blue) and SpinAware (red); dotted line, τ = 0.10 \tau= 0.10 τ = 0.10 ; dashed line, HOMO–LUMO boundary; letters, dominant orbital character. Lower panels: solid and hatched bars, τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07; dotted line, F 1 = 0.70 F_1 = 0.70 F 1 = 0.70 .
Figure S2. Single-orbital entropies of 2, [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − (Tier I). Singlet and triplet, the triplet on its lowest spin-clean UHF solution. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
The second regime is weak correlation concentrated near the threshold . In the high-spin quintet of [ C o ( C H 3 ) 4 ] − [\mathrm{Co}(\mathrm{CH}_{3})_{4}]^{-} [ Co ( CH 3 ) 4 ] − and the triplet of [ V F 6 ] 3 − [\mathrm{VF}_{6}]^{3-} [ VF 6 ] 3 − (Figure 9 c and Supplementary Figure S2 ), the reference profiles are almost flat, with all entropies below 0.15. In these states the d d d electrons occupy separate orbitals with parallel spins. Exchange between parallel spins stabilizes a single high-spin deter-
Figure 9. [ C o ( C H 3 ) 4 ] − \mathrm{[Co(CH_3)_4]^-} [ Co ( C H 3 ) 4 ] − (3, Tier I): single-orbital entropies. Weak correlation near the threshold gives small errors but unstable active-space decisions in the high-spin states. Panels: singlet (a) and quintet (b) profiles; number of active orbitals (c) and F 1 F_1 F 1 (d). Upper panels: s 1 s_1 s 1 over the 40-orbital window from DMRG (blue) and SpinAware (red); dotted line, τ = 0.10 \tau= 0.10 τ = 0.10 ; dashed line, HOMO–LUMO boundary; letters, dominant orbital character. Lower panels: solid and hatched bars, τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07; dotted line, F 1 = 0.70 F_1 = 0.70 F 1 = 0.70 .
minant, and the Pauli principle forbids the pair excitations that create static correlation within the half-filled orbitals, so little static correlation remains to be resolved. SpinAware reproduces both profiles with small absolute errors (MAE 0.014–0.022), yet F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 falls to zero, because a few orbitals lying just above τ \tau τ in the reference fall just below it in the prediction, or vice versa. Small differences near the entropy threshold can therefore change the classification of an orbital. The consequences for the energies are assessed separately for the application systems. A hard threshold applied to a flat profile is an ill-conditioned decision, which is why profile-level accuracy and set-level agreement must be assessed together. The low entropies identify weakly correlated states within the reference window.
The failure cases share chemical features that are underrepresented in the training data, which suggests where the learned representation is incomplete. (i) Electron count and metal size. The training metals (Cr–Ni in oxidation states +2 and +3) span d 3 d^3 d 3 –d 8 d^8 d 8 , so the d 2 d^2 d 2 configurations of the vanadium(III) and titanium(II) complexes are unseen. Early transition metals also have more diffuse 3d orbitals and smaller intra-atomic exchange integrals, which shift the energetic and exchange descriptors on which the model relies. That [Ti(η⁶-C₆H₆)(AlCl₄)₂] is nonetheless well predicted, while the vanadium(III) singlets are not, indicates that the bottleneck is the strongly correlated t 2 g 2 t_{2g}^2 t 2 g 2 singlet manifold rather than the d 2 d^2 d 2 count itself. (ii) σ-alkyl donors. [Co(CH₃)₄]⁻ binds through
four strongly covalent Co–C σ bonds with no π component, a bonding motif absent from training, where all carbon donors are π-acceptors. Its predicted profiles are uniformly flatter than the reference in all three spin states. (iii) Spin-state discrimination is learned per metal. For the metals represented in training, the model reproduces the dependence of correlation on spin state: for [Co(CH₃)₄]⁻, both the predicted and the DMRG maximal entropies decrease from the singlet to the quintet, as expected when exchange stabilization grows with the number of unpaired electrons. For vanadium, which is absent from training, the predicted singlet is no more correlated than the triplet (maximal predicted s 1 ≈ 0.12 s_1 \approx0.12 s 1 ≈ 0.12 and 0.21 for [VCl₆]³⁻), whereas DMRG shows a strongly correlated singlet and a weakly correlated triplet. Spin awareness therefore transfers across ligands but not, at present, to unseen metals. This identifies early transition metals, and strongly correlated low-spin states in general, as the most valuable additions to future training sets. The states at risk cannot be recognized from the UHF reference alone: the closed-shell solutions of both vanadium singlets show no significant broken-symmetry instability (lowering below 0.1 mE h E_{\mathrm{h}} E h ). They are, however, recognized from their chemistry (a low-spin state of a metal absent from training) before any correlated calculation. Lowering the threshold from 0.10 to 0.07 raises the mean F 1 F_1 F 1 from 0.65 to 0.75 (Table III), and it does so in every tier, most strongly in Tier I (0.53 to 0.66). In the back-bonding complexes [Cr(CO)₄(H₂NCH₂CO₂)]⁻ and [Ti(η⁶-C₆H₆)(AlCl₄)₂] the model places the entropy of several correlated metal–ligand orbitals just below 0.10, and the permissive threshold recovers them. The average, however, conceals the opposite effect in weakly correlated states. In the triplet of [VCl₆]³⁻, F 1 F_1 F 1 falls from 0.80 to 0.62, and in the triplet of [VF₆]³⁻ the lower threshold admits 13 predicted orbitals against two in the reference, because in flat profiles a lower threshold admits orbitals that carry no static correlation. A fixed entropy threshold therefore cannot serve every state equally well.
The choice also carries an asymmetric cost. Omitting a strongly correlated orbital causes a qualitative failure of the subsequent multireference calculation, whereas including a weakly correlated one only increases its cost. A permissive entropy threshold, which favors recall, is therefore the safer screening criterion, provided the resulting candidate set can be refined by a criterion that adapts to the individual system.
Single-orbital entropy alone cannot provide this refinement: it measures how correlated each orbital is, not which orbitals are correlated with each other. Chemically, static correlation always involves at least two orbitals (the orbital that loses electron density and the one that receives it, such as the bonding and antibonding partners of a metal–ligand bond), and an orbital of modest s 1 s_1 s 1 can be indispensable if it is the partner of a strongly correlated one. Pairwise mutual information I i j I_{ij} I ij supplies this missing information. We therefore do not use any
single threshold to fix the active space. Instead, we combine the predicted entropies and mutual information into a compact pool of candidate orbitals, from which the final active space is selected by a NOON scan.
The mutual information I i j I_{ij} I ij (Eq. (20) ) measures how strongly two orbitals are correlated with each other. The pairwise SpinAware network (Sec. II B 3), trained on the DMRG pair labels, predicts I i j I_{ij} I ij for all 780 orbital pairs of the 40-orbital window. Its predictions reproduce the pattern of pairwise correlation even for chemistry absent from training. For [Ti(η 6 \eta^6 η 6 -C₆H₆)(AlCl₄)₂], the metal–arene and Ti–Cl blocks of the reference, around window orbitals 19–25 and 8–10, appear in the prediction for both spin states (Figure 7 ). For [Cr(CO)₄(H₂NCH₂CO₂)]⁻, the coupled block of metal–ligand orbitals between window orbitals 17 and 30, which describes σ \sigma σ -donation and π \pi π -back-donation between chromium and its carbonyl, amine and carboxylate ligands, is recovered (Figure 10 ). The model therefore learns not only how correlated each orbital is but also which orbitals are correlated with each other. As for the single-orbital entropies, the magnitude of the strongest correlations is underestimated. In the vanadium(III) singlets, DMRG shows a few isolated pairs
Figure 10. [Cr(CO)₄(H₂NCH₂CO₂)]⁻ (6, Tier II): mutual information. The coupled metal–ligand block is recovered; in the singlet the pool holds all 18 DMRG-active orbitals, against 13 from the entropy alone. Panels: singlet (left) and triplet (right); DMRG (top row) and SpinAware (bottom row). Logarithmic color scale over the 40-orbital window; red circles, rescue pairs (predicted s 1 < 0.10 s_1 < 0.10 s 1 < 0.10 and I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 with an orbital of predicted s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 ).
of very large mutual information (the t 2 g t_{2g} t 2 g orbitals that share the two d d d electrons), which the prediction places correctly but with much smaller values (Figure 11 and Supplementary Figure S3 ). The prediction still identifies the correct partners of the most correlated orbitals, which is the information needed to assemble an active space.
Figure 11. [VCl₆]³⁻ (1, Tier I): mutual information. The isolated, strongly correlated pairs of the singlet are placed correctly but with much smaller values. Panels: singlet (left) and triplet (right); DMRG (top row) and SpinAware (bottom row). Logarithmic color scale over the 40-orbital window; red circles, rescue pairs (predicted s 1 < 0.10 s_1 < 0.10 s 1 < 0.10 and I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 with an orbital of predicted s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 ).
Figure S3. Mutual information of 2, [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − (Tier I). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
We build the candidate pool P \mathcal{P} P from two complementary criteria. The first is the permissive entropy threshold, which recovers orbitals whose predicted entropy falls just below the selection threshold. The second is a mutual-information rescue, which adds orbitals that are only weakly correlated on their own but strongly correlated with an active orbital, such as the partner that completes a correlated pair. With A = { i : s 1 , i M L ≥ 0.10 } \mathcal{A} = \{i : s^{\mathrm{ML}}_{1,i} \ge0.10\} A = { i : s 1 , i ML ≥ 0.10 } ,
P = { i : s 1 , i M L ≥ 0.07 } ∪ { i ∉ A : max j ∈ A I i j M L > 0.005 } . (25) \mathcal{P} = \{i : s^{\mathrm{ML}}_{1,i} \ge0.07\} \cup\{i \notin\mathcal{A} : \max_{j \in\mathcal{A}} I^{\mathrm{ML}}_{ij} > 0.005\}. \tag*{(25)} P = { i : s 1 , i ML ≥ 0.07 } ∪ { i ∈ / A : j ∈ A max I ij ML > 0.005 } . (25)
The rescue threshold reflects the scale of mutual information: values of order 10 − 1 10^{-1} 1 0 − 1 , 10 − 2 10^{-2} 1 0 − 2 and 10 − 3 10^{-3} 1 0 − 3 are characteristic of nondynamic, static and dynamic correlation, respectively [35 ] , so I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 selects couplings that stand above the dynamic-correlation background. The pool is designed for completeness rather than precision: its purpose is to contain every orbital that the final active space may need, while remaining much smaller than the full window. All quantities in the pool are predicted, and no correlated calculation is required to construct it.
A fixed-budget control compares the candidate pool with entropy ranking at the same state-specific number of orbitals. The archived controls for the earlier checkpoints are reported in the Supplementary Information. With the final models and the spin-checked references, over the 11 states of Table IV, the pool and the entropy ranking at the same budget both recover 82 of the 83 DMRG-active orbitals, and for the application systems the pool recovers 28 of 29 against 25 for the entropy ranking (§ 2.0.2 C).
Summed over the 11 spin states, DMRG identifies 83 active orbitals: between 0 and 18 per state, out of the 40 window orbitals of each state (Table IV). The predicted entropy alone, at τ = 0.10 \tau= 0.10 τ = 0.10 , recovers 66 of these 83 orbitals (80%). The candidate pool recovers 82 of them (99%), while containing on average only 15.2 of the 40 window orbitals of a state. The gains are concentrated where the entropy analysis located the main weakness of the model: the strongly correlated singlets, in which the predicted entropies are compressed below the threshold. For the singlet of [Cr(CO)4 _4 4 (H2 _2 2 NCH2 _2 2 CO2 _2 2 )]− ^- − , the pool raises the number of recovered orbitals from 13 to all 18, and for the singlets of [VCl6 _6 6 ]3 − ^{3-} 3 − and [VF6 _6 6 ]3 − ^{3-} 3 − from 4 and 4 to all 8 and 6, respectively. On the benchmark, however, most of this gain is already obtained by the lower entropy threshold alone: orbitals with s 1 M L ≥ 0.07 s_1^{\mathrm{ML}} \ge0.07 s 1 ML ≥ 0.07 recover 81 of the 83 orbitals with 13.5 orbitals per state, and ranking the orbitals by predicted entropy alone, at the same budget as the pool, recovers 82, as many as the pool. The mutual information is obtained from the same DMRG calculations at no additional cost and acts as a safety net: it recovers weakly correlated partners of strongly correlated orbitals where they occur, which is the case for the application systems (§ 2.0.2 C). The reason is chemical: an orbital whose own entropy is underestimated is still the partner of a correlated pair, and the predicted mutual information with that partner remains large.
The size of the pool adapts to the electronic structure without any system-specific adjustment. It contains 10–19 orbitals for the vanadium ions, the singlet of [Co(CH3 _3 3 )4 _4 4 ]− ^- − , and the titanium complex, whose correlation is confined to the metal d shell and its nearest ligand orbitals, and 19–27 orbitals for the chromium carbonyls, whose correlation extends over the full metal–ligand σ \sigma σ - and π \pi π -manifold of six ligands. In the weakly correlated triplet of [VF6 _6 6 ]3 − ^{3-} 3 − , the rescue adds six orbitals although DMRG places no orbital above the threshold. The pool then over-covers, which costs efficiency but not completeness. No orbital is rescued in the high-spin quintet of [Co(CH3 _3 3 )4 _4 4 ]− ^- − , consistent with the absence of static correlation when all d electrons occupy separate orbitals with parallel spins. The four orbitals that DMRG places marginally above the threshold are nevertheless in the pool, because the model predicts them above 0.07 (Figure 12 ). The single orbital missed by the pool belongs to the quintet of [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)]. It is ranked 23rd by the predicted entropy and is not coupled to an orbital of the core.
Figure 12. [Co(CH3 _3 3 )4 _4 4 ]− ^- − (3, Tier I): mutual information. The coupled blocks of the reference appear in the prediction for both spin states. Panels: singlet (left) and quintet (right); DMRG (top row) and SpinAware (bottom row). Logarithmic color scale over the 40-orbital window; red circles, rescue pairs (predicted s 1 < 0.10 s_1 < 0.10 s 1 < 0.10 and I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 with an orbital of predicted s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 ).
The pool is not itself the active space: it is a compact, nearly complete set of candidates. The final active space is selected from the pool by a NOON scan, in which correlated calculations are performed for active spaces of increasing size drawn from the pool, and orbitals are retained only while their natural occupation numbers deviate significantly from 0 and 2 (§ 2.0.2 D). Because the pool contains on average about 38% of the window, the scan explores a small, chemically targeted space. In the following section we apply the complete procedure (prediction of entropies and mutual information, construction of the pool and NOON scan) to [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + (4) and [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + (8), and use the resulting active spaces in SQD.
Predicted active spaces for two application systems
We next apply the complete prediction to two systems that were withheld from the analysis above and are used in the remainder of this work for the NOON scan and for SQD. The first is the benchmark complex 4 of Tier II (Table II ), the cationic nickel(II) allyl complex [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + , a planar d8 ^8 8 complex with an η 3 \eta^3 η 3 -crotyl ligand. The second, [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + (8), is a model of CO2 _2 2 binding to cobalt that has recently served as a benchmark for SQD on quantum hardware [26 ] . C O 2 \mathrm{CO}_2 CO 2 binds end-on through one oxygen atom (η 1 \eta^1 η 1 -O, κ 1 \kappa^1 κ 1 -O, Figure 4 ), with a linear Co–O–C unit (179 ∘ 179^\circ 17 9 ∘ ) and an essentially undistorted, linear C O 2 \mathrm{CO}_2 CO 2 (O–C–O 180 ∘ 180^\circ 18 0 ∘ , C–O 1.18 1.18 1.18 Å to the bound and 1.10 1.10 1.10 Å to the free oxygen). Monodentate binding through a single donor atom is the dominant topology of the training data, and water is part of the constructed ligand library. Only the C O 2 \mathrm{CO}_2 CO 2 ligand is new, so N O O D = 1 N_{\mathrm{OOD}} = 1 N OOD = 1 (Tier I). We used the equilibrium geometry of the published Co–O dissociation series (Co–O 1.97 1.97 1.97 Å) and considered the singlet and quintet of cobalt(III) (q = + 3 q = +3 q = + 3 ) and the quartet of cobalt(II) (q = + 2 q = +2 q = + 2 ). Together, the two systems cover nickel(II) d 8 d^8 d 8 , cobalt(III) d 6 d^6 d 6 and cobalt(II) d 7 d^7 d 7 centers and multiplicities from singlet to quintet. For every state, the reference is the lowest-energy UHF solution that satisfies the spin-contamination criterion, and the prediction-quality metrics are summarized in Table V . For the singlets of 4 and 8, the stability search finds a lower UHF solution only by breaking spin symmetry (⟨ S ^ 2 ⟩ = 0.74 \langle\hat{S}^2 \rangle= 0.74 ⟨ S ^ 2 ⟩ = 0.74 and 1.15 1.15 1.15 , 10.3 10.3 10.3 kcal mol− 1 ^{-1} − 1 below the closed-shell solution for 4). Such a solution mixes singlet and triplet character and fails the spin criterion, so the closed-shell solution is the lowest spin-clean reference and is used, as for every other state. The broken-symmetry solution signals open-shell (diradical) character of these singlets, consistent with the large T 1 T_1 T 1 diagnostic of 4 (0.062). This static correlation is recovered by the multireference treatment starting from the spin-pure orbitals (Supplementary Table S8).
Table V. Predicted active spaces for the application systems. Metrics as in Table III ; pool coverage and pool size as in Table IV. q q q is the total charge. ‡The missing orbital is the singly occupied Co 3d orbital, which enters every active space through the chemical seed of the NOON scan (§ 2.0.2 D).
Table 4. TABLE IV. Coverage of the DMRG active space by the candidate pool. For each spin state, n D M R G n_{\mathrm{DMRG}} n DMRG is the number of window orbitals with DMRG s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 (the DMRG active space, at most 40 orbitals per state). The next columns give how many of these orbitals are found by the predicted entropy alone (s 1 M L ≥ 0.10 s_1^{\mathrm{ML}} \ge0.10 s 1 ML ≥ 0.10 , and ≥ 0.07 \ge0.07 ≥ 0.07 without the mutual-information rescue) and by the candidate pool P \mathcal{P} P , and the size of the pool out of the 40 orbitals of the window. The last row sums the active orbitals over the 11 spin states and averages the pool size. In the column for s 1 M L ≥ 0.07 s_1^{\mathrm{ML}} \ge0.07 s 1 ML ≥ 0.07 , the number of orbitals kept is given in parentheses (mean in the last row).
In the nickel(II) allyl complex 4 (Figure 13 ), the singlet is dominated by two orbitals of mixed Ni d d d /allyl π \pi π character, the occupied frontier orbital (s 1 = 0.41 s_1 = 0.41 s 1 = 0.41 ) and a virtual partner (s 1 = 0.37 s_1 = 0.37 s 1 = 0.37 ), as expected for the partially diradical character of an η 3 \eta^3 η 3 -allyl–metal interaction. The prediction locates both orbitals but compresses their entropies (0.18 and 0.23), and it places several moderately correlated orbitals just below the threshold. The active set is therefore incomplete (six of ten orbitals at τ = 0.10 \tau= 0.10 τ = 0.10 , F 1 = 0.67 F_1 = 0.67 F 1 = 0.67 ), and the permissive threshold recovers 12 of 20 (F 1 = 0.73 F_1 = 0.73 F 1 = 0.73 ). The triplet is reproduced closely (MAE 0.016, F 1 = 0.67 F_1 = 0.67 F 1 = 0.67 and 0.70).
Figure 13. [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + (4, Tier II): single-orbital entropies. The two dominant Ni d d d /allyl orbitals of the singlet are located but under-predicted. Panels: singlet (a) and triplet (b) profiles; number of active orbitals (c) and F 1 F_1 F 1 (d). Upper panels: s 1 s_1 s 1 over the 40-orbital window from DMRG (blue) and SpinAware (red); dotted line, τ = 0.10 \tau= 0.10 τ = 0.10 ; dashed line, HOMO–LUMO boundary; letters, dominant orbital character. Lower panels: solid and hatched bars, τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07; dotted line, F 1 = 0.70 F_1 = 0.70 F 1 = 0.70 .
For 8 (Figure 14 ), the cobalt(III) singlet is recovered with F 1 = 0.60 F_1 = 0.60 F 1 = 0.60 and 0.86, but the prediction over-estimates the entropies of the two lowest virtual orbitals, which DMRG places at or below the threshold, and under-estimates those of the two highest occupied orbitals (DMRG s 1 = 0.13 s_1 = 0.13 s 1 = 0.13 , predicted below 0.04). The cobalt(II) quartet and the cobalt(III) quintet are examples of the near-threshold regime identified for the benchmark: DMRG assigns three orbitals entropies above 0.10 in each, with maxima of 0.12 and 0.18. The prediction follows these flat profiles closely (MAE 0.013 and 0.011) but places only one orbital above the threshold (F 1 = 0.50 F_1 = 0.50 F 1 = 0.50 at τ = 0.10 \tau= 0.10 τ = 0.10 , 0.73 and 0.55 at τ = 0.07 \tau= 0.07 τ = 0.07 ).
Figure 14. [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + (8, Tier I): single-orbital entropies. The flat profiles of the cobalt(II) quartet and the cobalt(III) quintet are reproduced closely, but only their strongest orbital is placed above the threshold. Panels: cobalt(III) singlet (a), cobalt(II) quartet (b) and cobalt(III) quintet (c) profiles; number of active orbitals (d) and F 1 F_1 F 1 (e). Upper panels: s 1 s_1 s 1 over the 40-orbital window from DMRG (blue) and SpinAware (red); dotted line, τ = 0.10 \tau= 0.10 τ = 0.10 ; dashed line, HOMO–LUMO boundary; letters, dominant orbital character. Lower panels: solid and hatched bars, τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07; dotted line, F 1 = 0.70 F_1 = 0.70 F 1 = 0.70 .
The predicted mutual-information matrices reproduce the coupling patterns of the reference in both systems (Figure 15 and Figure 16 ). For the singlet of 4, the strongest pair of the reference, which couples the two dominant Ni d d d /allyl orbitals, appears at the same position in the prediction, and the rescue pairs run along the rows and columns of these two orbitals (Figure 15 ). The rescue thus recovers the partners of the orbitals whose entropies are compressed, the case in which the mutual information is most needed. For the quartet of 8, the rescued pairs coincide with the coupled blocks of the reference matrix (Figure 16 ).
Figure 15. [Ni(η 3 \eta^{3} η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^{+} + (4, Tier II): mutual information. The strongest pair of the singlet is reproduced, and the rescues follow the rows of its two dominant orbitals. Panels: singlet (left) and triplet (right); DMRG (top row) and SpinAware (bottom row). Logarithmic color scale over the 40-orbital window; red circles, rescue pairs (predicted s 1 < 0.10 s_1 < 0.10 s 1 < 0.10 and I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 with an orbital of predicted s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 ).
Figure 16. [ C o ( H 2 O ) 5 ( C O 2 ) ] 2 + / 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{2+/3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 2 + /3 + (8, Tier I): mutual information. In the cobalt(II) quartet, the rescued pairs coincide with the coupled blocks of the reference. Panels: cobalt(III) singlet (left), cobalt(II) quartet (center) and cobalt(III) quintet (right); DMRG (top row) and SpinAware (bottom row). Logarithmic color scale over the 40-orbital window; red circles, rescue pairs (predicted s 1 < 0.10 s_1 < 0.10 s 1 < 0.10 and I i j > 0.005 I_{ij} > 0.005 I ij > 0.005 with an orbital of predicted s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 ).
For the five states, the candidate pool recovers 28 of the 29 DMRG-active orbitals with 12.2 orbitals on average (Table V), compared with 15 (52%) from the entropy alone. The remaining one, the singly occupied Co 3d orbital of the quintet of 8, is included by the chemical seed. The largest gain occurs for the singlet of 4, in which the pool recovers all ten active orbitals against six from the entropy alone, because the rescue adds the partners of the two dominant orbitals whose entropies are compressed. Here the mutual information adds information that the entropies do not contain: at the same budget, ranking by predicted entropy alone recovers 25 of the 29 orbitals and the lower threshold alone 23, whereas the pool recovers 28, including the Ni–allyl bond orbital of the singlet of 4 (ranked 22nd by entropy). In an application, however, no DMRG reference is available to tell which pool orbitals are required. The pool must therefore be reduced to a final active space by a criterion that relies only on the correlated calculation itself. This reduction is also a practical necessity for quantum hardware, where every spatial orbital in the active space requires two qubits in the Jordan–Wigner mapping, so that the pools of 6–20 orbitals would require 12–40 qubits. The following sub-section describes the NOON scan that selects the final active space from the pool.
Natural-orbital occupation scans
The candidate pool tells us which orbitals may need to be active, but in an application there is no DMRG reference to decide which of them actually are. That decision must come from the correlated wavefunction itself. The most direct measure is provided by the NOONs, the eigenvalues n k n_k n k of the one-particle reduced density matrix
γ p q = ∑ σ ⟨ Ψ ∣ a ^ p σ † a ^ q σ ∣ Ψ ⟩ (26) \gamma_{pq} = \sum_{\sigma} \langle\Psi|\hat{a}^{\dagger}_{p\sigma}\hat{a}_{q\sigma}|\Psi\rangle \tag*{(26)} γ pq = σ ∑ ⟨ Ψ∣ a ^ p σ † a ^ q σ ∣Ψ ⟩ (26)
with γ u k = n k u k \gamma\,u_k = n_k u_k γ u k = n k u k and 0 ≤ n k ≤ 2 0 \le n_k \le2 0 ≤ n k ≤ 2 . For a single determinant every orbital is either doubly occupied (n k = 2 n_k = 2 n k = 2 ) or empty (n k = 0 n_k = 0 n k = 0 ), apart from the singly occupied orbitals of an open-shell state (n k = 1 n_k = 1 n k = 1 ). Static correlation moves electron density from occupied to virtual natural orbitals, so that occupations depart from these integer values. The size of the departure measures how strongly an orbital participates in the correlation, and occupations outside the range 0.02–1.98 are the classic signature of an orbital that must be treated as active [31 , 36 , 86 ] . Unlike the single-orbital entropy, which is computed in a fixed window, the NOONs are properties of the wavefunction in a given active space. They therefore test directly whether adding an orbital to that space changes the correlated description.
A single calculation in the full pool would show the final occupations, but not which orbitals are responsible for them. Correlation also involves pairs of orbitals: an occupied orbital can lose density only if a suitable virtual partner is present to receive it. An orbital added without its partner therefore leaves the occupations unchanged, however important it is. We therefore grow the active space one orbital at a time, in the order given by the SpinAware prediction, and follow the NOONs after each step. A sudden change in the occupations marks the step at which a correlated pair is completed, a flat step marks an orbital that is only a passenger, and a plateau indicates that the occupation-based diagnostics have stabilized over the sampled range. The scan turns the predicted ranking into a testable statement: a good ranking produces its jumps early and reaches the plateau with few orbitals.
Each scan starts from the lowest spin-clean UHF reference on which the predictions were made (def2-SVP) and works in the fixed basis of its UNOs [31 ] . The initial active space is fixed by a chemical rule rather than by the model. It contains (i) every UNO with a fractional occupation 0.10 < n < 1.90 0.10 < n < 1.90 0.10 < n < 1.90 , which includes the singly occupied orbitals of open-shell states, and (ii) the five orbitals of largest metal-3d character. A singly occupied orbital has a definite occupation in the reference determinant and therefore zero single-orbital entropy there in the M s = S M_s = S M s = S component used for the labels (in other M s M_s M s components of the same state it carries up to ln 2 \ln2 ln 2 ). In the correlated state its entropy is usually small (below about 0.03) but can reach the threshold, as for one singly occupied site of the cobalt(III) quintet of 8 (s 1 = 0.106 s_1 = 0.106 s 1 = 0.106 ), so no entropy-based criterion, predicted or DMRG, can select it reliably, yet it must be active to describe the
spin coupling.
The 3d orbitals are identified by projecting the doubly occupied and the empty UNOs, separately, onto the metal 3d atomic orbitals, in the spirit of AVAS [30 ] . This rotation within each block leaves the UHF density unchanged. Every space must contain five orbitals with more than 50% metal-3d projection, otherwise it is rejected. Beyond this seed, all orbitals are taken from the SpinAware prediction, in the order defined by the pool: first orbitals with predicted s 1 ≥ 0.07 s_1 \ge0.07 s 1 ≥ 0.07 in decreasing order of s 1 s_1 s 1 , then the mutual-information-rescued orbitals, and finally the remaining window orbitals. Each predicted window orbital is mapped onto the UNO of maximum overlap (∣ S ∣ ≥ 0.3 |S| \ge0.3 ∣ S ∣ ≥ 0.3 ), and orbitals already in the space, such as the singly occupied ones, are skipped. After each addition, the wavefunction is obtained by exact CASCI up to 16 orbitals and by DMRG-CASCI from 17 to 20 orbitals (block2, spin adapted, Fiedler orbital ordering, 20 sweeps with bond dimension increased from 250 to a final M = 1000 M = 1000 M = 1000 ). For the 17-orbital spaces of all five states, the first DMRG step, DMRG reproduces the exact CASCI energy to 10 − 7 10^{-7} 1 0 − 7 E h E_{\mathrm{h}} E h and the NOONs to 3 × 10 − 6 3 \times10^{-6} 3 × 1 0 − 6 . The scripts and results of all NOON scans, CASCI, NEVPT2 and DMRG calculations are available from the corresponding authors upon reasonable request.
To follow the scan, we monitor the NOONs n n n (0 ≤ n ≤ 2 0 \le n \le2 0 ≤ n ≤ 2 ) of the active space. In a single-determinant description every occupation is an integer: n = 2 n = 2 n = 2 for a doubly occupied orbital, n = 0 n = 0 n = 0 for an empty (virtual) one and n = 1 n = 1 n = 1 for a singly occupied one. Static correlation mixes in further configurations and moves electron density from occupied into virtual orbitals, so the occupations deviate from these integers. We quantify the deviation in three ways: the depletion 2 − n 2 - n 2 − n of an occupied natural orbital (the density it has lost), the population n n n of a virtual natural orbital (the density it has gained), and, for open-shell states, the deviation ∣ n − 1 ∣ |n - 1| ∣ n − 1∣ of a singly occupied orbital from exactly one electron. For each step we report the strongest correlated pair, i.e. the largest depletion among the occupied and the largest population among the virtual natural orbitals, together with the second-strongest pair and the largest ∣ n − 1 ∣ |n - 1| ∣ n − 1∣ . A deviation above 0.02 (occupations below 1.98 or above 0.02), the conventional criterion for an active orbital and close to the occupation equivalent of τ = 0.07 \tau= 0.07 τ = 0.07 (§ 2.0.2 A), indicates clear static correlation. The two members of a pair need not have equal values, because density can flow from several occupied orbitals into one virtual orbital, or from one occupied orbital into several virtual ones. For example, in the 16-orbital space of the singlet of [Ni(η 3 \eta^{3} η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^{+} + (4), the most depleted occupied orbital has n = 1.955 n = 1.955 n = 1.955 (depletion 0.045), whereas the most populated virtual orbital has n = 0.064 n = 0.064 n = 0.064 .
In the closed-shell singlet of [Ni(η 3 \eta^{3} η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^{+} + (4), the correlation resides in the bond between nickel and the allyl ligand, and a rule based on the metal d shell alone misses most of it (Figure 17 a). The five Ni d orbitals already show moderate correlation (pair strength 0.026), from a filled Ni d orbital to an empty one. The first three predicted orbitals, windows 19, 21 and 25, raise it gradually to 0.038. The fourth, window 22, completes the allyl→ \rightarrow → Ni pair at nine orbitals and raises the strongest pair from 0.038 to 0.049 (empty orbital) and from 0.015 to 0.026 (filled orbital). This sequence illustrates the pair principle: the orbitals with the highest entropy are the right ones, but the correlation appears in full only once both partners of the pair are present. The two values differ because the empty Ni d orbital receives density from several filled orbitals, Ni d and allyl, at once. Further orbitals add small contributions, and a final step at 16 orbitals (window 30, a mutual-information-rescued orbital) brings the pair strength to 0.064, within 0.003 of its 20-orbital value. In this closed-shell state, the reference wavefunction offers no spin-polarized pair from which the correlated ligand and orbitals could have been identified. The predicted ranking places them among its first four picks.
Figure 17. [Ni(η 3 \eta^3 η 3 -C 4 H 7 \mathrm{C}_4\mathrm{H}_7 C 4 H 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + (4): NOON scans. Singlet (a) and triplet (b). The active space grows from the five Ni d orbitals by one predicted orbital per step (window index of each added orbital on top; numbered from 0, HOMO = 19). Solid lines: strongest correlated pair, shown as 2 − n 2-n 2 − n of the occupied (blue) and n n n of the virtual (red) natural orbital; dashed lines: second pair; purple: ∣ n − 1 ∣ \lvert n-1\rvert ∣ n − 1 ∣ of singly occupied orbitals. Dotted line: 0.02. Gray region: DMRG-CASCI; otherwise exact CASCI.
[Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + ^{2+} 2 + /3 + ^{3+} 3 + (8) is a control in which the correlation is weaker (Figure 18 ). In the cobalt(III) singlet, the d shell alone gives a pair strength of 0.024 (t 2 g → e g t_{2g}\rightarrow e_g t 2 g → e g ), and the first eight predicted orbitals raise it only to 0.027. The two highest occupied orbitals, windows 19 and 18, whose entropies the prediction underestimates (§ 2.0.2 C), enter only at 14 and 16 orbitals, and the second raises the strongest pair to 0.038, within
Figure 18. [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + (8): NOON scans. Cobalt(III) singlet (a), cobalt(II) quartet (b) and cobalt(III) quintet (c). In the singlet the strongest pair grows when the two highest occupied orbitals are added (14 and 16 orbitals); in the quartet the weak ligand pair forms when the mutual-information-rescued occupied orbitals are added. Lines and shading as in Figure 17 .
0.001 of its 20-orbital value. In the cobalt(III) quintet all values stay below 0.01, and the singly occupied orbitals remain at one electron. The high-spin triplet of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) behaves similarly: all pair strengths stay below 0.02, and the two largest steps, at 10 and 17 orbitals, add only weak correlation between a filled and an empty orbital (Figure 17 b). In these states the predicted orbitals do not create spurious correlation but only enlarge the space.
The cobalt(II) quartet of [ C o ( H 2 O ) 5 ( C O 2 ) ] 2 + / 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{2+/3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 2 + /3 + (8) provides a direct test of the mutual-information rescue (Figure 18 b). The scan starts with a weak coupling (0.006) between a filled Co d d d orbital and the singly occupied orbitals. The orbitals ranked highest by entropy, windows 26, 25, 23 and 13, leave the occupations almost unchanged (0.007). The weak ligand pair forms only when the rescued occupied partners, windows 16, 3 and 4, are added, raising the strongest pair to 0.019 at 13 orbitals, and two further rescued orbitals, windows 15 and 14, bring it to 0.027 at 18 orbitals. Without the rescue, these partners would have been ranked among the remaining window orbitals, and the pair would have appeared much later in the scan. This is the pair principle seen from the other side: the entropy identifies the virtual half of the pair, and the mutual information supplies its occupied partner.
Extending every scan to 20 orbitals with DMRG-CASCI tests the stability of the occupation-based correlation diagnostics (Table VI ). We characterize this stability by the change in the strongest pair strength between a given size and the largest scanned space. This diagnostic tracks the dominant static correlation, rather than providing a bound on the spin-gap error. For both complexes the occupations reach a plateau. The singlet of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) reaches it at 16 orbitals, when window orbital 30 is added (0.064 at 16 and 0.067 at 20 orbitals), and the triplet takes a last step at 17 orbitals, from 0.011 to 0.018. For [ C o ( H 2 O ) 5 ( C O 2 ) ] 2 + / 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{2+/3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 2 + /3 + (8), the singlet reaches its plateau at 16 orbitals (0.038, against 0.039 at 20), the quintet already at 12, and the quartet takes a last step at 18 orbitals (from 0.021 to 0.027). At 16 orbitals the pair strength of every state lies within 0.009 of its 20-orbital value, below the 0.02 that marks clear static correlation. The size of the energy calculations is set by NEVPT2: 16 orbitals is the largest active space for which the strongly contracted NEVPT2 correction remains tractable with the full four-particle reduced density matrix, so the Ni triplet and the Co quartet are used at 16 orbitals although their occupations take a last small step at 17 and 18 orbitals. We therefore use 16 orbitals for all spin states of a complex in the energy calculations: CAS(16,16) for the singlet and triplet of [Ni(η 3 \eta^3 η 3 -C 4 H 7 \mathrm{C}_4\mathrm{H}_7 C 4 H 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + , and CAS(22,16), CAS(19,16) and CAS(20,16) for the singlet, quartet and quintet of [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + . Energetic convergence is assessed separately below and in the Supplementary Information.
Table VI. Summary of the NOON scans. Pair strength is the larger of 2 − n 2-n 2 − n (occupied) and n n n (virtual) for the strongest correlated pair. Window orbitals are numbered from 0 (HOMO = 19). The last column gives the active space used for the energy calculations (same number of orbitals for all spin states of a complex) and, in parentheses, the size at which the strongest pair takes its last step.
Converged occupations are a necessary condition for an adequate active space, but not a sufficient one. We therefore test the selected spaces directly on the quantity of interest, the spin-state gap. Table VII reports the gaps obtained by CASCI and by CASCI with strongly contracted NEVPT2 for active spaces of equal size in both spin states, built from the same predicted ranking, and compares them with UCCSD(T) for the full molecule. The CASCI energies are exact at all sizes, and we use NEVPT2 evaluated from the DMRG one-to four-particle reduced density matrices (bond dimension 1000), which reproduces conventional NEVPT2 to 2 × 10 − 7 E h 2 \times10^{-7}\ E_{\mathrm{h}} 2 × 1 0 − 7 E h at 13 orbitals, and at 16 orbitals the DMRG energy matches exact CASCI to 3 × 10 − 7 E h 3 \times10^{-7}\ E_{\mathrm{h}} 3 × 1 0 − 7 E h . The CASCI gaps lie far from the reference, by about 36–45 kcal mol− 1 ^{-1} − 1 for [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + (4) at all sizes and 21–24 kcal mol− 1 ^{-1} − 1 for [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + (8) up to 13 orbitals, because the dynamic correlation outside the active space, which CASCI omits, differs strongly between the two spin states. With NEVPT2 the gaps change by at most 4.5 kcal mol− 1 ^{-1} − 1 between 9 and 16 orbitals but do not approach the coupled-cluster reference systematically: at 16 orbitals they differ from UCCSD(T) by 6.3 kcal mol− 1 ^{-1} − 1 for [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + and 5.3 kcal mol− 1 ^{-1} − 1 for [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + , and at 9–13 orbitals by 4.2–5.0 and 0.9–3.6 kcal mol− 1 ^{-1} − 1 . For [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + the NEVPT2 gap changes by 1.8 kcal mol− 1 ^{-1} − 1 between 13 and 16 orbitals, and for [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + it still increases by 2.8 kcal mol− 1 ^{-1} − 1 over this range. Comparison between multireference perturbation theory in the predicted spaces and full-molecule coupled cluster quantifies the cross-method differences at the present basis and fixed geometries. In particular, the remaining change for [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + warrants caution. The largest changes occur at the last step. For [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + , the CASCI gap drops by 21.7 kcal mol− 1 ^{-1} − 1 from 13 to 16 orbitals, when the two highest occupied orbitals enter the space. NEVPT2 absorbs most of this change, but the corrected gap still rises by 2.8 kcal mol− 1 ^{-1} − 1 . For [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + , adding the rescued orbital 30 at 16 orbitals changes the corrected gap by 1.8 kcal mol− 1 ^{-1} − 1 . At 16 orbitals the corrected gap therefore still depends on which ligand orbitals lie at the boundary of the active space. The T1 diagnostic of the singlet of [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + (0.062, above the usual 0.05 guideline) is consistent with the partly multiconfigurational character identified in the scan and also limits the status of UCCSD(T) as a definitive reference.
Table VII. Spin-state gaps for active spaces of equal size. Δ E = E \Delta E = E Δ E = E (low spin) − E E E (high spin) in kcal mol− 1 ^{-1} − 1 ; negative values mean that the low-spin state is lower. CASCI gaps from exact CASCI; NEVPT2 is strongly contracted, from DMRG reduced density matrices (bond dimension 1000). CC: UCCSD(T) for the full molecule (frozen core). S, T, Q: singlet, triplet, quintet. def2-SVP throughout.
An archived fixed-budget comparison of the earlier model spaces with metal-d d d plus frontier completion and a specified ligand-aware AVAS selection is reported in the Supplementary Information. All three selections match the active-electron allocation of the earlier ML spaces. The final-model spaces use different active-electron allocations, so the interpretation of those controls is restricted to the earlier spaces.
The scans support a simple protocol that combines chemical rules with the prediction. The active space is started from the metal d d d shell and the singly occupied orbitals, and it is completed in the order of the predicted pool until the strongest pair strength lies within 0.01 of its value in the largest scanned space, with the same number of orbitals for all spin states of a complex. When the correlation extends into the metal–ligand bonds, as in the singlet of [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + (4), the prediction supplies a ligand-orbital ranking: a metal-d d d shell rule alone does not identify the correlated ligand orbitals, whereas the predicted ranking places them among its first picks. When the correlation is confined to the d d d shell, as in the quartet and quintet of [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + (8) and the triplet of [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + (2), the rule alone suffices, and the predicted orbitals remain passengers without distorting the result. In the cobalt(III) singlet, two occupied orbitals whose entropies are underpredicted enter only at 14 and 16 orbitals, so the scan detects their contribution only at the end of the range used for the energies. Throughout, the scans confirm that high entropy alone does not make an orbital important in a given space: correlation appears only when both partners of a pair are present, which is why the pairwise information in the pool is essential. For every scanned space we retain the active-space Hamiltonian, constructed exactly as in the CASCI calculations, so that quantum and classical solvers address the same problem: 80 Hamiltonians for [Ni(η 3 \eta^3 η 3 -C4 _4 4 H7 _7 7 )(NCCH3 _3 3 )2 _2 2 ]+ ^+ + and [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + , from 5 to 20 spatial orbitals (10–40 qubits).
Every spatial orbital requires two qubits in the Jordan–Wigner mapping, so the 9-, 11-, 13- and 16-orbital spaces of the scans correspond to 18, 22, 26 and 32 qubits. We carry all four sizes to quantum hardware, with the same number of orbitals for both states of a complex (Sec. III E). Because the quantum calculation and CASCI solve the same active-space Hamiltonian, the comparison between them isolates the error of the quantum solver, while the change of the gap with the size of the space is the active-space truncation, quantified classically in Table VII.
Sample-based quantum diagonalization on quantum hardware
The active space selected by SpinAware determines the size of the quantum problem: an active space of n a c t n_{\mathrm{act}} n act orbitals requires 2 n a c t 2n_{\mathrm{act}} 2 n act qubits in the Jordan–Wigner mapping. A compact but complete active space is therefore what makes a transition-metal complex accessible to current quantum processors, whose circuits are limited by gate errors and decoherence. The aim of this section is to show that the predicted active spaces can be carried through to such a processor and still reproduce the spin-state gap of the exact classical calculation in the same space. We use SQD [22 , 23 ] , currently the main approach for chemistry on pre-fault-tolerant hardware beyond 30 qubits, because it needs no variational optimization on the device and is robust to noise. Variational algorithms [87 , 88 , 89 ] , by contrast, must optimize circuit parameters on the device, where flat energy landscapes slow the optimization [90 , 91 ] , and they need error mitigation and verification to reach chemical accuracy [92 , 93 , 94 , 95 , 96 ] . All active spaces studied here (18–32 qubits) can still be solved exactly by CASCI. This is deliberate: the exact energy of the same Hamiltonian is available as a reference, so every hardware result can be checked. The calculations are a proof of concept for the workflow and make no claim of quantum advantage. SQD is a hybrid procedure in which the quantum computer only supplies electronic configurations and a classical computer does the diagonalization. Each measurement of the 2 n a c t 2n_{\mathrm{act}} 2 n act qubits yields a bitstring x ∈ { 0 , 1 } 2 n a c t x \in\{0,1\}^{2n_{\mathrm{act}}} x ∈ { 0 , 1 } 2 n act that encodes one occupation-number vector, i.e. one Slater determinant, with one bit per spin orbital. Bitstrings with the wrong numbers of α \alpha α and β \beta β electrons, which arise mainly from hardware noise, are repaired by self-consistent configuration recovery, in which bits are flipped according to the current average orbital occupations. The distinct α \alpha α and β \beta β strings of a batch of samples are combined as a tensor product, and the active-space Hamiltonian is diagonalized in the resulting subspace with a constraint on ⟨ S ^ 2 ⟩ \langle\hat{S}^{2}\rangle ⟨ S ^ 2 ⟩ . Spin symmetry can alternatively be imposed on the quantum side, by spin-preserving circuits [97 , 98 ] or by working in a spin-adapted representation [99 ] . Because the diagonalization is variational, the SQD energy is an upper bound to the exact energy E C A S C I E_{\mathrm{CASCI}} E CASCI of the same Hamiltonian, which serves as the reference throughout. It approaches E C A S C I E_{\mathrm{CASCI}} E CASCI when the subspace contains the determinants that carry the ground-state wavefunction. The classical part is identical for all calculations below. What differs is the source of the bitstrings, and the accuracy of SQD depends entirely on how well these sources find the important determinants. The Hamiltonians are those of the NOON scans (see § 2.0.2 D), and both spin states of a complex are treated in active spaces of equal size.
We compare three sources of samples. (i) LUCJ. The local unitary cluster Jastrow (LUCJ) ansatz [100 ] prepares an approximation to the ground state on the quantum computer, and measuring it gives determinants with probabilities close to their weights in that state. Even on ideal hardware, such ground-state sampling finds new determinants inefficiently [101 ] . The circuit consists of orbital rotations and diagonal (number–number) Jastrow terms, and in the local variant the Jastrow terms act only between neighboring qubits, which matches the hardware connectivity. We use one repetition with parameters taken from the CCSD amplitudes of the active space, without variational optimization (ffsim [81 ] ), and draw 40 000 shots. (ii) SqDRIFT. SqDRIFT [102 ] uses no ansatz and no parameters. It builds on sample-based Krylov quantum diagonalization [103 ] : the time-evolved states e − i H ^ t ∣ H F ⟩ e^{-i\hat{H}t}\lvert\mathrm{HF}\rangle e − i H ^ t ∣ HF ⟩ of the Hartree–Fock determinant spread over the determinants that couple to it through the Hamiltonian, and therefore sample the determinants that matter for the ground state. The time evolution is approximated by the qDRIFT protocol [104 ] : each circuit applies a random sequence of L = 10 L = 10 L = 10 exponentials of individual Hamiltonian terms, drawn with probabilities proportional to their coefficients. Many such random circuits are run, and their samples are pooled. Under mild assumptions this gives a convergence guarantee for the SQD energy [102 ] . We use 200 random circuits with 100 shots each at two evolution times, t = 1 t = 1 t = 1 and 10 ℏ / E h 10 \hbar/E_{\mathrm{h}} 10ℏ/ E h , giving 40 000 shots per state. (iii) Uniformly random bitstrings. Every bit is set to 0 or 1 with probability one half, with the same number of samples and the same SQD settings, and without any quantum computer. This is a control experiment: configuration recovery and the tensor-product subspace can by themselves produce a useful subspace from unstructured input, so a hardware result is only meaningful if it is better than random input.
The hardware calculations used ibm_phoenix (IBM Nighthawk r2, a square lattice in which each qubit couples to four neighbors) and, for one LUCJ comparison, ibm_fez (IBM Heron r2, a heavy-hexagon lattice with at most three neighbors per qubit), without error mitigation. The active-space Hamiltonians, the sampling and SQD scripts and the raw measurement counts of all hardware and simulator runs are available from the corresponding authors upon reasonable request. The cost of a circuit on hardware is characterized by three quantities, given in Table VIII for the largest case, the singlet of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) at 32 qubits: the number of two-qubit gates, the dominant source of error; the depth, i.e. the number of sequential layers of native gates after transpilation, which sets the run time and the exposure to decoherence; and the estimated success probability (ESP), the product of all gate and readout fidelities, which approximates the fraction of shots that are free of any error. The two circuit types differ characteristically. Each SqDRIFT exponential acts on a string of qubits set by the Jordan–Wigner mapping and is applied one after the other, so the circuit is deep (about 1170 layers on average) but contains few two-qubit gates (518). The LUCJ circuit is built from layers of gates that act in parallel, so it is shallower (430–570 layers) but contains about twice as many two-qubit gates (928 on ibm_phoenix and 1077 on ibm_fez, where the sparser heavy-hexagon lattice requires additional swap gates). Because the gate count dominates the error, the ESP of SqDRIFT (0.38) is 3 and 30 times higher than that of LUCJ (0.12 and 0.012). Even for SqDRIFT most shots therefore contain at least one error. SQD tolerates this because configuration recovery and the variational diagonalization do not require error-free samples. Each source used 5–13 s of processor time. The processors, the qubit layout and further circuit details are given in Supplementary Section S7. All sources were processed with the same SQD settings (qiskit-addon-sqd [82 ] ; three batches of 300 samples), and every value is the mean and standard deviation over six SQD seeds or six independent random sets.
Table VIII. Circuits for the singlet of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) at 32 qubits, CAS(16,16). Two-qubit gates (2q), depth after transpilation to the native gates of the processor, estimated success probability (ESP) and processor time per source (40 000 shots). SqDRIFT values: mean over the 400 random circuits (Supplementary Table S13).
At 32 qubits (16 orbitals), the largest spaces selected by the NOON scans, SqDRIFT on ibm_phoenix reaches chemical accuracy for both states of both complexes already with 25 circuits per evolution time (Figure 19 ). For the Ni singlet the error falls from 0.31 to 0.27 kcal mol− 1 ^{-1} − 1 between 25 and 200 circuits, whereas noiseless SqDRIFT starts at 3.5 kcal mol− 1 ^{-1} − 1 and reaches 0.58 ± 0.08 0.58 \pm0.08 0.58 ± 0.08 kcal mol− 1 ^{-1} − 1
Figure 19. SqDRIFT on 32 qubits: noiseless simulation and ibm_phoenix. 16-orbital active spaces of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4; a) and [ C o ( H 2 O ) 5 ( C O 2 ) ] 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 3 + (8; b). Left: SQD error relative to CASCI for the low-spin (filled symbols) and high-spin (open symbols) state as a function of the number of random circuits per evolution time. Center: spin-state gap compared with CASCI. Right: gap with the NEVPT2 correction of the same active space, compared with CASCI+NEVPT2 and full-molecule UCCSD(T). Gray diamonds: uniformly random bitstrings with the same total number of samples. Green band: within 1 kcal mol− 1 ^{-1} − 1 of the reference. Error bars: standard deviation over six SQD seeds or six random sets.
(six SQD seeds) only at 200 circuits. The Co singlet behaves in the same way (3.4 kcal mol− 1 ^{-1} − 1 at 25 and 0.82 ± 0.07 0.82 \pm0.07 0.82 ± 0.07 kcal mol− 1 ^{-1} − 1 at 200 noiseless circuits, against 0.16 kcal mol− 1 ^{-1} − 1 on the hardware). The reason is sample diversity: noiseless randomized evolutions from the Hartree–Fock state remain concentrated on few configurations, while hardware noise spreads the samples over many more determinants, which the tensor-product subspace and configuration recovery turn into a larger and better subspace. For the weakly correlated high-spin states the noiseless samples suffice as well (0.22 and 0.06 kcal mol− 1 ^{-1} − 1 at 200 circuits). With 200 circuits, the hardware spin gaps are − 5.34 ± 0.06 -5.34 \pm0.06 − 5.34 ± 0.06 kcal mol− 1 ^{-1} − 1 for [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4, CASCI − 5.44 -5.44 − 5.44 ) and 0.10 0.10 0.10 kcal mol− 1 ^{-1} − 1 for [ C o ( H 2 O ) 5 ( C O 2 ) ] 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 3 + (8, CASCI 13.96). With the NEVPT2 correction they become − 35.46 -35.46 − 35.46 and 20.21 20.21 20.21 kcal mol− 1 ^{-1} − 1 , against CASCI+NEVPT2 values of − 35.57 -35.57 − 35.57 and 20.17 20.17 20.17 kcal mol− 1 ^{-1} − 1 and UCCSD(T) values of − 41.89 -41.89 − 41.89 and 14.83 14.83 14.83 kcal mol− 1 ^{-1} − 1 .
We then repeated the hardware calculation for the 9-, 11-, 13- and 16-orbital spaces of the predicted ranking (Figure 20 ). At every size and for every state, the SqDRIFT error stays below 0.3 kcal mol− 1 ^{-1} − 1 . It is below 0.025 kcal mol− 1 ^{-1} − 1 up to 26 qubits and grows to 0.12–0.27 kcal mol− 1 ^{-1} − 1 at 32 qubits. The hardware gaps follow the CASCI gaps to within 0.10 kcal mol− 1 ^{-1} − 1 at all sizes, so that the hardware reproduces the size dependence of the gap analyzed in Table VII : with the NEVPT2 correction, the gap of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) varies between − 37.7 -37.7 − 37.7 and − 35.5 -35.5 − 35.5 kcal mol− 1 ^{-1} − 1 and that of [ C o ( H 2 O ) 5 ( C O 2 ) ] 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 3 + (8) between 15.7 and 20.2 kcal mol− 1 ^{-1} − 1 as the space grows from 9 to 16 orbitals. Up to 26 qubits, uniformly random bitstrings are as accurate as the hardware samples (errors below 0.01 kcal mol− 1 ^{-1} − 1 ), because the subspace they span covers most of the configuration space. At 32 qubits this is no longer the case for [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + : random sampling gives errors of 0.84 ± 0.33 0.84 \pm0.33 0.84 ± 0.33 and 0.35 ± 0.08 0.35 \pm0.08 0.35 ± 0.08 kcal mol− 1 ^{-1} − 1 for the singlet and triplet, against 0.27 ± 0.03 0.27 \pm0.03 0.27 ± 0.03 and 0.17 ± 0.05 0.17 \pm0.05 0.17 ± 0.05 kcal mol− 1 ^{-1} − 1 on the hardware, and the random gap (− 4.96 ± 0.33 -4.96 \pm0.33 − 4.96 ± 0.33 kcal mol− 1 ^{-1} − 1 ) deviates from CASCI by 0.5 kcal mol− 1 ^{-1} − 1 . For the more weakly correlated [ C o ( H 2 O ) 5 ( C O 2 ) ] 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 3 + , random bitstrings remain comparable to the hardware at 32 qubits (0.11 ± \pm ± 0.02 and 0.13 ± \pm ± 0.16 kcal mol− 1 ^{-1} − 1 , against 0.16 ± \pm ± 0.02 and 0.12 ± \pm ± 0.09 kcal mol− 1 ^{-1} − 1 ). The hardware samples are also more often physically valid: at 32 qubits 2.4–5.4% of them carry the correct particle numbers, against 0.4–3.9% of random bitstrings.
Figure 20. Spin-state gaps from ibm-phoenix for active spaces of 9 to 16 orbitals (18–32 qubits). [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^{3}\text{-}\mathrm{C}_{4}\mathrm{H}_{7})(\mathrm{NCCH}_{3})_{2}]^{+} [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4; a) and [ C o ( H 2 O ) 5 ( C O 2 ) ] 3 + [\mathrm{Co}(\mathrm{H}_{2}\mathrm{O})_{5}(\mathrm{CO}_{2})]^{3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 3 + (8; b); both states of a complex use active spaces of the same size from the predicted ranking. Left: SQD error relative to CASCI for each state. Center: spin-state gap compared with CASCI. Right: gap with the NEVPT2 correction, compared with CASCI+NEVPT2 and UCCSD(T). Red: SqDRIFT on ibm-phoenix (L = 10 L = 10 L = 10 , 200 circuits per evolution time, 100 shots per circuit); gray: uniformly random bitstrings. Error bars: standard deviation over six SQD seeds or six random sets.
Figure 21 compares the sample sources for the most demanding case, the singlet of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3\text{-}\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) at 32 qubits, whose correlation extends into the Ni–allyl bond (§ 2.0.2 D). Noiseless LUCJ is the least accurate source (11.2 kcal mol− 1 ^{-1} − 1 ): its CCSD-parametrized state is dominated by the Hartree–Fock determinant, so the samples fill the occupied orbitals and leave the virtual orbitals empty (Figure 21 b), and the 40 000 samples contain only 74 distinct configurations. Uniformly random bitstrings have the opposite problem: every orbital is occupied with probability one half, most samples have the wrong particle numbers, and the error is 0.84 ± \pm ± 0.33 kcal mol− 1 ^{-1} − 1 . The hardware samples lie between these limits. Noise spreads them over many configurations but preserves the filled/empty pattern of the reference, with average occupations of about 0.5–0.8 for the occupied and 0.25–0.5 for the virtual orbitals. As a result, all three hardware sources reach chemical accuracy, and SqDRIFT gives the lowest error, 0.27 ± \pm ± 0.03 kcal mol− 1 ^{-1} − 1 , with the fewest two-qubit gates and the highest success probability (Table VIII ). LUCJ gives 0.35 ± \pm ± 0.05 kcal mol− 1 ^{-1} − 1 on ibm_fez but 0.66 ± \pm ± 0.24 kcal mol− 1 ^{-1} − 1 on ibm_phoenix, where its circuit runs with a ten times higher ESP: more of the phoenix samples reproduce the narrow noiseless distribution (35 642 distinct configurations among the 40 000 shots, against 39 999 on ibm_fez), and their occupations keep more of the Hartree–Fock pattern. For LUCJ in this case, the lower error rate of ibm_phoenix therefore did not give a better SQD energy.
Figure 21. Sample sources for the singlet of [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^{3}\text{-}\mathrm{C}_{4}\mathrm{H}_{7})(\mathrm{NCCH}_{3})_{2}]^{+} [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + (4) at 32 qubits, CAS(16,16). a, SQD error relative to CASCI with 40 000 samples from each source and identical SQD settings (logarithmic scale; mean and standard deviation over six SQD seeds or random sets). Boxes give the chip, the number of two-qubit gates (2q), the circuit depth, the estimated success probability (ESP) and the processor time; for SqDRIFT the mean over the 400 circuits. Dotted line: 1 kcal mol− 1 ^{-1} − 1 . b, Average occupation of each active spin orbital in the samples, with the orbitals sorted by natural occupation; the shaded region marks the orbitals occupied in the Hartree–Fock determinant.
This benefit of hardware noise refers to the systems and circuits studied here, and whether it persists for larger active spaces or other noise levels remains open.
These results carry the complete workflow, from a sin-
gle UHF calculation through the predicted active space to a quantum processor, up to 32 qubits for two open-shell transition-metal complexes, and recover spin-state gaps that agree with exact CASCI for the same active-space Hamiltonians to 0.10 0.10 0.10 kcal mol− 1 ^{-1} − 1 . The separate classical CASCI+NEVPT2 calculations give spin-state gaps that differ from full-molecule UCCSD(T) by 5.3 5.3 5.3 –6.3 6.3 6.3 kcal mol− 1 ^{-1} − 1 at 16 orbitals. This cross-method comparison remains subject to active-space and reference-method uncertainties. The role of the hardware is to provide samples that are diverse but structured. At 32 qubits this is the first size at which uniformly random sampling becomes clearly worse for the more strongly correlated nickel complex. All spaces studied here remain within the reach of exact diagonalization, so the calculations validate the workflow and do not indicate a quantum advantage. A 22-qubit LUCJ calculation for the three states of [Co(H2 _2 2 O)5 _5 5 (CO2 _2 2 )]2 + / 3 + ^{2+/3+} 2 + /3 + (8) on ibm_phoenix reproduces CASCI to within 0.02 0.02 0.02 kcal mol− 1 ^{-1} − 1 (Supplementary Section S7).
Scope and limitations
Five limitations follow from the results. First, the model compresses strong static correlation: entropies of about 0.6, as in the open-shell singlets of the vanadium(III) ions, are predicted at the correct orbitals but with much smaller magnitude. These states are rare in the training distribution, and they are recognized from their chemistry, low-spin states of metals absent from training, rather than from the UHF reference, which shows no significant instability. Second, in weakly correlated high-spin states the entropy profile is flat near any threshold, so the absolute predicted values are less reliable than the predicted ranking. In the quintet of 3, no orbital is predicted above 0.10, yet the four DMRG-active orbitals appear at ranks 1, 3, 5 and 6 of the prediction, and a rank-based or relative criterion would recover them. Third, the singly occupied d d d orbitals of high-spin states carry little entanglement in any entropy-based measure, DMRG or predicted, and are included by a chemical rule rather than by the model. Fourth, the workflow requires a spin-clean UHF reference. Several intermediate-spin states of the wider benchmark do not meet this criterion, so the most strongly spin-contaminated open-shell states are at present outside its scope. Fifth, the model works in a fixed window of 40 orbitals around the frontier gap. The boundary descriptors indicate coupling to orbitals outside the window but cannot recover them, so complexes with extended ligand systems may need a larger window, which we have not tested. All results refer to the def2-SVP basis, gas-phase species at fixed geometries, and elements up to Kr. The basis-set error could be reduced without enlarging the active space, and hence the qubit count, with the transcorrelated method, which builds short-range electron correlation into an effective Hamiltonian through a Jastrow factor [105 , 106 , 107 , 108 , 109 , 110 ] .
Conclusions
We have presented SpinAware, a model that predicts the orbital entanglement of transition-metal complexes (DMRG single-orbital entropies and orbital-pair mutual information) from a single UHF calculation, and we have used it to build active spaces for classical and quantum multireference calculations. The model combines spin-resolved descriptors that encode the spin state through α \alpha α –β \beta β occupation and energy differences, continuous metal–ligand composition in place of binary shell flags, and a transformer that scores each orbital in the context of all other orbitals of the window. Most of the gain over earlier models comes from the descriptors. Processing the orbitals jointly adds a smaller improvement, and the exchange information enters through the descriptors rather than through the attention bias.
On a benchmark of experimentally characterized complexes defined independently of the model, no systematic decrease in accuracy is observed across the compositional novelty tiers considered here: complexes that depart from the training data in metal, ligand, binding topology and oxidation state at once are among the best predicted. The residual errors are electronic rather than chemical in origin, and occur in two regimes: strong static correlation, whose magnitude is compressed, and weak correlation near the selection threshold, where a hard cut is ill-conditioned. A candidate pool of predicted entropies and mutual-information partners contains 99% of the DMRG-active orbitals of the eleven benchmark spin states (82 of 83) while holding, per state, 38% of the window, and 28 of the 29 active orbitals for the two application systems.
A NOON scan selects the final active space from this pool. Where the correlation extends into the metal–ligand bonds, the predicted ranking places the decisive ligand orbitals among its first picks. Where it is confined to the d d d shell, the predicted orbitals remain passengers and do not distort the result. From 13 to 16 orbitals, the corrected gaps change by 1.8 1.8 1.8 kcal mol− 1 ^{-1} − 1 for nickel and 2.8 2.8 2.8 kcal mol− 1 ^{-1} − 1 for cobalt. The selected spaces give spin-state gaps that, after classically evaluated NEVPT2 corrections, differ from full-molecule UCCSD(T) by 0.9 0.9 0.9 –6.3 6.3 6.3 kcal mol− 1 ^{-1} − 1 for active spaces of 9–16 orbitals, and at 16 orbitals the corrected gap still depends on the ligand orbitals at the boundary of the space. This comparison remains subject to residual active-space and reference-method uncertainties. On IBM quantum processors, SQD with samples from randomized time evolutions (Sq-DRIFT) reproduces exact CASCI energies for the same active-space Hamiltonians to within 0.3 0.3 0.3 kcal mol− 1 ^{-1} − 1 up to 32 qubits, where noisy hardware samples converge faster than noiseless ones and, for the nickel complex, outperform uniformly random bitstrings. Extending the training data to early transition metals and strongly correlated low-spin states, using the predicted ranking rather than absolute thresholds, and applying the workflow to active spaces beyond the reach of exact diagonalization are the natural next steps.
Beyond these steps, the prediction is cheap enough to serve as a component of larger automated workflows. High-throughput explorations of reaction mechanisms already launch multiconfigurational calculations automatically where coupled-cluster diagnostics indicate strong correlation [111 ] , and an end-to-end demonstration combining high-performance computing, entanglement-based active-space selection and logical qubits has shown how such a pipeline can deliver the active site of a catalytic reaction to a quantum computer [112 ] . In both cases the strongly correlated configurations are identified with correlated calculations. SpinAware could provide the same orbital-level information from the UHF calculation of each spin state, and thereby carry these pipelines to the spin-state chemistry of open-shell catalysts. Spin-forbidden reactions are a natural target: free-energy simulations along a spin-gap collective variable now reach crossing regions between surfaces of different multiplicity [113 ] , and spin-state-consistent active spaces could supply multireference reference points along such paths. A second direction is embedding. Active-space embedding treats a correlated fragment of a molecule or solid with a quantum solver in a classical environment [114 ] , and the orbital-embedding method eCAS assembles localized active spaces that can be solved independently, adding environment orbitals according to a mutual-correlation criterion [115 ] . Predicted orbital entropies and mutual information could supply such criteria at mean-field cost, once the training data are extended from isolated complexes to metal sites in larger environments, such as metalloenzymes and defects in solids. Finally, agentic frameworks for multireference calculations reduce errors relative to fixed protocols but vary widely in strategy and cost from run to run [46 ] . A fast, quantitative predictor of orbital entanglement could serve these agents as a tool that supplies a reproducible starting point for the active space, while AI-driven discovery of quantum algorithms [116 ] could then be applied to the Hamiltonians of the selected spaces.
Acknowledgments
S.K. acknowledges funding from the Quantum Science Center, a National Quantum Information Science Research Center of the U.S. Department of Energy (DOE), operated at Oak Ridge National Laboratory (ORNL). W.D. acknowledges funding from the German Federal Ministry of Research, Technology and Space (BMFTR) under the research program Quantensysteme and funding measure Quantum Futur 3 for project No. 13N17229, as well as from the Helmholtz Association via the Initiative and Networking Fund for project No. VH-NG-21-08 under the Helmholtz Investigator program and for projects KA-QUS-02 (qFLOW) and KA-QUS-03 (QT-Batt) under the Helmholtz Quantum Use case call. We thank the Scultetus Center at CASUS for funding the research stay of M.A.A.A., which enabled this work.
We gratefully acknowledge the computing time made available to us on the high-performance computers at the NHR Center Paderborn Center for Parallel Computing (PC2) (projects hpc-prf-qehpc and hpc-prf-nniap). PC2 is jointly supported by the Federal Ministry of Research, Technology and Space of Germany and the state governments participating in the NHR (www.nhr-verein.de/unsere-partner). We also acknowledge computing time on the Hazel cluster at North Carolina State University and access to the IBM Quantum systems ibm_phoenix and ibm_fez. The views expressed are those of the authors and do not reflect the official policy or position of IBM or the IBM Quantum team.
Author contributions. M.A.A.A. conceived the method, generated the training and benchmark data, developed and trained the models, performed the DMRG, NOON-scan, CASCI, NEVPT2 and quantum-hardware calculations, analyzed the results and wrote the original draft. W.D. and S.K. contributed equally: both supervised the project, acquired funding and contributed to writing the manuscript. W.D. additionally provided access to high-performance computing resources and S.K. to quantum-computing resources. T.D.K. acquired funding and contributed to writing the manuscript. All authors reviewed and approved the final manuscript.
Data availability. The trained models, the benchmark structures with their DMRG reference labels, the results of the NOON scans and of the CASCI, NEVPT2, DMRG and UCCSD(T) calculations, the active-space Hamiltonians, the raw measurement counts of all quantum-hardware runs and the training data set are available from the corresponding authors upon reasonable request.
Code availability. The SpinAware code for feature generation and prediction, together with the scripts for the NOON scans, the classical reference calculations and the SQD, SqDRIFT and LUCJ workflows, the five trained Model 1 and Model 2 pairs, their training scripts, the split by complex and a script that recomputes the benchmark tables from the released models,is available from the corresponding authors upon reasonable request.
supplementary.pdf
Benchmark complexes and reference calculations
Composition of the benchmark. The out-of-distribution benchmark comprises thirteen mononuclear complexes taken from experimental crystal structures in the Cambridge Structural Database (CSD). Seven of them (1–7), chosen to span the three difficulty tiers, are analyzed in detail in the main text; complex 4 is also used there as an application system. The remaining six (9–14) complement this set with further metals and coordination environments and are analyzed in § 3.4 and S5; their composition is given in Table S2 . Complex 8, [ C o ( H 2 O ) 5 ( C O 2 ) ] 2 + / 3 + [\mathrm{Co}(\mathrm{H}_2\mathrm{O})_5(\mathrm{CO}_2)]^{2+/3+} [ Co ( H 2 O ) 5 ( CO 2 ) ] 2 + /3 + , is not a crystal structure and is described in the main text (§ 2.0.2 C). The source crystal structures are listed in Table S1 . In every case the calculation was performed on the isolated metal-containing species; counter-ions and lattice solvent were omitted.
Table S2 : Complementary benchmark complexes 9–14. Atom counts refer to the isolated metal-containing species; q q q is the total charge. Novelty directions as in the main text (M, metal; L, ligand; T, binding topology; O, oxidation state).
Table S2.
Scope. The training set contains only elements up to Kr, for which the def2-SVP basis is all-electron. Elements beyond Kr are described with an effective core potential, which replaces the core electrons and thereby changes the one- and two-electron descriptors of the valence orbitals in a way that is not represented in the training data. Complexes containing such elements lie outside the scope of the present model, and all benchmark complexes were therefore restricted to elements up to Kr.
Reference wavefunctions. Because UHF solutions depend on the initial guess, the reference of every spin state was obtained from five starting points (restricted-open-shell or restricted starts from minimal-basis and Hückel guesses, and unrestricted minimal-basis, Hückel and atomic guesses), each followed through internal stability analysis. The lowest solution satisfying ∣ ⟨ S ^ 2 ⟩ − S ( S + 1 ) ∣ ≤ max [ 0.1 , 0.1 S ( S + 1 ) ] |\langle\hat{S}^{2}\rangle- S(S + 1)| \le\max[0.1, 0.1 S(S + 1)] ∣ ⟨ S ^ 2 ⟩ − S ( S + 1 ) ∣ ≤ max [ 0.1 , 0.1 S ( S + 1 )] was retained, and all DMRG labels and predictions refer to this solution. States for which no such solution exists are marked * and excluded from all averages. Closed-shell singlet solutions that satisfy this criterion but are internally unstable towards a broken-symmetry solution with larger spin contamination are retained; this is the case for the singlets of 4 and 8 (Section S6). Table S22 lists, for every benchmark state, the UHF energy of the reference used for the labels and predictions and the outcome of this screening. Where the screening found a lower spin-clean solution than the first calculation, the descriptors and the DMRG label were recomputed on it.
Table S22. UHF references of the benchmark states. E U H F E_{\mathrm{UHF}} E UHF of the reference used (in E h E_{\mathrm{h}} E h ) and the result of the stability screening: “confirmed” means that the screening returned the stored solution (within 0.1 mE h E_{\mathrm{h}} E h ); otherwise the lowering relative to the first calculation is given, and the labels refer to the lower solution.
Compound names.
1 hexachloridovanadate(III)
2 hexafluoridovanadate(III)
3 tetramethylcobaltate(III)
4 bis(acetonitrile-κ N \kappa N κ N )(η 3 \eta^{3} η 3 -syn-but-2-en-1-yl)nickel(II)
5 pentacarbonyl(morpholine-κ N \kappa N κ N )chromium(0)
6 tetracarbonyl(glycinato-κ 2 N , O \kappa^{2}N,O κ 2 N , O )chromate(0)
7 (η 6 \eta^{6} η 6 -benzene)bis(tetrachloridoaluminato-κ 2 C l , C l ′ \kappa^{2}Cl,Cl' κ 2 C l , C l ′ )titanium(II)
8 pentaaqua(carbon dioxide-κ O \kappa O κ O )cobalt(III) and cobalt(II)
Structures of complexes 1–7. Tier I. Complex 1 is a near-ideal octahedron (V–Cl 2.384–2.389 Å, Σ = 2.5 ∘ \Sigma= 2.5^\circ Σ = 2. 5 ∘ ) and differs from the training set only in its metal. Complex 2 contains the same d 2 d^{2} d 2 vanadium(III) center in a strongly distorted fluoride environment (V–F 1.899–2.018 Å, Σ = 29.0 ∘ \Sigma= 29.0^\circ Σ = 29. 0 ∘ , trans angles 173.3–175.4°), imposed in the crystal by N–H⋯F hydrogen bonding to the triammonium counter-ion. A second, more symmetric crystal structure of [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − (complex 10, § 3.4 ) gives the same picture. Complex 3 introduces σ \sigma σ -methyl carbon donors, which occur in only two training calculations (Co–C 1.951–2.011 Å). Its geometry lies far from a tetrahedron (τ 4 = 0.39 \tau_4 = 0.39 τ 4 = 0.39 ; largest angles 147.8 and 156.5°).
Tier II. Complex 4 is planar (sum of angles at Ni 359.9°) and contains an asymmetrically bound η 3 \eta^{3} η 3 -syn-crotyl ligand (Ni–C 1.854–2.133 Å, Ni–centroid 1.751 Å) trans to two acetonitrile donors
( N i − N 1.882 – 1.890 A ˚ , N − N i − N 95.6 ∘ ) (\mathrm{Ni-N}\ 1.882\text{--}1.890\ \text{\AA},\ \mathrm{N-Ni-N}\ 95.6^\circ) ( Ni − N 1.882 – 1.890 A ˚ , N − Ni − N 95. 6 ∘ ) . Complex 5 is a zero-valent chromium carbonyl bearing a single N-bound morpholine (Cr–N 2.196 Å, Σ = 23.3 ∘ \Sigma= 23.3^\circ Σ = 23. 3 ∘ ); the carbonyl trans to nitrogen is markedly shorter than the four cis carbonyls (Cr–C 1.820 versus 1.886–1.907 Å), consistent with enhanced back-donation to the CO opposite the weak π \pi π -acceptor. Complex 6 is an anionic chromium(0) carbonylate bearing an N,O-chelating glycinate. Its small bite angle (75.1°; Cr–N 2.150 Å, Cr–O 2.218 Å) makes it the most distorted octahedron in the set (Σ = 55.2 ∘ \Sigma= 55.2^\circ Σ = 55. 2 ∘ ), and the Cr–C distances span 1.731–1.891 Å. Because chelating ligands are common in training, its chelation does not count as a departure; its amino-acidate ligand (rare) and its zero-valent carbonyl chemistry (absent) do.
Tier III. Complex 7 departs from training in metal, ligand and topology simultaneously: an η 6 \eta^6 η 6 -benzene titanium(II) four-legged piano stool (Ti–centroid 2.069 Å, Ti–C 2.466–2.547 Å) with two κ 2 \kappa^2 κ 2 -bound tetrachloridoaluminate ligands (Ti–Cl 2.589–2.611 Å, Cl–Ti–Cl 77.7–78.1°, Ti···Al 3.40–3.44 Å). The coordinated ring shows pronounced C–C bond-length alternation (1.330–1.473 Å), consistent with substantial metal-to-arene back-donation.
Complementary complexes. Table S2 lists the composition of complexes 9–14 as used in the calculations. Complexes 10 and 13 are second crystal structures of the ions in 2 and 7, respectively. The novelty directions follow the definitions of the main text.
Training data
The training set comprises 3 454 spin-pure calculations (UHF/def2-SVP reference, DMRG with BLOCK2, bond dimension M = 1024 M = 1024 M = 1024 , 16 sweeps, 40-orbital window of 20 occupied and 20 virtual orbitals). Its composition is given in Table S3 . States whose UHF reference deviates from the exact ⟨ S ^ 2 ⟩ \langle\hat{S}^2\rangle ⟨ S ^ 2 ⟩ by more than 0.5 were excluded. The total spin of every DMRG label was verified with the exact ⟨ S ^ 2 ⟩ \langle\hat{S}^2\rangle ⟨ S ^ 2 ⟩ (including the overlaps between α \alpha α and β \beta β orbitals and the UHF core); 787 labels were obtained with an energy penalty on S ^ 2 − S ( S + 1 ) \hat{S}^2 - S(S+1) S ^ 2 − S ( S + 1 ) (Sec. IIC). The training references were not
Table S3 : Composition of the training set. Metals Cr, Mn, Fe, Co and Ni; spin multiplicities 1–6; total charges − 9 -9 − 9 to + 3 +3 + 3 , with all but nine calculations between − 5 -5 − 5 and + 3 +3 + 3 . Ligands: F − \mathrm{F}^{-} F − , C l − \mathrm{Cl}^{-} Cl − , B r − \mathrm{Br}^{-} Br − , O H − \mathrm{OH}^{-} OH − , S 2 − \mathrm{S}^{2-} S 2 − , N C S − \mathrm{NCS}^{-} NCS − , N 3 − \mathrm{N}_{3}^{-} N 3 − , H 2 O \mathrm{H}_{2}\mathrm{O} H 2 O , N H 3 \mathrm{NH}_{3} NH 3 , CO, and the ligands of the crystal-structure subset.
Table S3.
stability-screened: the model learns the map from a given UHF reference to the DMRG entropies in its orbitals, and descriptors and labels always refer to the same reference. All calculations use the ionic charge implied by the metal oxidation state and the formal ligand charges. None of the benchmark complexes of the main text or of Table S2 is part of the training data.
Figure S23 summarizes the training set and the in-distribution test. The 3 454 states are distributed over Cr (524), Mn (747), Fe (875), Co (727) and Ni (581) and over multiplicities 1–6 (595, 387, 813, 488, 808 and 363 states). The maximal single-orbital entropy per state has a median of 0.22; 87.5% of the states contain at least one orbital with s ( 1 ) > 0.10 s^{(1)} > 0.10 s ( 1 ) > 0.10 , and 18.9%, 8.0% and 4.0% reach 0.6, 0.8 and 1.0, up to the limit of ln 4 \ln4 ln 4 . In the crystal-structure subset the metal is bound to N, C, O, Cl, S, F, Br and P donors, with a few B, I, Se, Si and Sb contacts. On the test set (866 states of 402 complexes), the five-model ensemble of Model 1 reproduces the DMRG entropies with an MAE of 0.025 (0.080 on orbitals with s ( 1 ) > 0.10 s^{(1)} > 0.10 s ( 1 ) > 0.10 ), and the five-model ensemble of Model 2 reproduces log I i j \log I_{ij} log I ij with a correlation coefficient of 0.88 over 643 500 orbital pairs.
Figure S23. Training set and in-distribution test. (a)–(c) Numbers of states by metal, spin multiplicity and total charge (3 454 states). (d) Maximal single-orbital entropy per state; dotted line: selection threshold 0.10. (e) Predicted versus DMRG single-orbital entropies on the test set (866 states of 402 held-out complexes; mean of the five Model 1 runs). (f) Predicted versus DMRG mutual information on the same test complexes (643 500 orbital pairs; mean of the five Model 2 runs), on a logarithmic scale; values below 10 − 7 10^{-7} 1 0 − 7 are shown at 10 − 7 10^{-7} 1 0 − 7 . Large entropies are underestimated in (e), the compression of strong correlation discussed in the main text. Color: number of orbitals or pairs per bin (logarithmic).
Model, training and software settings
Exchange statistics in an idealized model
The way the exchange statistics of Eqs. (7) and (8) of the main text resolve the three archetypes of Figure 2 a–Figure 2 c of the main text can be seen in an idealized model. Let orbital A couple with strength K 0 K_{0} K 0 to m m m partners, let orbital B couple with the same strength K 0 K_{0} K 0 to a single partner, and let orbital C couple with a small strength δ ≪ K 0 \delta\ll K_{0} δ ≪ K 0 to all N − 1 N-1 N − 1 orbitals; all remaining couplings are neglected. The leading coupling is K 0 K_{0} K 0 for both A and B, so the sorted list sees no difference. The statistics of
Table S5: Software, quantum hardware and computing resources.
Table S5.
Eqs. (7) and (8) do (n i n_i n i for K 0 > κ > δ K_0 > \kappa> \delta K 0 > κ > δ ):
Table 2.
For a d d d orbital coupled to the four other d d d orbitals of its center, m = 4 m = 4 m = 4 : the total coupling, the partner count and the participation ratio are all four times larger than for the pair orbital, although their strongest couplings are equal. The relative spread s i / μ i s_i/\mu_i s i / μ i tells the same story from the other side. It is large for B, whose coupling is a single spike, and smaller for A, whose coupling is spread over several partners. The spectator C is separated from both by its small max j K i j \max_j K_{ij} max j K ij and S i S_i S i and by n i = 0 n_i = 0 n i = 0 . Its participation ratio is nevertheless the largest of the three, because weak coupling spread evenly is also “delocalized”. The shape descriptors C i C_i C i and P R i \mathrm{PR}_i PR i are therefore meaningful only together with the magnitudes S i S_i S i and max j K i j \max_j K_{ij} max j K ij , which is why both kinds of statistic are supplied to the model.
Prediction quality of complexes 9–14 and of spin-contaminated states
The prediction quality for complexes 1–7 is reported in Table III of the main text; Table S6 reports the same quantities for complexes 9–14, whose structures are shown in Figure S1 . Over the eight spin states of these complexes, the predicted active space agrees with DMRG with F 1 = 0.61 F_1 = 0.61 F 1 = 0.61 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.83 at τ = 0.07 \tau= 0.07 τ = 0.07 (MAE 0.019), close to the values for the eleven spin states of complexes 1–7 (0.65, 0.75 and 0.027). The behavior of the two groups is therefore consistent, and the two electronic regimes described in the main text appear in both.
Table S6 : Prediction quality for complexes 9–14. n D M R G n_{\mathrm{DMRG}} n DMRG and n M L n_{\mathrm{ML}} n ML are the numbers of orbitals with s 1 ≥ τ s_1 \ge\tau s 1 ≥ τ in the reference and predicted profiles. The state marked * has no spin-clean UHF reference and is excluded from the mean. The last row gives the septet of 5, which requires an electron outside the d d d shell and is reported here as a high-spin test case; it is not included in the mean.
Table S6. Prediction quality for complexes 9–14. n D M R G n_{\mathrm{DMRG}} n DMRG and n M L n_{\mathrm{ML}} n ML are the numbers of orbitals with s 1 ≥ τ s_1 \ge\tau s 1 ≥ τ in the reference and predicted profiles. The state marked * has no spin-clean UHF reference and is excluded from the mean. The last row gives the septet of 5, which requires an electron outside the d d d shell and is reported here as a high-spin test case; it is not included in the mean.
Figure S1. Structures of the additional benchmark complexes 9, 11, 12 and 14 , taken from experimental crystal structures (counter-ions and lattice solvent omitted). Complexes 10 and 13 are second crystal structures of 2 and 7 and are shown in Fig. 4 of the main text.
Spin states without a spin-clean reference. For three spin states of the main benchmark, the triplets of [ C o ( C H 3 ) 4 ] − [\mathrm{Co}(\mathrm{CH}_3)_4]^- [ Co ( CH 3 ) 4 ] − (3), [ C r ( C O ) 5 ( H N C 4 H 8 O ) ] [\mathrm{Cr}(\mathrm{CO})_5(\mathrm{HNC}_4\mathrm{H}_8\mathrm{O})] [ Cr ( CO ) 5 ( HNC 4 H 8 O )] (5) and [ C r ( C O ) 4 ( H 2 N C H 2 C O 2 ) ] − [\mathrm{Cr}(\mathrm{CO})_4(\mathrm{H}_2\mathrm{NCH}_2\mathrm{CO}_2)]^- [ Cr ( CO ) 4 ( H 2 NCH 2 CO 2 ) ] − (6), no UHF solution satisfying the spin-contamination criterion of the main text was found. A spin-contaminated reference mixes in higher spin states, so its orbitals, its exchange matrix and the DMRG entropies computed on them describe a mixture rather than the target state; neither the descriptors nor the reference labels are then meaningful, and these states are left out of every average in the main text. For completeness, Table S7 lists the results obtained on the lowest-energy contaminated solutions. The numbers are similar to those of the spin-clean states, but they are not used in any conclusion. For the triplet of 3 (⟨ S ^ 2 ⟩ U H F = 2.47 \langle\hat{S}^2\rangle_{\mathrm{UHF}} = 2.47 ⟨ S ^ 2 ⟩ UHF = 2.47 , above the limit of 2.2) a DMRG label exists and is included in the released data, but the state was excluded before the predictions were made and is therefore not evaluated.
Spin states excluded from the main-text statistics. Results on the lowest-energy, spin-contaminated UHF solutions. Columns as in Tables III and IV of the main text. The triplet of 3 was not evaluated(see text).
Table S7. Spin states excluded from the main-text statistics. Results on the lowest-energy, spin-contaminated UHF solutions. Columns as in Tables III and IV of the main text. The triplet of 3 was not evaluated(see text).
Weak correlation near the threshold. Every state with F 1 = 0 F_1 = 0 F 1 = 0 is a weakly correlated high-spin state, the triplets of 10 and 14. DMRG places no orbital above the threshold, and the prediction reproduces the flat profile with a small error (MAE 0.005–0.019) but places one to four orbitals marginally above 0.10. The quintet of 9 and the quartet of 11 behave in the same way, with errors of 0.006 and 0.018.
Strong correlation and metal identity. The singlets of the vanadium(III) ion 10 and of the nickel(II) ion 14 are strongly correlated (maximal DMRG s 1 = 0.65 s_1 = 0.65 s 1 = 0.65 and 0.46 0.46 0.46 ), and in both the correlated orbitals are located correctly (F 1 = 0.91 F_1 = 0.91 F 1 = 0.91 and 0.77 0.77 0.77 ). Their entropies, however, are compressed to different degrees: the largest predicted entropy is 0.18 for vanadium, which is absent from training, but 0.34 for nickel, which is part of the training set. Together with the vanadium singlets of 1 and 2, this shows that the compression of strong correlation discussed in the main text is mainly a property of metals absent from training.
Geometry and further metals. The two crystal structures of [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − , the distorted 2 and the more symmetric 10, give the same picture: a strongly correlated singlet and a weakly correlated triplet, with similar F 1 F_1 F 1 values, so the results do not depend on a particular geometry. The manganese complex 12, the only manganese complex of the benchmark, and the titanium complex 13 are predicted with F 1 = 0.89 F_1 = 0.89 F 1 = 0.89 and 0.95 0.95 0.95 . Complex 13, a second crystal structure of the titanium arene complex 7, confirms the accurate transfer to titanium found for 7 in the main text. The novelty directions and tiers of 9–14 are given in Table S2 .
The main text shows the entropy profiles and mutual information of 1, 3, 6 and 7 (Figure 8 –Figure 10 ) and of the application systems 4 and 8. Figure S2 –Figure S17 show the same analysis for the remaining benchmark complexes, one complex per page. In the entropy profiles, the gray dotted line marks the selection threshold τ = 0.10 \tau= 0.10 τ = 0.10 , the magenta dashed line the HOMO–LUMO boundary, and the letters above each panel the dominant atomic-orbital character (s, p, d) of each orbital.
Figure S17. Mutual information of 14, [ N i C l 4 ] 2 − [\mathrm{NiCl}_{4}]^{2-} [ NiCl 4 ] 2 − (VEBPUN). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
2, [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − (Figure S2 and Figure S3 ). This ion contains the same d 2 d^2 d 2 vanadium(III) center as 1 in a distorted fluoride environment. As for 1, the singlet is strongly correlated: DMRG assigns entropies of up to 0.66 to frontier orbitals with V d character, whereas the largest predicted entropy is 0.15. The correlated orbitals are nevertheless partly recovered (five predicted against six in DMRG; F 1 = 0.73 F_1 = 0.73 F 1 = 0.73 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.61 at τ = 0.07 \tau= 0.07 τ = 0.07 ; MAE 0.044), and the mutual information places the strongly correlated pairs correctly but with much smaller values. The triplet is weakly correlated: no DMRG entropy reaches 0.10, and the prediction reproduces the flat profile (MAE 0.022) but places five orbitals marginally above the threshold (F 1 = 0 F_1 = 0 F 1 = 0 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.27 at τ = 0.07 \tau= 0.07 τ = 0.07 ). Here the mutual-information rescue adds six orbitals, so that the candidate pool (16 orbitals) over-covers, which costs efficiency but not completeness.
5, [ C r ( C O ) 5 ( H N C 4 H 8 O ) ] [\mathrm{Cr}(\mathrm{CO})_5(\mathrm{HNC}_4\mathrm{H}_8\mathrm{O})] [ Cr ( CO ) 5 ( HNC 4 H 8 O )] (Figure S4 and Figure S5 ). The singlet is the ground state of this 18-electron chromium(0) carbonyl: it lies 50 kcal mol− 1 ^{-1} − 1 below the triplet at the UHF level and 72 kcal mol− 1 ^{-1} − 1 in the window DMRG (each state in its own orbitals). Its active space (11 orbitals at τ = 0.10 \tau= 0.10 τ = 0.10 , 19 at τ = 0.07 \tau= 0.07 τ = 0.07 ) is formed by the metal-d and carbonyl π ∗ \pi^* π ∗ orbitals on either side of the frontier gap and is reproduced closely (F 1 = 0.88 F_1 = 0.88 F 1 = 0.88 and 0.95; MAE 0.024); the candidate pool (20 orbitals) contains all eleven active orbitals. The quintet is predicted with F 1 = 0.82 F_1 = 0.82 F 1 = 0.82 and 0.79 (MAE 0.031); its pool (19 orbitals) contains ten of the eleven active orbitals. The septet requires an electron outside the d d d shell and is reported only here, as a high-spin test case (F 1 = 0.62 F_1 = 0.62 F 1 = 0.62 and 0.79; MAE 0.022; Table S6 ). The triplet, whose reference is spin-contaminated, is shown for completeness but excluded from all averages.
Figure S4. Figure S4: Single-orbital entropies of 5, [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)] (Tier II). Triplet (spin-contaminated reference, excluded from averages), quintet and septet. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
Figure S5. Figure S5: Mutual information of 5, [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)] (Tier II). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
9, [ C r C l 4 ] 2 − [\mathrm{CrCl}_4]^{2-} [ CrCl 4 ] 2 − (Figure S6 and Figure S7 ). The tetrahedral chromium(II) quintet is weakly correlated (maximal DMRG s 1 = 0.12 s_1 = 0.12 s 1 = 0.12 ). Two orbitals lie marginally above the threshold in DMRG and one in the prediction (F 1 = 0.67 F_1 = 0.67 F 1 = 0.67 ), and the agreement improves at the permissive threshold (F 1 = 1.00 F_1 = 1.00 F 1 = 1.00 ), with the smallest error of the benchmark (MAE 0.006).
Figure S6. Single-orbital entropies of 9 , [ C r C l 4 ] 2 − 9,[\mathrm{CrCl}_4]^{2-} 9 , [ CrCl 4 ] 2 − (ADUGIT). Quintet. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
Figure S7. Mutual information of 9 , [ C r C l 4 ] 2 − 9,[\mathrm{CrCl}_4]^{2-} 9 , [ CrCl 4 ] 2 − (ADUGIT). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
10, [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − , second structure (Figure S8 and Figure S9 ). The more symmetric structure of the vanadium(III) ion reproduces the behavior of 2. The singlet is strongly correlated (maximal DMRG s 1 = 0.65 s_1 = 0.65 s 1 = 0.65 ); all five active orbitals are recovered (F 1 = 0.91 F_1 = 0.91 F 1 = 0.91 and 0.88), with entropies compressed to at most 0.18. The triplet is weakly correlated (maximum 0.07) and belongs to the near-threshold regime (F 1 = 0 F_1 = 0 F 1 = 0 and 0.40; MAE 0.019).
Figure S8. Single-orbital entropies of 10, [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − , second crystal structure (HEG-NUE). Singlet and triplet. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07 0.07 0.07 .
Figure S9. Mutual information of 10, [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − , second crystal structure (HEG-NUE). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
11, [ C o B r 2 C l ( N C C H 3 ) ] − [\mathrm{CoBr}_2\mathrm{Cl}(\mathrm{NCCH}_3)]^{-} [ CoBr 2 Cl ( NCCH 3 ) ] − (Figure S10 and Figure S11 ). The cobalt(II) quartet is weakly correlated (maximal DMRG s 1 = 0.13 s_1 = 0.13 s 1 = 0.13 ). The prediction follows the flat profile (MAE 0.018) but places seven orbitals above the threshold against four in DMRG (F 1 = 0.73 F_1 = 0.73 F 1 = 0.73 and 0.71).
Figure S10. Single-orbital entropies of 11, [CoBr2 _2 2 Cl(NCCH3 _3 3 )]− ^- − (QOVWIH). Quartet. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
Figure S11. Mutual information of 11, [CoBr2 _2 2 Cl(NCCH3 _3 3 )]− ^- − (QOVWIH). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
12, [ M n B r ( C 13 H 10 N 2 O 4 ) ] [\mathrm{MnBr}(\mathrm{C}_{13}\mathrm{H}_{10}\mathrm{N}_2\mathrm{O}_4)] [ MnBr ( C 13 H 10 N 2 O 4 )] (Figure S12 and Figure S13 ). The manganese singlet has a large active space of 14 orbitals (maximal DMRG s 1 = 0.27 s_1 = 0.27 s 1 = 0.27 ), which the prediction reproduces closely (F 1 = 0.89 F_1 = 0.89 F 1 = 0.89 and 0.94; MAE 0.021). The triplet has no spin-clean UHF solution and is not shown.
Figure S12. Single-orbital entropies of 12, [MnBr(C13 _{13} 13 H10 _{10} 10 N2 _2 2 O4 _4 4 )] (SEDGAN). Singlet; the triplet has no spin-clean reference. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
Figure S13. Mutual information of 12, [MnBr(C13 _{13} 13 H10 _{10} 10 N2 _2 2 O4 _4 4 )] (SEDGAN). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
13, [ T i ( η 6 - C 6 H 6 ) ( A l C l 4 ) 2 ] [\mathrm{Ti}(\eta^6\text{-}\mathrm{C}_6\mathrm{H}_6)(\mathrm{AlCl}_4)_2] [ Ti ( η 6 - C 6 H 6 ) ( AlCl 4 ) 2 ] , second structure (Figure S14 and Figure S15 ). The titanium singlet (maximal DMRG s 1 = 0.38 s_1 = 0.38 s 1 = 0.38 ) is the most accurately predicted state of the benchmark at τ = 0.10 \tau= 0.10 τ = 0.10 : all nine DMRG-active orbitals are recovered with one additional orbital (F 1 = 0.95 F_1 = 0.95 F 1 = 0.95 ; MAE 0.029).
Figure S14. Single-orbital entropies of 13, [Ti(η 6 \eta^{6} η 6 -C6 _6 6 H6 _6 6 )(AlCl4 _4 4 )2 _2 2 ], second structure (SESDUP). Singlet. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
Figure S15. Mutual information of 13, [Ti(η 6 \eta^{6} η 6 -C6 _6 6 H6 _6 6 )(AlCl4 _4 4 )2 _2 2 ], second structure (SESDUP). DMRG (upper row) and predicted (lower row) mutual information on a logarithmic scale; red circles mark mutual-information rescue pairs (I > 0.005 I > 0.005 I > 0.005 ).
14, [ N i C l 4 ] 2 − [\mathrm{NiCl}_4]^{2-} [ NiCl 4 ] 2 − (Figure S16 and Figure S17 ). The tetrahedral nickel(II) singlet is strongly correlated (maximal DMRG s 1 = 0.46 s_1 = 0.46 s 1 = 0.46 ). Because nickel is part of the training set, its entropies are compressed far less than those of the vanadium singlets (maximal predicted s 1 = 0.34 s_1 = 0.34 s 1 = 0.34 ); five of the seven active orbitals are recovered (F 1 = 0.77 F_1 = 0.77 F 1 = 0.77 and 0.90; MAE 0.016), and the mutual-information rescue adds further partner orbitals. The triplet is weakly correlated (maximum 0.08) and belongs to the near-threshold regime (F 1 = 0 F_1 = 0 F 1 = 0 and 0.92; MAE 0.005).
Figure S16. Single-orbital entropies of 14, [ N i C l 4 ] 2 − [\mathrm{NiCl}_{4}]^{2-} [ NiCl 4 ] 2 − (VEBPUN). Singlet and triplet. DMRG (blue circles) and predicted (red squares) entropies over the 40-orbital window, with the numbers of active orbitals and F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 and 0.07.
Application systems: reference wavefunctions and energies
Table S8 lists the UHF references and the energies of the 11-orbital active spaces of the earlier scans (for the final-model spaces see Table S9 ). For both complexes, the closed-shell singlet UHF solution is internally unstable towards a broken-symmetry solution with large spin contamination, which the spin criterion of the main text rejects. For 4 the broken-symmetry solution lies 16.4 mE h E_\mathrm{h} E h (10.3 kcal mol− 1 ^{-1} − 1 ) below the closed-shell solution, with ⟨ S ^ 2 ⟩ = 0.737 \langle\hat{S}^2 \rangle= 0.737 ⟨ S ^ 2 ⟩ = 0.737 , which indicates substantial open-shell character of the singlet; the triplet nevertheless remains 12.4 kcal mol− 1 ^{-1} − 1 below even the broken-symmetry singlet at the UHF level. For 8 the lower solution has ⟨ S ^ 2 ⟩ = 1.15 \langle\hat{S}^2 \rangle= 1.15 ⟨ S ^ 2 ⟩ = 1.15 . All singlet results refer to the closed-shell solution. The UCCSD(T) T 1 T_1 T 1 diagnostic is 0.062 for the singlet of 4 and 0.033/0.030 for the singlet/quintet of 8.
Table S8: Reference energies of the earlier application spaces (def2-SVP, E h E_{\mathrm{h}} E h ). CASCI and NEVPT2 for the 11-orbital spaces built from UHF natural orbitals (the Hamiltonians used for SQD); “corr.” is the strongly contracted NEVPT2 correlation energy. Complex 8 at r ( C o − O C O 2 ) = 1.97 r(\mathrm{Co}-\mathrm{OCO}_{2}) = 1.97 r ( Co − OCO 2 ) = 1.97 Å.
Table S8.
Table S9: Energies of the final-model active spaces used in the main text (E h E_{\mathrm{h}} E h , def2-SVP). 11 orbitals: spaces of the 22-qubit hardware calculations; 16 orbitals: spaces of the 32-qubit calculations and of Table VII of the main text. CASCI exact; “corr.” is the strongly contracted NEVPT2 correlation energy from DMRG reduced density matrices.
Table S9.
Distinct accuracy comparisons. The SQD errors in Section S7 compare SQD with exact CASCI for the same active-space Hamiltonian and therefore quantify the solver error within a fixed space. The comparison of CASCI+NEVPT2 gaps with full-molecule UCCSD(T) in the main text instead tests agreement between different correlation treatments. The NEVPT2 corrections are evaluated classically. This agreement neither measures the SQD error nor establishes absolute gap accuracy or complete active-space convergence.
The archived UCCSD(T) total energies of 4 are − 1927.29704203 E h -1927.29704203\,E_\mathrm{h} − 1927.29704203 E h (singlet) and − 1927.23028785 E h -1927.23028785\,E_\mathrm{h} − 1927.23028785 E h (triplet). Expanded active-space energies and fixed-budget selection controls are given in § 3.8 . The full-molecule UCCSD(T) total energies of 8 (cobalt(III), 17 frozen core orbitals) are − 1949.18164270 E h -1949.18164270\,E_\mathrm{h} − 1949.18164270 E h (singlet) and − 1949.20528028 E h -1949.20528028\,E_\mathrm{h} − 1949.20528028 E h (quintet), with T 1 T_1 T 1 diagnostics of 0.033 and 0.030. The spaces above were built from UHF natural orbitals before the final models; they are kept for reference in the archived controls. All energies of the main text and the hardware calculations use the spaces of the final models, whose 11- and 16-orbital energies are listed in Table S9 ; all sizes from 9 to 20 orbitals are in Table S21 .
Table S21. Active-space energies of the final-model spaces (E h E_h E h ). n n n and N e N_e N e are spatial orbitals and active electrons. CASCI is exact up to 16 orbitals and DMRG-CASCI (bond dimension 1000) from 17 orbitals. NEVPT2 corrections are strongly contracted, from DMRG reduced density matrices; dashes: not evaluated. S, T, Q4 and Q: singlet, triplet, quartet and quintet.
Sample-based quantum diagonalization: details
SqDRIFT and LUCJ on ibm_phoenix and ibm_fez. Table S10 lists the SQD errors and spin gaps of Figure 20 for every active-space size. SqDRIFT circuits used random sequences of L = 10 L = 10 L = 10 Hamiltonian exponentials, 200 circuits per evolution time and 100 shots per circuit; SQD used three batches of 300 samples, with six SQD seeds (19–24) for the hardware samples and six independent sets of uniformly random bitstrings. A sweep of the sequence length (L = 5 L = 5 L = 5 , 10, 20 and 40) at 22 qubits is shown in Figure S18 . Each state was sampled at two evolution times, t = 1 t = 1 t = 1 and t = 10 t = 10 t = 10 (atomic units), with 200 circuits per time; the qDRIFT sequences were drawn with seed 19, so every hardware circuit has a noiseless twin. Circuits were compiled with the Qiskit preset pass manager (optimization level 3, transpiler seed 42) against the backend calibration at the time of preparation.
Table S10: SQD with SqDRIFT samples from ibm_phoenix and with uniformly random bitstrings (mean over six seeds or random sets; final-model spaces). Errors E S Q D − E C A S C I E_{\mathrm{SQD}} - E_{\mathrm{CASCI}} E SQD − E CASCI and gaps E ( l o w ) − E ( h i g h ) E(\mathrm{low}) - E(\mathrm{high}) E ( low ) − E ( high ) in kcal mol− 1 ^{-1} − 1 ; high spin is the triplet for 4 and the quintet for 8. “Valid” is the percentage of singlet samples with the correct particle numbers (hardware) and the corresponding fraction for uniformly random bitstrings.
Table S10.
Figure S18. Noiseless SqDRIFT at 22 qubits (11 orbitals): sequence length and number of circuits. Rows: 4 (singlet and triplet) and 8 (singlet and quintet). (a) Energy error relative to CASCI; (b) error of the spin gap; (c) CASCI weight of the configurations that were not sampled, versus the number of circuits per evolution time (100 shots each), for sequence lengths L = 5 L = 5 L = 5 , 10, 20 and 40. Dotted lines: 1 kcal mol− 1 ^{-1} − 1 .
Quantum processors. All hardware calculations used IBM superconducting processors with fixed-frequency transmon qubits and tunable couplers (Table S11 ). ibm_phoenix belongs to the
Table S11. Quantum processors. Median error rates and coherence times over all qubits and couplers, from the calibration stored with the compiled circuits.
Nighthawk family: 120 qubits on a 10 × 12 10 \times12 10 × 12 square lattice in which each qubit couples to up to four neighbors (218 couplers; Figure S19 ). ibm_fez is a Heron r2 processor with 156 qubits on a heavy-hexagon lattice, in which each qubit couples to at most three neighbors. The square lattice accommodates the two spin chains of the Jordan–Wigner ordering as a ladder—the α \alpha α and β \beta β spin orbitals of the same spatial orbital on neighboring qubits—so that the opposite-spin interactions of LUCJ act on the rungs without swap gates. The heavy-hexagon lattice lacks such ladders, which explains the larger two-qubit gate count of the same LUCJ circuit on The SqDRIFT circuits were compiled one by one, and the compiler chose a different placement for each, so the 400 circuits of one state together touch most of the qubits of ibm_phoenix (95 and 99 of the 120 qubits for the 32-qubit quintet and singlet of 8). A single layout therefore does not exist for these runs. The 32-qubit LUCJ circuit of 4 used physical qubits 64–66, 70–72, 74–76, 80–86, 90–96, 100–104 and 110–114 of ibm_phoenix, and qubits 87–93, 97, 98, 105–111, 117, 118, 123–129, 136, 137 and 143–147 of ibm_fez.
Figure S19. Qubit layout of the IBM Nighthawk processors. 120 qubits on a square lattice with nearest-neighbor tunable couplers. Coloured: the 22-qubit ladder of the LUCJ circuits for 8 on ibm_phoenix, with the α \alpha α spin orbitals on qubits 11, 21, … \ldots … , 111 and the β \beta β spin orbitals on qubits 10, 20, … \ldots … , 110; red rungs carry the opposite-spin interactions.
LUCJ calculation for 8 at 22 qubits. The three states of 8 were also computed with LUCJ circuits on ibm_phoenix, in the 11-orbital spaces of the predicted ranking, CAS(14,11), CAS(13,11) and CAS(12,11) for the singlet, quartet and quintet, with 100 000 shots per state. An earlier run on ibm_miami in spaces built from UHF natural orbitals is superseded by this calculation.
Circuits. The LUCJ circuits (one repetition) were parametrized from the CCSD t 1 t_1 t 1 and t 2 t_2 t 2 amplitudes of each active space by double factorization of t 2 t_2 t 2 , without variational optimization (ffsim UCJOpSpinBalanced for the singlet, UCJOpSpinUnbalanced for the quartet and quintet). In the local variant, same-spin Jastrow terms act between nearest neighbors along each spin chain and opposite-spin terms only on the ladder rungs; a full LUCJ variant with all pairs was run on the simulator as a test. The circuits were transpiled with the Qiskit preset pass manager at optimization level 3 together with the ffsim pre-initialization passes, on the ladder α \alpha α : [11, 21, … \ldots … , 111],
Table S13. Circuits for the singlet of 4 at 32 qubits, CAS(16,16). Two-qubit gate count (2q), depth and estimated success probability (ESP, the product of the gate and readout fidelities of the physical qubits used) after transpilation for the listed processor. SqDRIFT: two evolution times (t = 1 t = 1 t = 1 and 10) × \times × 200 random circuits × \times × 100 shots; mean over all 400 circuits (the means for t = 1 t = 1 t = 1 and t = 10 t = 10 t = 10 differ by at most 1 two-qubit gate, 14 layers and 0.002 in ESP).
β \beta β : [10, 20, … \ldots … , 110] of ibm_phoenix (summed CZ error of the layout 0.060), without dynamical decoupling or error suppression (Table S14). The transpilation added no swap gates and kept every qubit on the ladder.
Table S14. Executed LUCJ circuits for 8 (ibm-phoenix, 11 November 2026). ESP estimated with the calibration of Table S11.
SQD settings. Configuration recovery, subspace construction and diagonalization used qiskit-addon-sqd with the settings of Table S15 . The self-consistency loop was stopped when the energy changed by less than 10 − 5 E h 10^{-5}\ E_{\mathrm{h}} 1 0 − 5 E h and the average occupations by less than 10 − 7 10^{-7} 1 0 − 7 . Table S16 compares the different sources of samples.
Table S15. SQD settings for the sources of samples of 8.
Table S16. SQD error E S Q D − E C A S E_{\mathrm{SQD}}-E_{\mathrm{CAS}} E SQD − E CAS (kcal mol− 1 ^{-1} − 1 ) and sample statistics for the three states of 8 (final-model 11-orbital spaces, 10 5 10^{5} 1 0 5 samples per source). Chemical accuracy is 1 kcal mol− 1 ^{-1} − 1 = 1.6 mE h E_{\mathrm{h}} E h . Noiseless LUCJ was not repeated for these spaces.
Subsampling. To determine how many shots are needed, and whether the hardware samples differ from uniformly random bitstrings, random subsets of the measured bitstrings and the same numbers of random bitstrings were processed with the settings of Table S15 (Table S17 and Table S24 , Figure S20 ); no further quantum time was used. All runs with 10 3 10^{3} 1 0 3 to 10 5 10^{5} 1 0 5 shots are far within chemical accuracy (largest error 0.16 kcal mol− 1 ^{-1} − 1 , quartet at 10 3 10^{3} 1 0 3 shots). The hardware samples are 15–27 times more often physically valid than random bitstrings, but random bitstrings give larger subspaces and equal or smaller errors: in these 22-qubit spaces, configuration recovery turns 10 5 10^{5} 1 0 5 random bitstrings into the complete singlet space (108 900 configurations), so that SQD is exact. The phoenix samples are concentrated on few configurations (22–27% of all shots are the Hartree–Fock determinant), and small subsets therefore give small subspaces. At 50–100 shots, random samples sometimes contain no valid configuration (one of three runs for the singlet and quartet and two of three for the quintet at 50 shots; one of three for the quintet at 100 shots), so configuration recovery cannot start, whereas the hardware samples always can. Chemical accuracy in all seeds is reached from 500 (hardware) and 100 (random) shots for the singlet, from 200 and 100 shots for the quartet, and from 100 and 200 shots for the quintet; chemical accuracy is taken here as 1 kcal mol− 1 ^{-1} − 1 . With the full data the error of the quintet depends on the SQD seed (1.2 × 10 − 2 \times10^{-2} × 1 0 − 2 with the seed of Table S16 , 3 × 10 − 5 \times10^{-5} × 1 0 − 5 with another), which shows the scatter of SQD at this level.
Table S17. Subsampling of the measured data (ibm_phoenix, final-model 11-orbital spaces; range over two seeds, one set at 10 5 10^{5} 1 0 5 QPU shots). Error in kcal mol− 1 ^{-1} − 1 ; subspace dimension out of 108 900 (singlet), 76 230 (quartet) and 54 450 (quintet).
Figure S20. Shots needed for SQD of 8 (22 qubits, final-model 11-orbital spaces). Error of SQD relative to exact diagonalization versus the number of shots, for subsamples of the ibm_phoenix LUCJ samples (red) and for uniformly random bitstrings (gray). Symbols: median over the seeds (three below 10 3 10^{3} 1 0 3 shots, two above; one QPU set at 10 5 10^{5} 1 0 5 ); shaded: range. Dotted: 1 kcal mol− 1 ^{-1} − 1 . Runs in which no valid configuration was sampled are omitted (Table S24 ).
Table S24. SQD of 8 with few shots (22 qubits, ibm-phoenix LUCJ samples and uniformly random bitstrings, final-model 11-orbital spaces). Error relative to exact diagonalization in kcal mol− 1 ^{-1} − 1 for three subsampling seeds; n.v.: no valid configuration in the sample, so configuration recovery cannot start.
Archived fixed-reference and fixed-budget controls
Released-model inference and reference matching
These archived controls use the earlier released entropy and mutual-information checkpoints, without retraining, threshold optimization or fitting to benchmark labels. All eleven states of the benchmark at the time of these controls were reevaluated. Nine reproduce the printed active-orbital counts, F 1 F_1 F 1 scores and mean absolute errors to the stated precision. The released triplet references for complexes 1 and 2 differ from those underlying the main tables and are excluded from these paired controls rather than combined across references. The control cohort therefore contains nine states of six complexes and 83 DMRG-active orbitals. All selection methods within a comparison use the same 40-orbital reference window and its DMRG labels.
The stored Box–Cox exponent is 0.10910732265471183, with ϵ = 10 − 8 \epsilon= 10^{-8} ϵ = 1 0 − 8 . For Model 1, the stored affine min–max transformation acts on zero-based descriptor columns 0–3, 30, 31, 44–49, 58 and 59. The remaining columns are unchanged. The exchange matrix is divided by 2.6162941455841064 E h E_{\mathrm{h}} E h . These are the fitted transformations distributed with the checkpoints, not transformations refitted for this control. Inference used Python 3.13.5, PyTorch 2.5.1 (CPU), NumPy 2.4.6, SciPy 1.17.1 and scikit-learn 1.9.0. Figures used Matplotlib 3.11.2. These versions describe the independent inference control, not an inferred training environment.
Information gain at equal orbital budget
The entropy-only comparator selects the ∣ P ∣ |\mathcal{P}| ∣ P ∣ orbitals with the largest predicted single-orbital entropy in each state. It therefore has exactly the same state-specific orbital budget as the entropy-plus-MI pool. The pool recovers 79 of 83 DMRG-active orbitals (95.2%), compared with 77 (92.8%) for equal-budget entropy ranking and 76 (91.6%) for the fixed permissive entropy threshold of 0.07. The mean pool size is 16.9 spatial orbitals. The two equal-budget gains occur in the quintet of 5 (13 versus 12 recovered orbitals) and the singlet of 6 (18 versus 17). The quintet of 3 remains a failure case, with zero of four active orbitals recovered by either selection.
The paired increase in pooled recall is 2.4 percentage points. A cluster bootstrap resampling the six complexes, retaining all their spin states together, gives a 95% percentile interval of 0.0–4.1 percentage points (20 000 resamples, random seed 73291). This small cohort does not establish a universal or statistically resolved advantage. It does establish that the two additional recoveries are not caused by a larger orbital budget. They are additional recovered reference orbitals at unchanged pool size, not additional selected orbitals or a demonstrated reduction in qubit count.
For budget curves, orbitals above the permissive entropy threshold are ranked by predicted entropy, followed by MI-rescued orbitals ranked by their largest predicted coupling to a predicted active orbital. The remaining orbitals follow in entropy order. The frontier-energy comparator ranks the same window by distance from the midpoint of its central occupied and virtual orbital energies. Random selection has expected recall m / 40 m/40 m /40 at budget m m m . Orbital recall tests screening completeness, not spin-gap accuracy.
Final models. The numbers above refer to the released (paper) checkpoints and the original labels. With the final five-model ensembles and the spin-checked labels, over the eleven states of Table IV (83 DMRG-active orbitals), the pool recovers 82 orbitals with 15.2 orbitals per state, equal-budget entropy ranking also 82, and the fixed threshold of 0.07 alone 81 with 13.5 orbitals per state. For the application systems (29 active orbitals) the pool recovers 28, equal-budget entropy ranking 25 and the threshold of 0.07 alone 23. The quintet of 3 is no longer a failure case: all four of its active orbitals are in the pool. The released benchmark data have been regenerated from the final label files, so the triplet references of 1 and 2 now agree with the tables.
Figure S21. Fixed-reference screening controls for nine benchmark states. a, Pooled recall at the same spatial-orbital budget for entropy-plus-MI ranking, entropy ranking and frontier-energy ranking. The dashed line is the expectation for random selection. b, Recovered DMRG-active orbitals at the state-specific pool budget, identical for both selections. Black marks give the reference counts. Labels denote complex number and multiplicity. The empty bars for 3, quintet, retain the negative case.
Attention sensitivity is not a retraining ablation
The four fitted exchange-bias coefficients are − 0.1333 -0.1333 − 0.1333 , − 2.1200 -2.1200 − 2.1200 , − 2.9219 -2.9219 − 2.9219 and − 3.0672 -3.0672 − 3.0672 . In the five final models all four coefficients are negative as well, from − 0.37 -0.37 − 0.37 to − 0.81 -0.81 − 0.81 in the first and from − 2.0 -2.0 − 2.0 to − 2.8 -2.8 − 2.8 in the last layer. Thus, exchange bias does not necessarily increase attention between strongly exchange-coupled orbitals. Setting only the attention-bias matrix to zero at inference, while retaining all exchange-derived descriptors and the trained weights, changes the mean entropy error of the nine-state control cohort from 0.03589 to 0.03593. This small response is a sensitivity result for the fitted checkpoint. It does not remove exchange information from the input, and it does not establish the result of training a model without exchange bias. Retraining controls with and without the exchange bias are given in § 3.10 .
Equal-budget energetic controls for the nickel complex
For 4, three selection methods were compared for the earlier model spaces at 9, 11, 13 and 16 spatial orbitals (Figure S22 ). The singlet/triplet active-electron counts are 10/12, 14/16, 14/18 and 18/22, respectively, inherited from the released scans and identical for every method within a spin state. Geometry, def2-SVP basis, UHF spin state, inactive-electron count and strongly contracted NEVPT2 treatment were also matched. The comparison tests orbital selection at these fixed electron allocations, rather than optimization over all allocations. All calculations used PySCF 2.13.1.
Figure S22. Archived fixed-budget nickel selection controls. Singlet–triplet gaps for the earlier model spaces using ML-guided selection, metal-d d d plus frontier completion and the specified Ni 3d d d /ligand C,N 2p p p AVAS choice. a, CASCI. b, CASCI with strongly contracted NEVPT2. The dashed line is the full-molecule UCCSD(T) comparison, not an exact reference. All selections use the same orbital and active-electron budgets within each spin state.
The nine- and eleven-orbital CASCI problems were solved exactly. The nine-orbital ML spaces reconstruct the archived CASCI energies within 3.9 × 10 − 7 E h 3.9 \times10^{-7}\ E_{\mathrm{h}} 3.9 × 1 0 − 7 E h , with spin expectation values within 2 × 10 − 9 2 \times10^{-9} 2 × 1 0 − 9 of their targets.
The chemical comparator retains the same metal-d d d and fractionally occupied orbitals, then fills the remaining occupied and virtual slots in Fock-energy order towards the frontier. The AVAS comparator uses Ni 3 d 3d 3 d and ligand C/N 2 p 2p 2 p minimal-basis target functions. Occupied and virtual blocks are rotated separately to diagonalize the target projector, and the highest-overlap orbitals are retained with the same occupied, fractional and virtual counts as the ML space. This is one specified ligand-aware AVAS choice, not an optimization over AVAS targets.
The thirteen- and sixteen-orbital controls use spin-adapted DMRG with block2 0.5.4 and full one- through four-particle density matrices for strongly contracted NEVPT2. A fixed random seed of 4382 and a sweep schedule of 4 sweeps at bond dimension 250, 4 at 500 and up to 12 at 1000 were followed by a 6-sweep check at 2000. The allowed energy change was 10 − 6 E h 10^{-6}\ E_{\mathrm{h}} 1 0 − 6 E h . If the one-particle density changed by more than 10 − 4 10^{-4} 1 0 − 4 in Frobenius norm, the higher-bond state was refined until its energy and density changes fell below 10 − 8 E h 10^{-8}\ E_{\mathrm{h}} 1 0 − 8 E h and 10 − 5 10^{-5} 1 0 − 5 , respectively. Orbital reordering permutes the same selected space. Density-matrix electron traces were checked, and the active orbitals were kept unchanged during inactive/virtual semicanonicalization. Tests against exact eleven-orbital CASCI+NEVPT2 agree within 6.9 × 10 − 7 E h 6.9 \times10^{-7}\ E_{\mathrm{h}} 6.9 × 1 0 − 7 E h for both spin states. The sixteen-orbital ML corrected totals reproduce the archived values within 1.2 × 10 − 6 E h 1.2 \times10^{-6}\ E_{\mathrm{h}} 1.2 × 1 0 − 6 E h . Full state energies are listed below.
For these earlier spaces, chemical frontier completion is closer to UCCSD(T) at nine and eleven orbitals. ML is closest at thirteen orbitals, whereas ML and chemical selection are similarly close at sixteen. From thirteen to sixteen orbitals, the corrected gaps change by + 0.0512 +0.0512 + 0.0512 , + 3.1129 +3.1129 + 3.1129 and − 5.1918 -5.1918 − 5.1918 kcal mol− 1 ^{-1} − 1 for ML, chemical and AVAS selection, respectively. Within this archived comparison, the evidence supports local budget stability of the ML ranking, not universal superiority or a unique energetic optimum. Screening recovery, spin-gap stability and agreement with a comparison method remain distinct tests.
Occupation-based stability does not establish complete energetic convergence. The archived ML CASCI gaps at eighteen and twenty orbitals are − 3.5259 -3.5259 − 3.5259 and − 9.5058 -9.5058 − 9.5058 kcal mol− 1 ^{-1} − 1 , compared with approximately − 0.24 -0.24 − 0.24 and − 0.31 -0.31 − 0.31 kcal mol− 1 ^{-1} − 1 at thirteen and sixteen orbitals. Their NEVPT2 corrections are not available. Thus the thirteen-to-sixteen-orbital plateau is local to the reported budget range. UCCSD(T) is a cross-method comparison with its own uncertainty, particularly for the nickel singlet with a T1 diagnostic of 0.062. These controls concern one complex, fixed geometries and basis, one chemical rule and one specified AVAS target choice. They do not isolate the energetic contribution of ML or constitute retraining ablations. The ML rows of Table S18 and S19 use the earlier ML spaces, with 18 and 22 active electrons at 16 orbitals. The final-model spaces use 16 active electrons in both spin states at that size (Table S21 ). Their CASCI+NEVPT2 gaps are − 37.72 -37.72 − 37.72 , − 36.90 -36.90 − 36.90 , − 37.33 -37.33 − 37.33 and − 35.57 -35.57 − 35.57 kcal mol− 1 ^{-1} − 1 at 9, 11, 13 and 16 orbitals, respectively, with absolute differences from UCCSD(T) of 4.17, 4.99, 4.56 and 6.32 kcal mol− 1 ^{-1} − 1 . The change from 13 to 16 orbitals is + 1.76 +1.76 + 1.76 kcal mol− 1 ^{-1} − 1 . The chemical and AVAS controls retain the earlier electron allocations, so their comparison applies to the earlier model spaces.
Table S18. Archived nickel spin gaps for equal-budget selection controls. Δ E = E ( S ) − E ( T ) \Delta E = E(\mathrm{S}) - E(\mathrm{T}) Δ E = E ( S ) − E ( T ) in kcal mol− 1 ^{-1} − 1 . The full-molecule UCCSD(T) comparison is − 41.8889 -41.8889 − 41.8889 kcal mol− 1 ^{-1} − 1 . The last column is the absolute difference between the corrected gap and this comparison, not an error relative to an exact reference.
Expanded active-space energy record
The table above records the earlier active spaces. The final-model energies used in the main text are listed in Table S21 , including NEVPT2 corrections at 9, 11, 13 and 16 orbitals. Corrections beyond 16 orbitals are unavailable for both sets of spaces. For the final-model spaces, the cobalt corrected gap rises by 2.8 kcal mol− 1 ^{-1} − 1 from 13 to 16 orbitals and lies 5.3 kcal mol− 1 ^{-1} − 1 above UCCSD(T) at 16 orbitals.
Stability screening of the UHF references
Retraining controls of the exchange bias
Model 1 was retrained, on the same data, split and settings as the final models and with the same five initial-weight seeds, (a) with the exchange bias held at zero in all layers and (b) with the exchange matrix of every molecule permuted by a fixed random permutation of its orbitals, so that the couplings are assigned to the wrong orbital pairs while all descriptors are unchanged. Table S23 compares the five-model ensembles. Neither control reduces the accuracy on the test complexes, and the model without exchange bias also reaches the same accuracy on the benchmark. The exchange bias in attention therefore adds no measurable accuracy in the trained model; the exchange information reaches the model through the exchange descriptors of each orbital.
Table S23. Retraining controls of the exchange bias (five-model ensembles). Test: 866 states of 402 held-out complexes. Benchmark: eleven states as in Table III, evaluated before the singlet of 5 was added (septet of 5 included); F 1 F_1 F 1 is the mean over states. The permuted-exchange control was evaluated on the test set only.
Table S19. Archived state energies for the fixed-budget nickel controls in E h E_{\mathrm{h}} E h . The total energy is the sum of CASCI and the strongly contracted NEVPT2 correction. Chemical denotes metal-d d d plus frontier completion, and AVAS uses the Ni 3 d 3d 3 d /ligand C,N 2 p 2p 2 p target defined above.
Comparison with the model of Golub et al.
The code and trained models of Golub et al. (J. Chem. Theory Comput. 17, 6053 [43 ] are not available, and their model was trained on spin-restricted orbitals of the SC1MC set, with one spin state per complex. We therefore retrained its architecture on the data of this work: a feed-forward network with five hidden layers of 896 units (ReLU, dropout 0.1) that scores each orbital from its own features, trained with an L1 loss on Box–Cox and min–max scaled s ( 1 ) s^{(1)} s ( 1 ) (Adam, learning rate 10 − 3 10^{-3} 1 0 − 3 with cosine decay, batch size 1024, 2000 epochs, best validation epoch kept). As for Model 1, the output layer is linear during training and clamped at zero afterwards. Two feature sets were used: the 29 features of Golub et al. (descriptors 0–18 and 40–49 of Table I, with identical definitions) and all 60 descriptors of this work. Data, split and the five initial seeds are those of Model 1, and the five-model ensembles are compared in Table S25 . With Golub’s features, recall on the test complexes drops to 0.63 (SpinAware 0.80), so about a third of the active orbitals are missed; the largest benchmark losses occur for open-shell states, e.g. the triplet of 7 (F 1 = 0.20 F_1 = 0.20 F 1 = 0.20 , against 0.67 with the 60 descriptors). Replacing the features by the 60 descriptors recovers most of the difference, and the transformer of SpinAware adds a smaller further improvement, mainly at the threshold of 0.07.
Table S25. Comparison with the model of Golub et al. Five-model ensembles trained on the same data and split. Test: 866 states of 402 held-out complexes, metrics pooled over all orbitals. Benchmark: the eleven states of Table III; F 1 F_1 F 1 at τ = 0.10 \tau= 0.10 τ = 0.10 / 0.07.
Orbital-resolved single-orbital entropies
For each complex and spin state, the tables below list every orbital of the 40-orbital window with its dominant atomic-orbital character, orbital energy, occupation in the UHF reference, and the DMRG and predicted single-orbital entropies. Atomic-orbital character is given for the α \alpha α component of each
Table S20: Archived active-space energies for the two application complexes in E h E_{\mathrm{h}} E h . n n n and N e N_{\mathrm{e}} N e are spatial orbitals and active electrons. CASCI is exact except at 16 orbitals, where the listed energy is DMRG-CASCI with bond dimension 1000. Corrections are from the archived DMRG-based strongly contracted NEVPT2 calculations. Dashes mean that no correction was available and must not be interpreted as zero. S, T and Q denote singlet, triplet and quintet.
Table S20.
orbital from Löwdin populations. These tables are numbered O1–O23.
Table O1: Single-orbital entropies for complex 1, [ V C l 6 ] 3 − [\mathrm{VCl}_{6}]^{3-} [ VCl 6 ] 3 − , singlet (2 S + 1 = 1 2S + 1 = 1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 3698.904480 E h E_{\mathrm{UHF}} = -3698.904480\ E_{\mathrm{h}} E UHF = − 3698.904480 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^{2} \rangle= 0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_{1} \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O1.
Table O2: Single-orbital entropies for complex 1, [VCl6 _6 6 ]3 − ^{3-} 3 − , triplet (2 S + 1 = 3 2S + 1 = 3 2 S + 1 = 3 ): DMRG reference and SpinAware prediction. E U H F = − 3699.000000 E h E_{\mathrm{UHF}} = -3699.000000\ E_{\mathrm{h}} E UHF = − 3699.000000 E h ; ⟨ S ^ 2 ⟩ = 2.010 \langle\hat{S}^{2} \rangle= 2.010 ⟨ S ^ 2 ⟩ = 2.010 (exact 2.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O2.
Table O3: Single-orbital entropies for complex 2, [ V F 6 ] 3 − [\mathrm{VF}_{6}]^{3-} [ VF 6 ] 3 − , singlet (2 S + 1 = 1 2S+1=1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 1538.709354 E h E_{\mathrm{UHF}}=-1538.709354\ E_{\mathrm{h}} E UHF = − 1538.709354 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^{2}\rangle=0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_{1}\geq0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \geq0.08 ≥ 0.08 ).
Table O3.
Table O4. Single-orbital entropies for complex 2, [V F 6 \mathrm{VF}_6 VF 6 ]3 − ^{3-} 3 − , triplet (2 S + 1 = 3 2S + 1 = 3 2 S + 1 = 3 ): DMRG reference and SpinAware prediction. E U H F = − 1538.805934 E_{\mathrm{UHF}} = -1538.805934 E UHF = − 1538.805934 E h E_{\mathrm{h}} E h ; ⟨ S ^ 2 ⟩ = 2.004 \langle\hat{S}^2 \rangle= 2.004 ⟨ S ^ 2 ⟩ = 2.004 (exact 2.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O5. Single-orbital entropies for complex 3, [ C o ( C H 3 ) 4 ] − [\mathrm{Co}(\mathrm{CH}_3)_4]^- [ Co ( CH 3 ) 4 ] − , singlet (2 S + 1 = 1 2S+1=1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 1539.288110 E h E_{\mathrm{UHF}}=-1539.288110\ E_{\mathrm{h}} E UHF = − 1539.288110 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^2\rangle=0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O6. Single-orbital entropies for complex 3, [ C o ( C H 3 ) 4 ] − [\mathrm{Co}(\mathrm{CH}_{3})_{4}]^{-} [ Co ( CH 3 ) 4 ] − , quintet (2 S + 1 = 5 2S+1=5 2 S + 1 = 5 ): DMRG reference and SpinAware prediction. E U H F = − 1539.376178 E h E_{\mathrm{UHF}}=-1539.376178\ E_{\mathrm{h}} E UHF = − 1539.376178 E h ; ⟨ S ^ 2 ⟩ = 6.017 \langle\hat{S}^{2}\rangle=6.017 ⟨ S ^ 2 ⟩ = 6.017 (exact 6.000). Values of s 1 ≥ 0.10 s_{1}\geq0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \geq0.08 ≥ 0.08 ).
Table O7. Table O7: Single-orbital entropies for complex 4, [ N i ( η 3 − C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^3-\mathrm{C}_4\mathrm{H}_7)(\mathrm{NCCH}_3)_2]^+ [ Ni ( η 3 − C 4 H 7 ) ( NCCH 3 ) 2 ] + , singlet (2 S + 1 = 1 2S+1=1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 1925.292985 E_{\mathrm{UHF}}=-1925.292985 E UHF = − 1925.292985 E h E_{\mathrm{h}} E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^2\rangle=0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O8: Single-orbital entropies for complex 4, [ N i ( η 3 - C 4 H 7 ) ( N C C H 3 ) 2 ] + [\mathrm{Ni}(\eta^{3}\text{-}\mathrm{C}_{4}\mathrm{H}_{7})(\mathrm{NCCH}_{3})_{2}]^{+} [ Ni ( η 3 - C 4 H 7 ) ( NCCH 3 ) 2 ] + , triplet (2 S + 1 = 3 2S+1=3 2 S + 1 = 3 ): DMRG reference and SpinAware prediction. E U H F = − 1925.329155 E h E_{\mathrm{UHF}}=-1925.329155\ E_{\mathrm{h}} E UHF = − 1925.329155 E h ; ⟨ S ^ 2 ⟩ = 2.032 \langle\hat{S}^{2}\rangle=2.032 ⟨ S ^ 2 ⟩ = 2.032 (exact 2.000). Values of s 1 ≥ 0.10 s_{1}\geq0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \geq0.08 ≥ 0.08 ).
Table O8.
Table O9. Single-orbital entropies for complex 5, [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)], singlet (2 S + 1 = 1 2S + 1 = 1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 1892.070711 E h E_{\mathrm{UHF}} = -1892.070711 E_{\mathrm{h}} E UHF = − 1892.070711 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^2 \rangle= 0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 6.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O10. Single-orbital entropies for complex 5, [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)], triplet (2 S + 1 = 3 2S + 1 = 3 2 S + 1 = 3 ): DMRG reference and SpinAware prediction. E U H F = − 1891.990451 E h E_{\mathrm{UHF}} = -1891.990451\ E_{\mathrm{h}} E UHF = − 1891.990451 E h ; ⟨ S ^ 2 ⟩ = 2.276 \langle\hat{S}^2 \rangle= 2.276 ⟨ S ^ 2 ⟩ = 2.276 (exact 2.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ). The reference is spin-contaminated; this state is excluded from all averages.
Table O11: Single-orbital entropies for complex 5, [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)], quintet (2 S + 1 = 5 2S + 1 = 5 2 S + 1 = 5 ): DMRG reference and SpinAware prediction. E U H F = − 1891.945629 E h E_{\mathrm{UHF}} = -1891.945629\ E_{\mathrm{h}} E UHF = − 1891.945629 E h ; ⟨ S ^ 2 ⟩ = 6.462 \langle\hat{S}^{2} \rangle= 6.462 ⟨ S ^ 2 ⟩ = 6.462 (exact 6.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O11.
Table O12. Single-orbital entropies for complex 5, [Cr(CO)5 _5 5 (HNC4 _4 4 H8 _8 8 O)], septet (2 S + 1 = 7 2S + 1 = 7 2 S + 1 = 7 ): DMRG reference and SpinAware prediction. E U H F = − 1891.871014 E h E_{\mathrm{UHF}} = -1891.871014\,E_{\mathrm{h}} E UHF = − 1891.871014 E h ; ⟨ S ^ 2 ⟩ = 12.071 \langle\hat{S}^2 \rangle= 12.071 ⟨ S ^ 2 ⟩ = 12.071 (exact 12.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O13: Single-orbital entropies for complex 6, [ C r ( C O ) 4 ( H 2 N C H 2 C O 2 ) ] − [\mathrm{Cr}(\mathrm{CO})_4(\mathrm{H}_2\mathrm{NCH}_2\mathrm{CO}_2)]^- [ Cr ( CO ) 4 ( H 2 NCH 2 CO 2 ) ] − , singlet (2 S + 1 = 1 2S+1=1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 1775.684255 E h E_{\mathrm{UHF}}=-1775.684255\,E_{\mathrm{h}} E UHF = − 1775.684255 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^2\rangle=0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O13.
Table O14: Single-orbital entropies for complex 7, [ T i ( η 6 − C 6 H 6 ) ( A l C l 4 ) 2 ] [\mathrm{Ti}(\eta^{6}-\mathrm{C}_{6}\mathrm{H}_{6})(\mathrm{AlCl}_{4})_{2}] [ Ti ( η 6 − C 6 H 6 ) ( AlCl 4 ) 2 ] , singlet (2 S + 1 = 1 2S+1=1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 5238.034090 E h \mathrm{E}_{\mathrm{UHF}}=-5238.034090\ E_{\mathrm{h}} E UHF = − 5238.034090 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^{2}\rangle=0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_{1}\geq0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \geq0.08 ≥ 0.08 ).
Table O14.
Table O15. Single-orbital entropies for complex 7, [ T i ( η 6 − C 6 H 6 ) ( A l C l 4 ) 2 ] [\mathrm{Ti}(\eta^6-\mathrm{C}_6\mathrm{H}_6)(\mathrm{AlCl}_4)_2] [ Ti ( η 6 − C 6 H 6 ) ( AlCl 4 ) 2 ] , triplet (2 S + 1 = 3 2S+1=3 2 S + 1 = 3 ): DMRG reference and SpinAware prediction. E U H F = − 5238.101363 E h E_{\mathrm{UHF}}=-5238.101363\ E_{\mathrm{h}} E UHF = − 5238.101363 E h ; ⟨ S ^ 2 ⟩ = 2.139 \langle\hat{S}^2\rangle=2.139 ⟨ S ^ 2 ⟩ = 2.139 (exact 2.000). Values of s 1 ≥ 0.10 s_1\ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O16. Table O16: Single-orbital entropies for complex 9, [ C r C l 4 ] 2 − [\mathrm{CrCl}_4]^{2-} [ CrCl 4 ] 2 − , quintet (2 S + 1 = 5 2S + 1 = 5 2 S + 1 = 5 ): DMRG reference and SpinAware prediction. E U H F = − 2880.846959 E h E_{\mathrm{UHF}} = -2880.846959\ E_{\mathrm{h}} E UHF = − 2880.846959 E h ; ⟨ S ^ 2 ⟩ = 6.012 \langle\hat{S}^2 \rangle= 6.012 ⟨ S ^ 2 ⟩ = 6.012 (exact 6.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O17. Table O17: Single-orbital entropies for complex 10, [ V F 6 ] 3 − [\mathrm{VF}_{6}]^{3-} [ VF 6 ] 3 − , singlet (2 S + 1 = 1 2S+1=1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 1538.704175 E h E_{\mathrm{UHF}}=-1538.704175\ E_{\mathrm{h}} E UHF = − 1538.704175 E h ; ⟨ S ^ 2 ⟩ = − 0.000 \langle\hat{S}^{2}\rangle=-0.000 ⟨ S ^ 2 ⟩ = − 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_{1}\ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O18. Single-orbital entropies for complex 10, [ V F 6 ] 3 − [\mathrm{VF}_6]^{3-} [ VF 6 ] 3 − , triplet (2 S + 1 = 3 2S+1=3 2 S + 1 = 3 ): DMRG reference and SpinAware prediction. E U H F = − 1538.802165 E h E_{\mathrm{UHF} }=-1538.802165\ E_{\mathrm{h}} E UHF = − 1538.802165 E h ; ⟨ S ^ 2 ⟩ = 2.004 \langle\hat{S}^2\rangle=2.004 ⟨ S ^ 2 ⟩ = 2.004 (exact 2.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O19: Single-orbital entropies for complex 11, [ C o B r 2 C l ( N C C H 3 ) ] − [\mathrm{CoBr}_{2}\mathrm{Cl}(\mathrm{NCCH}_{3})]^{-} [ CoBr 2 Cl ( NCCH 3 ) ] − , quartet (2 S + 1 = 4 2S+1=4 2 S + 1 = 4 ): DMRG reference and SpinAware prediction. E U H F = − 7116.882956 E h E_{\mathrm{UHF}}=-7116.882956\,E_{\mathrm{h}} E UHF = − 7116.882956 E h ; ⟨ S ^ 2 ⟩ = 3.757 \langle\hat{S}^{2}\rangle=3.757 ⟨ S ^ 2 ⟩ = 3.757 (exact 3.750). Values of s 1 ≥ 0.10 s_{1}\geq0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \geq0.08 ≥ 0.08 ).
Table O19.
Table O20. Single-orbital entropies for complex 12, [ M n B r ( C 13 H 10 N 2 O 4 ) ] [\mathrm{MnBr}(\mathrm{C}_{13}\mathrm{H}_{10}\mathrm{N}_{2}\mathrm{O}_{4})] [ MnBr ( C 13 H 10 N 2 O 4 )] , singlet (2 S + 1 = 1 2S + 1 = 1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 4627.372240 E h E_{\mathrm{UHF}} = -4627.372240\ E_{\mathrm{h}} E UHF = − 4627.372240 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^{2} \rangle= 0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O21. Single-orbital entropies for complex 13, [Ti(η 6 \eta^6 η 6 -C6 _6 6 H6 _6 6 )(AlCl4 _4 4 )2 _2 2 ], singlet (2 S + 1 = 1 2S + 1 = 1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 5238.044961 E h E_{\mathrm{UHF}} = -5238.044961\ E_{\mathrm{h}} E UHF = − 5238.044961 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^2 \rangle= 0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O22. Table O22: Single-orbital entropies for complex 14, [ N i C l 4 ] 2 − [\mathrm{NiCl}_4]^{2-} [ NiCl 4 ] 2 − , singlet (2 S + 1 = 1 2S + 1 = 1 2 S + 1 = 1 ): DMRG reference and SpinAware prediction. E U H F = − 3344.166809 E h E_{\mathrm{UHF}} = -3344.166809\ E_{\mathrm{h}} E UHF = − 3344.166809 E h ; ⟨ S ^ 2 ⟩ = 0.000 \langle\hat{S}^2 \rangle= 0.000 ⟨ S ^ 2 ⟩ = 0.000 (exact 0.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).
Table O23. Single-orbital entropies for complex 14, [ N i C l 4 ] 2 − [\mathrm{NiCl}_4]^{2-} [ NiCl 4 ] 2 − , triplet (2 S + 1 = 3 2S+1=3 2 S + 1 = 3 ): DMRG reference and SpinAware prediction. E U H F = − 3344.290676 E h E_{\mathrm{UHF} }=-3344.290676\,E_{\mathrm{h}} E UHF = − 3344.290676 E h ; ⟨ S ^ 2 ⟩ = 2.004 \langle\hat{S}^{2}\rangle=2.004 ⟨ S ^ 2 ⟩ = 2.004 (exact 2.000). Values of s 1 ≥ 0.10 s_1 \ge0.10 s 1 ≥ 0.10 are in bold. Atomic-orbital character is given for the α \alpha α component of each orbital (Löwdin weights ≥ 0.08 \ge0.08 ≥ 0.08 ).