WorksPhysics and Astronomy
Spin-Adapted Neural-Network Backflow Wavefunctions for Excited States of Strongly Correlated Open-Shell Systems
Abstract
Building on the success of neural-network quantum states (NQS) for molecular ground states, multi-state frameworks such as natural excited states (NES) [Pfau et al., Science 2024, 385, eadn0137] extend NQS to chemically important excited states. However, without control over total spin, such calculations can require the inclusion of many intervening states of other spins, particularly in strongly correlated systems with dense low-energy spectra. We therefore generalize our previously developed spin-adapted neural network backflow (SA-NNBF) ansatz [Li et al., arXiv:2604.06841 (2026)] to improve its variational flexibility for excited-state calculations within a prescribed spin sector. We establish the universality of this extended ansatz within the chosen orbital space and spin sector, and use numerical benchmarks to assess how parameter allocation affects the ansatz's accuracy. With the spin-adapted ansatz, we are able to directly optimize the first excited state (a singlet) of without using any multi-state frameworks. Combining the extended ansatz with NES, we obtain low-lying triplet states of and singlet states of the cluster without including intervening states of other spins. Both the single-state and NES calculations achieve chemical accuracy relative to the exact energies. Our work provides a practical NQS framework for targeting excited states of prescribed spin in strongly correlated systems.
Introduction
Accurate calculations of molecular excited states are important for understanding electronic spectra and photochemical processes, but remain especially challenging in strongly correlated systems. Time-dependent density functional theory (TD-DFT)1 is most commonly used in practical applications for its low computational cost with reasonable accuracy for many excitations. However, the performance of a TD-DFT calculation depends on the choice of exchange–correlation functional and the character of the excitation [2, 3], which sometimes makes its results unreliable, especially for strongly correlated systems. Wavefunction methods, on the other hand, give overall more precise results with different ways to represent the many-body wavefunction, such as equation-of-motion coupled-cluster theory with single and double excitations (EOM-CCSD),4,5 selected configuration interaction (SCI)6–10 and density matrix renormalization group (DMRG)11,12 while these methods are usually more expensive and each of them has its specific limitations.[6, 7, 10, 12, 13, 14]
Recently, neural-network quantum states (NQS) have provided a new way to represent the wavefunction using the expressive power of neural networks.[15, 16, 17] In recent NQS studies, one of the most popular ansätze for fermions is the neural network backflow (NNBF), which uses determinants produced by a neural network to describe fermionic wavefunctions.[18, 19, 20] Related ansätze introduce additional degrees of freedom, as in the hidden fermion determinant state (HFDS).[21] Others replace the determinant with a Pfaffian, as in the hidden fermion Pfaffian state (HFPS).[22, 23] To solve for the ground state of a system, an NQS is usually optimized through variational Monte Carlo (VMC) where the energy and gradients are estimated by stochastic sampling and optimized by iteratively updating the trainable parameters in the wavefunction ansatz,[24] with a favorable low scaling with the system size while retaining a high accuracy for correlated states.
To extend NQS-VMC to excited states, a straightforward way is to optimize separate NQS sequentially or jointly with overlap penalties, but the penalty strengths require tuning and sampled overlaps can be difficult to estimate.[25, 26, 27, 28, 29, 30] The auxiliary wavefunction approach avoids penalty tuning by projecting out lower states, but still requires their prior optimization and overlap estimates.[31] To avoid explicit overlap estimates and penalty parameters, the natural excited states (NES) method recasts several low-lying states as a ground-state variational problem in an extended configuration space,[32, 33, 34] where the energies of individual states then follow from diagonalizing the sampled average of the matrix local energy.
Although the methods above provide access to low-lying excited states, solving for a specific target state with such methods requires including and optimizing all the states below the target one. This could make the calculation inefficient, especially for strongly correlated systems where many near-degenerate states with different spins intervene. For example, for the active space model of [FeS(SCH)] cluster (hereafter [2Fe-2S]),[35] the first excited singlet is the 29-th state in the spectrum.
The 28 lower-energy states, including the singlet ground state, must also be optimized to reach it with the methods above; see Figure 1 for an illustration. To reduce the number of states needed to be included, a well-established strategy which is already exploited in electronic structure methods, is to restrict the optimization in the target-spin sector to exclude the intervening states of other spins.[36, 37, 38, 39, 40, 41, 42] For NQS-VMC, a notable attempt is using spin penalties that bias excited-state NQS calculations toward a target total spin,[29] but the coefficients of the penalties require tuning and the purity of the target spin depends on the optimization.
!FCI excitation energies of [2Fe-2S] relative to its S = 0 ground state bbox=0.577,0.049,0.887,0.249
Figure 1. FCI excitation energies of [2Fe-2S] relative to its ground state. Each horizontal line represents a state in the space; colors indicate total spin, with in red. The left spectrum includes all spin sectors through the first excited singlet, while the right spectrum retains only the targeted sector. The labels and give energy-order indices in the full and sector, respectively.
Recently, inspired by the spin-adapted antisymmetrization method (SAAM) in real space,[43] we developed a spin-adapted NQS ansatz in occupied-number representation, namely the spin-adapted neural network backflow (SA-NNBF), which yields strictly spin-pure wavefunctions by construction.[44] In this work, we further extend SA-NNBF to a multi-determinant form and prove that an SA-NNBF is flexible enough to represent any state in the chosen orbital space and spin sector with an arbitrary fixed nonzero spin function. Moreover, we combine it with NES to optimize multiple low-lying states directly within fixed spin sectors, without any additional spin penalties. With spin-adaptation, the example of the 29-state calculation in [2Fe-2S] could be reduced to only 2 states as shown in Figure 1. We test such strategy on the example of O and [2Fe-2S], where several low-lying states of specific spins are obtained without additional states included in the NES calculation, and the energies of the resulting states are within chemical accuracy. These results demonstrate that the combination of SA-NNBF and NES provides an efficient tool to calculate excited states with high accuracy even for strongly correlated systems.
The remaining parts of this work are organized as follows. § 2 reviews variational Monte Carlo with neural quantum states, the natural excited states method, and the NNBF and SA-NNBF ansätze. § 3 explains how the spin function in SA-NNBF is chosen and constructed, and shows that the multi-determinant extension can represent any state in a chosen orbital space and spin sector. § 5 tests this extended ansatz on O and [2Fe-2S], including calculations of their low-lying excited states in chosen spin sectors. § 6 summarizes the findings.
Theory
Variational Monte Carlo (VMC) with neural quantum states (NQS)
For a ground-state calculation, the time-independent Schrödinger equation in the occupation-number representation is
Here, and are one-electron and antisymmetrized two-electron integrals in the spin-orbital basis, respectively, and is the fermionic annihilation (creation) operator for the -th spin-orbital. The NQS represents
where is referred to as a (single) wavefunction ansatz. Here, or , and is an occupation-number vector (ONV) labeling a basis state in . The number of spatial orbitals is ; the entries 0 and 1 denote an empty and an occupied spin orbital, respectively. For a single ansatz, the energy of can be estimated by averaging the local energy,
where , is the probability distribution, and is defined as .
For a real variational parameter , differentiating the energy expectation above gives[24]
In VMC, the expectation is estimated from sampled ONVs. In simple stochastic gradient descent, the parameters are then updated as , where is the learning rate. AdamW[45, 46] modifies this gradient-based update, while stochastic reconfiguration (SR)[47] and its minimum-norm formulation (minSR)[48, 49] precondition the gradient using the overlap matrix of wavefunction derivatives.
For the electronic Hamiltonian, the local energy sums over single and double excitations of . In this work, we use the semistochastic local energy evaluation developed in our previous work.[50] Specifically, a threshold decomposes the sum into a deterministic part and a stochastic part. The deterministic part sums the excitations with . The stochastic part samples configurations from , with . Then the estimator is
Its expectation equals Eq. (3), so the estimator is unbiased. Appropriate choices of and can reduce the computational cost dramatically.
Natural excited states (NES)
The NES method[32, 33, 34] transfers the problem of solving the lowest states of an -electron system into solving the ground state of an expanded system containing non-interacting replicas of the original system,
where is the Hamiltonian of the -th replica, and the wavefunction of the expanded system is defined as a determinant spanned by individual -electron wavefunctions over the electron configuration of the replicas,
where
Here with being the occupation number vector of the -th replica and each being a single -electron state.[32, 33] One could therefore minimize the NES energy,
where in the exact limit, the minimum would be the sum of the lowest eigenvalues of the original Hamiltonian, and the single-state functions would span the corresponding low-energy subspace. The individual variational energies and states are obtained by solving the generalized eigenvalue problem,
where , , and with being the energy of the -th state. The -th column of defines the corresponding state,
When solving the NES problem using the VMC framework, the single-state functions are parameterized and optimized, where the NES energy to be minimized is estimated through stochastic sampling by
where
is the matrix local energy with
and
In order to obtain the energy of each state after the variational optimization, one can estimate the matrix energy:
where it can be proved that , such as the eigenvalues of are the desired according to Eq. (9); see the Supporting Information for more details.
Wavefunction ansätze
Vanilla neural network backflow (NNBF)
To smoothly transfer to the SA-NNBF framework via NNBF, we start with reviewing the well-established generalized Hartree–Fock (GHF) ansatz. For an -electron system in occupation-number representation, a set of occupied GHF orbitals can be encoded by a coefficient matrix , where each column corresponds to an occupied GHF orbital represented as a combination of basis spin orbitals, containing both spin-up and spin-down components. Therefore, the transformed creation operators are
and the corresponding Slater determinant state is
and one can derive that the corresponding amplitude in the original basis can be written as
Here, the symbol represents the operation of selecting rows from the matrix on its right according to the occupied orbitals in .
The NNBF ansatz generalizes the GHF orbitals into backflow orbitals depending on the all-electron configuration, and further expands the ansatz from a single determinant into a combination of multiple determinants. It turns the constant matrix into a set of configuration-dependent matrices , generated by a neural network with as the input, which gives
with being the number of determinants.
In this work, we use a multilayer perceptron (MLP) to implement , which is specified by the number of layers and the number of hidden units . Accordingly, we denote an NNBF model as NNBF(, , ). Other architectures such as Transformer[51, 52] are also alternative, which are beyond the scope of this work.
Spin-adapted neural network backflow (SA-NNBF)
Here, we briefly introduce the SA-NNBF ansatz[44]. We consider a system with electrons in spatial orbitals, with total spin and spin projection . The electron spins can be coupled to form a spin function, denoted as , which satisfies
and can be expanded as a sum of products of one-electron spin functions,
Here, and denote the spin-up and spin-down one-electron spin eigenfunctions for the spin variable , respectively.[53] The matrix gives the coefficients of these one-electron spin functions in the -th product term, and denotes the number of product terms used to represent . In this work, we give a detailed discussion on the alternative choices of as well as the available implementations for Eq. (19) in § 3.
For a fixed spin function , the SA-NNBF ansatz uses a neural network to produce spatial-orbital transformation matrices, denoted as , with the spatial occupation number vector as the input. Here , where and denote the occupation numbers of the spin-up and spin-down orbitals associated with spatial orbital , respectively. With the spatial matrix and the spin function specified, the coefficient matrix of the -th product is combined with to form a spin-orbital coefficient matrix , whose components are
This matrix depends only on the spatial occupation , so configurations that differ only by a spin assignment share the same . The amplitude on a configuration is obtained by antisymmetrization: selects the rows of the spin orbitals occupied in , and the determinant of those rows gives
When Eq. (19) represents exactly, the state in Eq. (21) is an eigenfunction of both and for any spatial matrix . We denote this state by . As a function of the trainable matrix , the amplitude on configuration is written .
In analogy to NNBF and following Ref. 44, we simply choose an MLP to implement in SA-NNBF.
Spin functions in SA-NNBF
As introduced above, the construction of an SA-NNBF contains two main parts: the spin part and the spatial part. In this work, we give a brief discussion on the practical construction of the spin part, including the available choices of the spin eigenfunction and the implementations of the sum-of-product decomposition in Eq. (19). For the spatial part, we expand SA-NNBF with a combination of multiple sets of spatial orbitals to enhance its expressivity. With such expansion, we prove that given a sufficiently large number of spatial orbital sets, an SA-NNBF with arbitrary spin function is able to reproduce any spin eigenfunctions.
Choices of spin functions
Since the electrons are identical particles, for any given target and , there exist many ’s that satisfy Eq. (18), [54] where in SA-NNBF, the choice of may affect both the accuracy and efficiency in practical VMC calculations. Here, although a systematic numerical benchmark for all these choices is beyond the scope of this work, we briefly review some available methods to construct a spin eigenfunction.
The previous SA-NNBF work [44] constructed the spin functions through genealogical coupling, a systematic scheme where electrons are coupled one at a time and each electron raises or lowers the total spin by . For given , and , the spin functions generated by this method form a complete orthogonal basis for the corresponding -electron spin Hilbert subspace.[54, 55, 56] However, such scheme does not always offer an intuitive way to select one based on the chemical picture of a particular system, such as systems containing multiple high-spin transition-metal centers, where the spins of each cluster are coupled as a whole rather than one electron at a time.
The SAAM approach of Ref. 43 provides another way to construct a spin wavefunction, which is through a coupling tree where the target total spin is obtained by hierarchically coupling the high-spin leaves along a binary tree, with each leaf at the highest possible spin according to its contained electrons. This scheme enables one to choose a coupling tree using chemical intuition, such as for [2Fe-2S] where the chemical intuition suggests two local high-spin Fe(III) centers with each . In such system, we can use the hole-representation trick [44, 57] to simplify the (20o,30e) active space model into a 10-hole (10h) model. Following the two-center picture above, we divide the ten hole spins into two groups of five, couple each group to , and then couple the two local spins to the target total spin ; see Figure 2(a).
!Spin-coupling trees for (a) [2Fe-2S] in the ten-hole representation and (b,c) O2 in the 16-electron representation bbox=0.52,0.05,0.91,0.33
Figure 2. Spin-coupling trees for (a) [2Fe-2S] in the ten-hole representation and (b,c) O in the 16-electron representation. In (c), four singlet pairs are omitted from the drawing. The in (c) is calculated from Eq. (23).
It is worth noting that different methods above can sometimes construct the same spin function. A typical example of chemical interest is the valence bond (VB) spin function:[54]
where is the number of electron pairs. In genealogical coupling scheme, Eq. (22) corresponds to the coupling path along which the first electrons alternately raise and lower the intermediate total spin, while each remaining unpaired electron raises it.[54] Equivalently, Eq. (22) can also be represented by a coupling tree as the example of shown in Figure 2(c), which couples each electron pair into an node and successively couples them together with the unpaired electrons to the target spin.
Beside the spatial–spin structure, another alternative perspective to construct a pure-spin many-body wavefunction is through spin projection (SP), where a spin-projection operator is applied to a symmetry-broken state.[35, 54, 58, 59] Although the SP scheme does not explicitly introduce a spin eigenfunction, we can prove that such method is equivalent to the spatial–spin construction with an effective spin function corresponding to the simplest coupling tree; see the Supporting Information for more details.
Sum-of-product decomposition of spin functions
In SA-NNBF, the spin function needs to be encoded into a sum-of-product format as shown in Eq. (19) to apply antisymmetrization. However, in most cases, the construction methods introduced in § 3.1 do not directly give such format. Therefore, a practical problem is how to decompose a spin function into such sum-of-product format. Moreover, since the number of product terms could directly affect the computational cost, we desire to represent a spin function with as few terms as possible.
In principle, an optimal representation can be found through a projected CANDE-
COMP/PARAFAC (CP) decomposition.[44, 60] Nevertheless, since the spin wavefunction is a high-order tensor, a practical algorithm requires some special structures of the spin function to accelerate the evaluation of the loss function defined in the projected CP scheme. In Ref. 44, the spin functions are constructed by the genealogical coupling scheme, which could be written as a matrix product state (MPS), enabling efficient evaluation of the loss by tensor contraction.
For the spin functions with a simple binary tree structure, the original SAAM[43] provided an exact analytic decomposition, where the number of terms is at most
in which the coefficient of a few terms might be coincidentally zero. Here is the number of electrons in the -th leaf of the binary tree, and the product excludes the leaf containing the most electrons. Therefore, for a simple tree with only two leaves, such as Figure 2(a) and (b), is nearly half of the electron number, while for the tree in Figure 2(c), could scale exponentially. Since such decomposition is based on Fourier transformation, its resulting coefficients are mostly complex numbers. In this work, we improve this decomposition by replacing the complex-valued Fourier transform[43] with real-valued interpolation, thereby making the resulting coefficients real; see the Supporting Information for details. Although such improvement may still not reach the optimal decomposition from the perspective of minimizing , such scheme is efficient enough for the proof-of-concept calculations in this work.
An Example of a Spin Function
For molecular , although the tree in Figure 2(b) gives a much smaller than Figure 2(c), such coupling scheme does not seem to be consistent with the chemical intuition at a first glance, since the electrons in a molecule are usually believed to be mostly paired, rather than coupling into two high-spin clusters. However, SA-NNBF with a tree like Figure 2(b) can actually reproduce the paired-electron picture with a properly permuted spatial part.
To simplify the algebra, we present the derivation for the triplet of a 6-electron system, while the results extend straightforwardly to other spin states as well as systems with more electrons, such as the 16-electron molecule. Moreover, we simplify the backflow spatial orbitals into fixed spatial orbitals, including two doubly occupied ones and two singly occupied ones:
where – are spatial orbitals in the spatial product . One could verify that the SA-NNBF with as the spatial part and the VB spin wavefunction in Eq. (22) as the spin part, is identical to a restricted open-shell Hartree–Fock (ROHF) determinant up to normalization:
where
Such result is expected since the VB spin function and ROHF determinant correspond to a same paired-electron feature.
On the other hand, for the same system, all the possible 2-leaf simple trees are shown in Figure 3, where the corresponding spin function are noted as and . The two trees give the following spin functions.
Figure 3. All the possible 2-leaf coupling tree with six electrons in the sector, giving the spin functions and , respectively.
Here, and denote the one-electron spin functions on electron . Although Figure 3 seems to be different with the paired-electron picture, we can obtain exactly the same ROHF determinant by using a permuted spatial orbital set:
resulting in,
In practice, since is obtained by optimization, Eq. (24) and Eq. (25) show that SA-NNBF is always able to reproduce the paired-electron picture since the ROHF state in within the representable space, regardless of whether , or is chosen as the spin function.
The results above in the 6-electron example could be straightforwardly extend to general systems, which makes it in general reasonable for us to use a 2-leaf simple coupling tree even in closed-shell systems. Therefore, in this work, we use the tree in Figure 2(b) rather than (c), in order that the spin function can be decomposed into less terms while retaining the paired-electron chemical picture.
SA-NNBF with multiple spatial-orbital sets
Our previous work[61] proved that given a sufficiently large hidden length , vanilla NNBF Eq. (17) is universal with a single set of backflow orbitals, i.e. the NNBF is able to exactly represent an arbitrary target state. However, SA-NNBF Eq. (21) does not have such property in general.
To illustrate a potential limitation of the SA-NNBF ansatz with only a single set of backflow orbitals, one could focus on the subspace of the configurations corresponding to a given spatial , where no matter how complicated is the neural network architecture, the SA-NNBF on these configurations share a fixed same . Therefore, the SA-NNBF amplitudes within such subspace are completely determined by and matrix according to Eq. (20) and Eq. (21):
To reproduce the exact wavefunction , the SA-NNBF needs to satisfy the following set of equations:
The equations in Eq. (26) are indexed by the configurations whose spatial occupation is the chosen . Because is a spin eigenstate, the right-hand sides for this set of are linearly related. The number of independent equations in the set equals the dimension of the spin subspace formed by the singly occupied orbitals in , and this dimension still grows combinatorially with the number of singly occupied orbitals.[54] Every equation in the set uses the entries of one matrix, . When places electrons in distinct spatial orbitals, only the rows of those orbitals enter the determinants, so contributes at most unknown entries. For large , the number of independent equations exceeds . Eq. (26) then has no solution for a general target state, and an SA-NNBF with a single spatial-orbital matrix cannot reproduce every state in the chosen sector.
To overcome the limitation above, in this work, we expand the SA-NNBF with multiple sets of spatial orbitals in analogy to the form of Eq. (17):
where
with produced by the neural network. Following the notation for NNBF, we denote an SA-NNBF model configuration by SANNBF().
The expansion in Section 4 enables SA-NNBF to restore universality. In particular, we prove the following:
Theorem 1 (Universality of SA-NNBF for fixed spin function). For any FCI state in a fixed sector and any nonzero spin function , an SA-NNBF state with the fixed spin function can exactly represent when is sufficiently large.
A detailed proof is given in the Supporting Information. Theorem 1 also shows that the universality of SA-NNBF does not depend on the choice of , although in a practical calculation with a finite , some specific choices may achieve higher accuracy or make the corresponding fixed-spin-function ansatz easier to optimize.
Results
In all calculations reported here, both NNBF and SA-NNBF use . For a given system, the number of trainable parameters is then approximately proportional to . At the same and , NNBF has roughly twice as many trainable parameters as SA-NNBF since its backflow orbitals are spin orbitals rather than spatial orbitals. Increasing may provide another route to higher accuracy; its effect remains to be investigated systematically in future. Computational details are provided in the Supporting Information.
Benchmarks of SA-NNBF
To investigate the performance of SA-NNBF, we test different allocations of between and and scaling strategies for SA-NNBF and compare them with NNBF. We start with the triplet ground state of molecular O, as a simple case to benchmark different strategies. Figure 4(a) examines how the accuracy is affected as increases while decreases to keep approximately within a constant level. For both NNBF and SA-NNBF, most of the curves have local minima rather than being entirely monotonic. This suggests an optimal parameter allocation strategy at each level, striking a balance between and . However, the preferred strategies for the two ansätze differ substantially: the optimal values for SA-NNBF, typically , are consistently larger than those for NNBF, which are 1 or 2.
Figure 4. Energy errors of NNBF (dashed lines with open circles) and SA-NNBF (solid lines with filled squares) with different or relative to FCI for the triplet ground state of O in the 6-31G basis with an active space. For each vertically aligned pair of data points with the same color, there is in order that their parameter counts are comparable. (a) Errors versus different parameter allocations. Results of each color correspond approximately to the same parameter count, with , and at , 20000 and 40000, respectively. For the same color, is twice . (b) Errors versus hidden width at different fixed ’s where the lower and upper axes correspond to the hidden width for NNBF () and SA-NNBF (), respectively. (c) Errors versus at different fixed NNBF hidden widths.
We further compare two different strategies of systematically scaling the model: increasing at fixed in Figure 4(b), and increasing at fixed in Figure 4(c). In both panels, the energy error for each curve shows an overall decrease. Within each pair of curves with the same color, the SA-NNBF curve (solid) shows a gentler slope and ends at a larger error than the NNBF curve (dashed) in panel (b). (c) shows the opposite pattern, where the SA-NNBF curve declines more steeply. Such a contrast shows a preference for increasing over in SA-NNBF when scaling the model.
To test SA-NNBF in more different spin sectors, we calculate the lowest states of [2Fe-2S] in the sectors respectively. Since the energies of these states increase with , we can directly use ground-state VMC to calculate the lowest state in each spin sector with both vanilla NNBF and SA-NNBF by simply setting to the target , without invoking NES. As shown in Figure 5, the SA-NNBF(1,128,1) failed to achieve chemical accuracy for half of the six states. By increasing to 2, resulting in SA-NNBF(1,128,2), SA-NNBF shows a substantial improvement, making its error consistently within 1 kcal/mol. This error is comparable with or even better than vanilla NNBF(1,128,1). The two models have similar parameter counts. By contrast, doubling to obtain SA-NNBF(1,256,1) yields a final accuracy between those of SA-NNBF(1,128,1) and SA-NNBF(1,128,2), with only a modest improvement over SA-NNBF(1,128,1). Such results are consistent with the preference observed in O.
!Optimization curves of the energy error for the lowest [2Fe-2S] states bbox=0.14,0.05,0.91,0.43
Figure 5. Optimization curves of the energy error relative to the corresponding FCI energy for the lowest [2Fe-2S] states with and , obtained with vanilla NNBF(1,128,1) and SA-NNBF with different or . The active space is (20o,30e). The curves are smoothed using a 500-step moving average. The shaded area represents the chemical accuracy ( kcal/mol).
In summary, the O and [2Fe-2S] benchmarks show that, for the expanded SA-NNBF in this work, increasing is preferable to increasing when systematically scaling the model to improve accuracy. We believe this preference, compared with NNBF, is related to the universality results in § 4: NNBF can achieve universality with either a sufficiently large or a sufficiently large , whereas SA-NNBF can do so only by increasing sufficiently, highlighting the more important role of in SA-NNBF.
Spin-Adapted Excited-State Calculations
The preceding benchmarks concern the lowest state in each spin sector. We now use SA-NNBF to target excited states relative to the overall ground state, first with single-state VMC and then with NES. For O, the lowest state is the first excited state of the molecule but the lowest state in the singlet sector. SA-NNBF can therefore optimize it with a single-state VMC calculation, without an NES expansion.
sion. As shown in Table 1, SA-NNBF(1,64,1) reaches an error of 1.46 mHa, within the conventional chemical-accuracy threshold of 1.59 mHa; increasing reduces the error to 0.46 mHa at . By contrast, a vanilla NNBF calculation constrained only by also includes the lower-energy component of the triplet ground state and therefore cannot isolate this singlet by energy minimization alone.
| model | energy (Ha) | error (mHa) | |
| (1,64,1) | 19 937 | 1.46 | |
| (1,64,2) | 38 657 | 0.94 | |
| (1,64,3) | 57 377 | 0.72 | |
| (1,64,4) | 76 097 | 0.56 | |
| (1,64,5) | 94 817 | 0.46 |
Table 1. Energies and absolute errors relative to FCI for the lowest singlet state of O from single-state SA-NNBF calculations.
We further combine SA-NNBF with NES to calculate the lowest four triplets of O, including the ground state. The four single ansätze use the same fixed spin function shown in Figure 2. As shown in Table 2, for each of the four triplets, SA-NNBF could achieve a high accuracy of 1 mHa with for each single-state ansatz. Meanwhile, with NNBF, one would need to include all the lowest 8-th states in an NES calculation to reach the fourth triplet, which would be far more expensive than the
| state | energy (Ha) | error (mHa) |
| X | 0.91 | |
| A | 0.55 | |
| A | 0.97 | |
| A | 0.64 |
Table 2. Energies and absolute errors relative to FCI for the four lowest triplet states of O from a 4-state SA-NNBF+NES calculation.
Table 1: Energies and absolute errors relative to FCI for the lowest singlet state of O from single-state SA-NNBF calculations. The spin function is chosen as the one represented by Figure 2(b) with .
4-state SA-NNBF calculation since constructing the NES wavefunction and local-energy matrices requires single-state evaluations per sampled configuration, in addition to the matrix-operation costs.
Table 2: Energies and absolute errors relative to FCI for the four lowest triplet states of O from a 4-state SA-NNBF+NES calculation with four separately parameterized SA-NNBF(1,64,3) single-state ansätze using the fixed spin function illustrated by Figure 2(b) with . The two A rows correspond to its different components obtained in NES calculations.
Furthermore, we consider the [2Fe-2S] cluster, which is a typical example of correlated systems including many near-degenerate energy levels of different spins. In seeking the three lowest singlet states , , and , there are over 30 higher-spin states lying between and , all of which need to be included in a conventional NES calculation with NNBF to reach , making the computation unacceptably expensive. However, as shown in Figure 6, with SA-NNBF, we are able to compute the lowest three singlets with only a 3-state NES calculation. Moreover, each of the three energy levels is optimized to chemical accuracy with a total parameter number of 240 267, which demonstrates the ability of SA-NNBF combined with NES to provide accurate state-resolved energies for the ground and excited states of challenging systems.
Figure 6. Optimization curves for the three lowest singlet states (, , ) of [2Fe-2S] from a 3-state SA-NNBF+NES calculation. The spin function is constructed using the same simple spin-coupling tree shown in Figure 2(a) with . The optimization curves are smoothed using a 500-step moving average. The dashed lines indicate the FCI energies of the target states. The inset shows the absolute error of the same smoothed energies, for each state in mHa; the shaded area represents chemical accuracy ( kcal/mol).
Conclusion
In this work, we give a detailed discussion on SA-NNBF, including the choice and construction of spin functions. To overcome its potential limitations, we extended SA-NNBF to include sets of spatial orbitals and proved its universality for sufficiently large . Moreover, we performed a relatively systematic comparison of NNBF and SA-NNBF across the parameter allocation for the triplet ground state of and the lowest states of [2Fe-2S] in different spin sectors. Within the tested range, increasing improved the accuracy of SA-NNBF more effectively than increasing alone, highlighting the central role of in SA-NNBF and providing guidance for future parameter choices.
Without any multi-state frameworks, we used SA-NNBF to calculate the lowest singlet of , which is its first excited state. Chemical accuracy was reached already at , and increasing further improved the accuracy. We further combined SA-NNBF with NES to calculate low-energy triplets of and low-energy singlets of [2Fe-2S], achieving chemical accuracy in both cases. Because the calculation is restricted to the target spin sector, it does not require the simultaneous calculation of many lower-energy intermediate states in other spin sectors. While maintaining spin purity and energy accuracy, such spin restriction greatly reduces the computational resources required to obtain excited states of the target spin, making otherwise costly calculations feasible.
Although this work uses only MLP-based neural networks, NQSs with other architectures, such as the Transformer,[52, 62, 63, 64] can also benefit from the spin-adaptation framework to accelerate the excited-state calculations by sufficiently reducing the involved states. For more complex architectures, the optimal parameter allocations also deserve investigation, and the benchmark in this work could serve as a primary reference. Besides, several avenues for improving the spin-adapted multi-state framework remain open, including selecting optimal spin functions for complex correlated systems, finding more compact spin-function decompositions, and adopting a multi-state formulation that scales more favorably with the number of states. These could improve the efficiency and accuracy of this framework and enable us to calculate larger systems and therefore warrant further exploration in the future.
Acknowledgement
This work was supported by the Quantum Science and Technology-National Science and Technology Major Project (2023ZD0300200) and the Fundamental Research Funds for the Central Universities.
Supporting Information Available
The Supporting Information lists the symbols used here and proves the NES matrix local-energy identity. It then develops a real-valued decomposition of binary-tree spin functions, constructs the spin function of the simplest coupling tree, and examines its connection to spin projection. A proof of SA-NNBF exactness for a fixed spin function is followed by computational settings, FCI reference energies, and additional numerical results for [2Fe-2S] and .
References
References
- [2]Dreuw, A.; Head-Gordon, M. Single-Reference Ab Initio Methods for the Calculation of Excited States of Large Molecules. Chem. Rev. 2005, 105, 4009–4037.pubmed.ncbi.nlm.nih.gov/16277369
- [3]Liang, J.; Feng, X.; Hait, D.; Head-Gordon, M. Revisiting the Performance of Time-Dependent Density Functional Theory for Electronic Excitations: Assessment of 43 Popular and Recently Developed Functionals from Rungs One to Four. J. Chem. Theory Comput. 2022, 18, 3460–3473.DOI
- [6]Tubman, N. M.; Freeman, C. D.; Levine, D. S.; Hait, D.; Head-Gordon, M.; Whaley, K. B. Modern Approaches to Exact Diagonalization and Selected Configuration Interaction with the Adaptive Sampling CI Method. J. Chem. Theory Comput. 2020, 16, 2139–2159.DOI
- [7]Holmes, A. A.; Tubman, N. M.; Umrigar, C. J. Heat-Bath Configuration Interaction: An Efficient Selected Configuration Interaction Algorithm Inspired by Heat-Bath Sampling. J. Chem. Theory Comput. 2016, 12, 3674–3680.DOI
- [10]Zhang, N.; Liu, W.; Hoffmann, M. R. Further Development of iCIPT2 for Strongly Correlated Electrons. J. Chem. Theory Comput. 2021, 17, 949–964.arxiv.org/abs/2011.09096
- [12]Baiardi, A.; Reiher, M. The Density Matrix Renormalization Group in Chemistry and Molecular Physics: Recent Developments and New Challenges. J. Chem. Phys. 2020, 152, 040903.DOI
- [13]Krylov, A. I. Equation-of-Motion Coupled-Cluster Methods for Open-Shell and Electronically Excited Species: The Hitchhiker’s Guide to Fock Space. Annu. Rev. Phys. Chem. 2008, 59, 433–462.DOI
- [14]Chan, G. K.-L.; Keselman, A.; Nakatani, N.; Li, Z.; White, S. R. Matrix Product Operators, Matrix Product States, and Ab Initio Density Matrix Renormalization Group Algorithms. J. Chem. Phys. 2016, 145, 014102.arxiv.org/abs/1605.02611
- [15]Carleo, G.; Troyer, M. Solving the Quantum Many-Body Problem with Artificial Neural Networks. Science 2017, 355, 602–606.DOI
- [16]Hibat-Allah, M.; Ganahl, M.; Hayward, L. E.; Melko, R. G.; Carrasquilla, J. Recurrent Neural Network Wave Functions. Phys. Rev. Res. 2020, 2, 023358.DOI
- [17]Wu, D.; Rossi, R.; Vicentini, F.; Carleo, G. From Tensor-Network Quantum States to Tensorial Recurrent Neural Networks. Phys. Rev. Res. 2023, 5, L032001.DOI
- [18]Luo, D.; Clark, B. K. Backflow Transformations via Neural Networks for Quantum Many-Body Wave Functions. Phys. Rev. Lett. 2019, 122, 226401.DOI
- [19]Liu, A.-J.; Clark, B. K. Neural Network Backflow for Ab Initio Quantum Chemistry. Phys. Rev. B 2024, 110, 115137.DOI
- [20]Liu, Z.; Clark, B. K. Unifying View of Fermionic Neural Network Quantum States: From Neural Network Backflow to Hidden Fermion Determinant States. Phys. Rev. B 2024, 110, 115124.DOI
- [21]Robledo Moreno, J.; Carleo, G.; Georges, A.; Stokes, J. Fermionic Wave Functions from Neural-Network Constrained Hidden States. Proc. Natl. Acad. Sci. U. S. A. 2022, 119, e2122059119.DOI
- [22]Chen, A.; Wan, Z.-Q.; Sengupta, A.; Georges, A.; Roth, C. Neural Network-Augmented Pfaffian Wave-Functions for Scalable Simulations of Interacting Fermions. Proc. Natl. Acad. Sci. U. S. A. 2026, 123, e2535288123.DOI
- [23]Roth, C.; Chen, A.; Sengupta, A.; Georges, A. Superconductivity in the Two-Dimensional Hubbard Model Revealed by Neural Quantum States. arXiv preprint arXiv:2511.07566 2025,arxiv.org/abs/2511.07566
- [24]Becca, F.; Sorella, S. Quantum Monte Carlo approaches for correlated systems; Cambridge University Press, 2017.DOI
- [25]Pathak, S.; Busemeyer, B.; Rodrigues, J. N. B.; Wagner, L. K. Excited States in Variational Monte Carlo Using a Penalty Method. J. Chem. Phys. 2021, 154, 034101.DOI
- [26]Entwistle, M. T.; Schätzle, Z.; Erdman, P. A.; Hermann, J.; Noé, F. Electronic Excited States in Deep Variational Monte Carlo. Nat. Commun. 2023, 14, 274.DOI
- [27]Liu, J.; Qin, C.; He, X.; Jiang, G. Calculate Electronic Excited States Using Neural Networks with Effective Core Potential. arXiv preprint arXiv:2312.15216 2023,arxiv.org/abs/2312.15216
- [28]Wheeler, W. A.; Kleiner, K. G.; Wagner, L. K. Ensemble Variational Monte Carlo for Optimization of Correlated Excited State Wave Functions. Electron. Struct. 2024, 6, 025001.arxiv.org/abs/2312.00693
- [29]Szabó, P. B.; Schätzle, Z.; Entwistle, M. T.; Noé, F. An Improved Penalty-Based Excited-State Variational Monte Carlo Approach with Deep-Learning Ansatz. J. Chem. Theory Comput. 2024, 20, 7922–7935.DOI
- [30]Gao, N.; Grutschus, T.; Noé, F.; Günnemann, S. Excited Pfaffians: Generalized Neural Wave Functions across Structure and State. Proceedings of the 43rd International Conference on Machine Learning (PMLR 306). 2026; pp 33184–33213.
- [31]Lu, Z.; Fu, W. Penalty and Auxiliary Wave Function Methods for Electronic Excitation in Neural Network Variational Monte Carlo. arXiv preprint arXiv:2311.17595 2023,arxiv.org/abs/2311.17595
- [32]Pfau, D.; Axelrod, S.; Sutterud, H.; Von Glehn, I.; Spencer, J. S. Accurate Computation of Quantum Excited States with Neural Networks. Science 2024, 385, eadn0137.DOI
- [33]Ma, Y.; Liu, C.; Li, W.; Zhang, S.-Y.; Duan, L.-M.; Wu, Y.; Deng, D.-L. Solving Excited States for Long-Range Interacting Trapped Ions with Neural Networks. Sci. Bull. 2026, 71, 3881–3887.arxiv.org/abs/2506.08594
- [34]Hendry, D.; Sinibaldi, A.; Carleo, G. Grassmann Variational Monte Carlo with Neural Wave Functions. Phys. Rev. B 2026, 114, 134404.arxiv.org/abs/2507.10287
- [35]Li, Z.; Chan, G. K.-L. Spin-Projected Matrix Product States: Versatile Tool for Strongly Correlated Systems. J. Chem. Theory Comput. 2017, 13, 2681–2695.DOI
- [36]Sharma, S.; Chan, G. K.-L. Spin-Adapted Density Matrix Renormalization Group Algorithms for Quantum Chemistry. J. Chem. Phys. 2012, 136, 124121.DOI
- [37]Wouters, S.; Poelmans, W.; Ayers, P. W.; Van Neck, D. CheMPS2: A Free Open-Source Spin-Adapted Implementation of the Density Matrix Renormalization Group for Ab Initio Quantum Chemistry. Comput. Phys. Commun. 2014, 185, 1501–1514.DOI
- [38]Sharma, S.; Sivalingam, K.; Neese, F.; Chan, G. K.-L. Low-Energy Spectrum of Iron–Sulfur Clusters Directly from Many-Particle Quantum Mechanics. Nat. Chem. 2014, 6, 927–933.DOI
- [39]Dobrautz, W.; Smart, S. D.; Alavi, A. Efficient Formulation of Full Configuration Interaction Quantum Monte Carlo in a Spin Eigenbasis via the Graphical Unitary Group Approach. J. Chem. Phys. 2019, 151, 094104.DOI
- [40]Fales, B. S.; Hohenstein, E. G.; Levine, B. G. Robust and Efficient Spin Purification for Determinantal Configuration Interaction. J. Chem. Theory Comput. 2017, 13, 4162–4172.DOI
- [41]Weser, O.; Liebermann, N.; Kats, D.; Alavi, A.; Li Manni, G. Spin Purification in Full-CI Quantum Monte Carlo via a First-Order Penalty Approach. J. Phys. Chem. A 2022, 126, 2050–2060.DOI
- [42]Mejuto-Zaera, C.; Tzeli, D.; Williams-Young, D.; Tubman, N. M.; Matoušek, M.; Brabec, J.; Weis, L.; Xantheas, S. S.; De Jong, W. A. The Effect of Geometry, Spin, and Orbital Optimization in Achieving Accurate, Correlated Results for Iron–Sulfur Cubanes. J. Chem. Theory Comput. 2022, 18, 687–702.
- [43]Li, R.; Liu, Y.; Jiang, D.; Chen, Y.; Wen, X.; Li, W.; He, D.; Wang, L.; Chen, J.; Ren, W. Spin-Adapted Neural Network Wavefunctions in Real Space. arXiv preprint arXiv:2511.01671 2025,arxiv.org/abs/2511.01671
- [44]Li, Y.; Wu, Z.; Zhang, B.; Fang, W.-H.; Li, Z. Spin-Adapted Neural Network Backflow for Strongly Correlated Electrons. arXiv preprint arXiv:2604.06841 2026,arxiv.org/abs/2604.06841
- [45]Kingma, D. P.; Ba, J. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations. 2015.arxiv.org/abs/1412.6980
- [46]Loshchilov, I.; Hutter, F. Decoupled Weight Decay Regularization. International Conference on Learning Representations. 2019.
- [47]Sorella, S. Green Function Monte Carlo with Stochastic Reconfiguration. Phys. Rev. Lett. 1998, 80, 4558–4561.
- [48]Rende, R.; Viteritti, L. L.; Bardone, L.; Becca, F.; Goldt, S. A Simple Linear Algebra Identity to Optimize Large-Scale Neural Network Quantum States. Commun. Phys. 2024, 7, 260.
- [49]Chen, A.; Heyl, M. Empowering Deep Neural Quantum States through Efficient Optimization. Nat. Phys. 2024, 20, 1476–1481.DOI
- [50]Wu, Z.; Zhang, B.; Fang, W.-H.; Li, Z. Hybrid Tensor Network and Neural Network Quantum States for Quantum Chemistry. J. Chem. Theory Comput. 2025, 21, 10252–10262.arxiv.org/abs/2507.19276
- [51]Ma, H.; Kan, B.; Shang, H.; Yang, J. Transformer-Based Neural Networks Backflow for Strongly Correlated Electronic Structure. arXiv preprint arXiv:2509.25720 2025,arxiv.org/abs/2509.25720
- [52]Gu, Y.; Li, W.; Lin, H.; Zhan, B.; Li, R.; Huang, Y.; He, D.; Wu, Y.; Xiang, T.; Qin, M.; Wang, L.; Lv, D. Solving the Hubbard Model with Neural Quantum States. Nat. Commun. 2026, 17, 7838.arxiv.org/abs/2507.02644
- [53]Helgaker, T.; Jørgensen, P.; Olsen, J. Molecular Electronic-Structure Theory; John Wiley & Sons: Chichester, England, 2000.
- [54]Pauncz, R. Spin Eigenfunctions; Plenum Press: New York, 1979.
- [55]Salmon, W. I. Advances in Quantum Chemistry Volume 8; Advances in Quantum Chemistry; Elsevier, 1974; Vol. 8; pp 37–94.
- [56]Ugandi, M.; Roemelt, M. A Recursive Formulation of One-Electron Coupling Coefficients for Spin-Adapted Configuration Interaction Calculations Featuring Many Unpaired Electrons. Int. J. Quantum Chem. 2023, 123, e27045.DOI
- [57]Li, Z.; Chan, G. K.-L. Hilbert Space Renormalization for the Many-Electron Problem. J. Chem. Phys. 2016, 144, 084103.DOI
- [58]Löwdin, P.-O. Quantum Theory of Many-Particle Systems. III. Extension of the Hartree-Fock Scheme to Include Degenerate Systems and Correlation Effects. Phys. Rev. 1955, 97, 1509–1520.DOI
- [59]Jiménez-Hoyos, C. A.; Henderson, T. M.; Tsuchimochi, T.; Scuseria, G. E. Projected Hartree-Fock Theory. J. Chem. Phys. 2012, 136, 164109.DOI
- [60]Kolda, T. G.; Bader, B. W. Tensor decompositions and applications. SIAM Rev. 2009, 51, 455–500.DOI
- [61]Li, Z.; Zhao, T.; Zhang, B. Representational Power of Selected Neural Network Quantum States in Second Quantization. J. Chem. Phys. 2026, 164, 084105.arxiv.org/abs/2511.04932
- [62]Viteritti, L. L.; Rende, R.; Becca, F. Transformer Variational Wave Functions for Frustrated Quantum Spin Systems. Phys. Rev. Lett. 2023, 130, 236401.DOI
- [63]Shang, H.; Guo, C.; Wu, Y.; Li, Z.; Yang, J. Solving the Many-Electron Schrödinger Equation with a Transformer-Based Framework. Nat. Commun. 2025, 16, 8464.DOI
- [64]Rende, R.; Nikolaenko, A.; Viteritti, L. L.; Sachdev, S.; Zhang, Y.-H. Transformer Neural-Network Quantum States for Lattice Models of Spins and Fermions: Application to the Ancilla Layer Model. PRX Intelligence 2026, 1, 023007.DOI


