betaFeedback

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 O2O_2 without using any multi-state frameworks. Combining the extended ansatz with NES, we obtain low-lying triplet states of O2O_2 and singlet states of the [Fe2S2(SCH3)4]2−[\mathrm{Fe}_2\mathrm{S}_2(\mathrm{SCH}_3)_4]^{2-} 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 [Fe2_2S2_2(SCH3_3)4_4]2−^{2-} cluster (hereafter [2Fe-2S]),[35] the first excited singlet is the 29-th state in the MS=0M_S = 0 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.

Figure 1. FCI excitation energies of [2Fe-2S] relative to its S=0S = 0 ground state. Each horizontal line represents a state in the MS=0M_S = 0 space; colors indicate total spin, with S=0S = 0 in red. The left spectrum includes all spin sectors through the first excited singlet, while the right spectrum retains only the targeted (S,MS)=(0,0)(S, M_S) = (0, 0) sector. The labels nrootn_{\mathrm{root}} and nroot′n'_{\mathrm{root}} give energy-order indices in the full and S=0S = 0 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 O2_2 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 O2_2 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

H^∣Ψ⟩=E∣Ψ⟩,\hat{H}|\Psi\rangle= E|\Psi\rangle,
H^=∑pqhpqa^p†a^q+14∑pqrs⟨pq∥rs⟩a^p†a^q†a^sa^r.(1)\hat{H} = \sum_{pq} h_{pq}\hat{a}_{p}^{\dagger}\hat{a}_{q} + \frac{1}{4}\sum_{pqrs}\langle pq\Vert rs\rangle\hat{a}_{p}^{\dagger}\hat{a}_{q}^{\dagger}\hat{a}_{s}\hat{a}_{r}. \tag*{(1)}

Here, hpqh_{pq} and ⟨pq∥rs⟩\langle pq\Vert rs\rangle are one-electron and antisymmetrized two-electron integrals in the spin-orbital basis, respectively, and a^p(†)\hat{a}_{p}^{(\dagger)} is the fermionic annihilation (creation) operator for the pp-th spin-orbital. The NQS represents ∣Ψ⟩|\Psi\rangle

∣Ψ⟩=∑nΨ(n)∣n⟩,(2)|\Psi\rangle= \sum_n \Psi(n)|n\rangle, \tag*{(2)}

where Ψ(n)=⟨n∣Ψ⟩:{0,1}2K→F\Psi(n)=\langle n|\Psi\rangle:\{0,1\}^{2K}\to\mathbb{F} is referred to as a (single) wavefunction ansatz. Here, F=R\mathbb{F}=\mathbb{R} or C\mathbb{C}, and n∈{0,1}2Kn\in\{0,1\}^{2K} is an occupation-number vector (ONV) labeling a basis state in SS. The number of spatial orbitals is KK; the entries 0 and 1 denote an empty and an occupied spin orbital, respectively. For a single ansatz, the energy of H^\hat{H} can be estimated by averaging the local energy,

E=⟨Ψ∣H^∣Ψ⟩⟨Ψ∣Ψ⟩=⟨Eloc(n)⟩n∼P(n),E=\frac{\langle\Psi|\hat{H}|\Psi\rangle}{\langle\Psi|\Psi\rangle}=\langle E_{\mathrm{loc}}(n)\rangle_{n\sim P(n)},
Eloc(n)=⟨n∣H^∣Ψ⟩⟨n∣Ψ⟩=∑mHnmΨ(m)Ψ(n).(3)E_{\mathrm{loc}}(n)=\frac{\langle n|\hat{H}|\Psi\rangle}{\langle n|\Psi\rangle}=\sum_m H_{nm}\frac{\Psi(m)}{\Psi(n)}. \tag*{(3)}

where Hnm≡⟨n∣H^∣m⟩H_{nm}\equiv\langle n|\hat{H}|m\rangle, P(n)=∣Ψ(n)∣2/∑n∣Ψ(n)∣2P(n)=|\Psi(n)|^2/\sum_n|\Psi(n)|^2 is the probability distribution, and ⟨f(n)⟩n∼P(n)\langle f(n)\rangle_{n\sim P(n)} is defined as ∑nP(n)f(n)\sum_nP(n)f(n).

For a real variational parameter θi\theta_i, differentiating the energy expectation above gives[24]

gi=∂E∂θi=2Re⁡⟨∂ln⁡Ψ∗(n)∂θi(Eloc(n)−E)⟩n∼P(n).(4)\begin{aligned} g_i=\frac{\partial E}{\partial\theta_i} \\ &=2\operatorname{Re}\left\langle\frac{\partial\ln\Psi^*(n)}{\partial\theta_i}\left(E_{\mathrm{loc}}(n)-E\right)\right\rangle_{n\sim P(n)}. \tag*{(4)} \end{aligned}

In VMC, the expectation is estimated from sampled ONVs. In simple stochastic gradient descent, the parameters are then updated as θi←θi−ηgi\theta_i\leftarrow\theta_i-\eta g_i, where η>0\eta>0 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 Eloc(n)E_{\mathrm{loc}}(n) sums over single and double excitations of nn. In this work, we use the semistochastic local energy evaluation developed in our previous work.[50] Specifically, a threshold ϵ>0\epsilon>0 decomposes the sum into a deterministic part and a stochastic part. The deterministic part sums the excitations with ∣Hnm∣≥ϵ|H_{nm}|\ge\epsilon. The stochastic part samples NϵN_{\epsilon} configurations m′m' from Pn(m′)∝∣Hnm′∣P_n(m') \propto|H_{nm'}|, with ∣Hnm′∣<ϵ|H_{nm'}| < \epsilon. Then the estimator is

Eloc(n;ϵ,Nϵ)=∑∣Hnm∣≥ϵHnmΨ(m)Ψ(n)+1Nϵ∑m′Hnm′Pn(m′)Ψ(m′)Ψ(n).(5)E_{\mathrm{loc}}(n;\epsilon,N_\epsilon)=\sum_{|H_{nm}|\ge\epsilon}H_{nm}\frac{\Psi(m)}{\Psi(n)}+\frac{1}{N_\epsilon}\sum_{m'}\frac{H_{nm'}}{P_n(m')}\frac{\Psi(m')}{\Psi(n)}. \tag*{(5)}

Its expectation equals Eq. (3), so the estimator is unbiased. Appropriate choices of ϵ\epsilon and NϵN_\epsilon can reduce the computational cost dramatically.

Natural excited states (NES)

The NES method[32, 33, 34] transfers the problem of solving the lowest MM states of an NN-electron system into solving the ground state of an expanded system containing MM non-interacting replicas of the original system,

H^=∑I=1MH^I,(6)\hat{H}=\sum_{I=1}^{M}\hat{H}_I, \tag*{(6)}

where H^I\hat{H}_I is the Hamiltonian of the II-th replica, and the wavefunction of the expanded system is defined as a determinant spanned by MM individual NN-electron wavefunctions over the electron configuration of the MM replicas,

ΨNES(nNES)=det⁡ΨNES(nNES),(7)\Psi_{\mathrm{NES}}(n_{\mathrm{NES}})=\det\mathbf{\Psi}_{\mathrm{NES}}(n_{\mathrm{NES}}), \tag*{(7)}

where

ΨNES(nNES)=[Ψ1(n1)⋯ΨM(n1)⋮⋱⋮Ψ1(nM)⋯ΨM(nM)].\mathbf{\Psi}_{\mathrm{NES}}(n_{\mathrm{NES}})= \begin{bmatrix} \Psi_1(n_1) & \cdots& \Psi_M(n_1) \\ \vdots& \ddots& \vdots\\ \Psi_1(n_M) & \cdots& \Psi_M(n_M) \end{bmatrix}.

Here nNES≡(n1,n2,…,nM)n_{\mathrm{NES}}\equiv(n_1,n_2,\ldots,n_M) with nIn_I being the occupation number vector of the II-th replica and each ∣ΨI⟩|\Psi_I\rangle being a single NN-electron state.[32, 33] One could therefore minimize the NES energy,

ENES=⟨ΨNES∣H^∣ΨNES⟩⟨ΨNES∣ΨNES⟩,(8)E_{\mathrm{NES}}=\frac{\langle\Psi_{\mathrm{NES}}|\hat{H}|\Psi_{\mathrm{NES}}\rangle}{\langle\Psi_{\mathrm{NES}}|\Psi_{\mathrm{NES}}\rangle}, \tag*{(8)}

where in the exact limit, the minimum would be the sum of the lowest MM eigenvalues of the original Hamiltonian, and the single-state functions {ΨI}\{\Psi_I\} would span the corresponding low-energy subspace. The individual variational energies and states are obtained by solving the generalized eigenvalue problem,

HC=SCϵ,(9)HC=SC\epsilon, \tag*{(9)}

where HIJ=⟨ΨI∣H^∣ΨJ⟩H_{IJ}=\langle\Psi_I|\hat{H}|\Psi_J\rangle, SIJ=⟨ΨI∣ΨJ⟩S_{IJ}=\langle\Psi_I|\Psi_J\rangle, and ϵ=diag⁡(ϵ1eig,…,ϵMeig)\epsilon=\operatorname{diag}(\epsilon_1^{\mathrm{eig}},\ldots,\epsilon_M^{\mathrm{eig}}) with ϵIeig\epsilon_I^{\mathrm{eig}} being the energy of the II-th state. The II-th column of CC defines the corresponding state,

∣ΨIeig⟩=∑J=1MCJI∣ΨJ⟩.(10)|\Psi_I^{\mathrm{eig}}\rangle=\sum_{J=1}^{M}C_{JI}|\Psi_J\rangle. \tag*{(10)}

When solving the NES problem using the VMC framework, the single-state functions {ΨI}\{\Psi_I\} are parameterized and optimized, where the NES energy to be minimized is estimated through stochastic sampling by

ENES=Tr⁡[⟨Eloc(nNES)⟩nNES∼P(nNES)],(11)E_{\mathrm{NES}}=\operatorname{Tr}\left[\langle E_{\mathrm{loc}}(n_{\mathrm{NES}})\rangle_{n_{\mathrm{NES}}\sim P(n_{\mathrm{NES}})}\right], \tag*{(11)}

where

Eloc(nNES)=ΨNES−1(nNES)(H^ΨNES)(nNES)(12)E_{\mathrm{loc}}(n_{\mathrm{NES}})=\mathbf{\Psi}_{\mathrm{NES}}^{-1}(n_{\mathrm{NES}})(\hat{H}\mathbf{\Psi}_{\mathrm{NES}})(n_{\mathrm{NES}}) \tag*{(12)}

is the matrix local energy with

(H^ΨNES)(nNES)=[⟨n1∣H^∣Ψ1⟩⋯⟨n1∣H^∣ΨM⟩⋮⋱⋮⟨nM∣H^∣Ψ1⟩⋯⟨nM∣H^∣ΨM⟩],(\hat{H}\mathbf{\Psi}_{\mathrm{NES}})(n_{\mathrm{NES}})= \begin{bmatrix} \langle n_1|\hat{H}|\Psi_1\rangle& \cdots& \langle n_1|\hat{H}|\Psi_M\rangle\\ \vdots& \ddots& \vdots\\ \langle n_M|\hat{H}|\Psi_1\rangle& \cdots& \langle n_M|\hat{H}|\Psi_M\rangle \end{bmatrix},

and

P(nNES)=∣ΨNES(nNES)∣2∑nNES∣ΨNES(nNES)∣2.P(n_{\mathrm{NES}})=\frac{|\Psi_{\mathrm{NES}}(n_{\mathrm{NES}})|^2}{\sum_{n_{\mathrm{NES}}}|\Psi_{\mathrm{NES}}(n_{\mathrm{NES}})|^2}.

In order to obtain the energy of each state after the variational optimization, one can estimate the matrix energy:

Emat=⟨Eloc(nNES)⟩nNES∼P(nNES),(13)E_{\mathrm{mat}}=\langle E_{\mathrm{loc}}(n_{\mathrm{NES}})\rangle_{n_{\mathrm{NES}}\sim P(n_{\mathrm{NES}})}, \tag*{(13)}

where it can be proved that Emat=S−1HE_{\mathrm{mat}}=S^{-1}H, such as the eigenvalues of EmatE_{\mathrm{mat}} are the desired {ϵIeig}\{\epsilon_I^{\mathrm{eig}}\} 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 NN-electron system in occupation-number representation, a set of occupied GHF orbitals can be encoded by a coefficient matrix UGHF∈F2K×N\mathbf{U}^{\mathrm{GHF}} \in\mathbb{F}^{2K \times N}, where each column corresponds to an occupied GHF orbital represented as a combination of 2K2K basis spin orbitals, containing both spin-up and spin-down components. Therefore, the transformed creation operators are

b^j†=∑p=12KUpjGHFa^p†,j=1,…,N,(14)\hat{b}_j^\dagger= \sum_{p=1}^{2K} U_{pj}^{\mathrm{GHF}} \hat{a}_p^\dagger,\qquad j=1,\ldots,N, \tag*{(14)}

and the corresponding Slater determinant state is

∣ΨGHF⟩=∏j=1Nb^j†∣vac⟩,(15)|\Psi^{\mathrm{GHF}}\rangle= \prod_{j=1}^{N}\hat{b}_j^\dagger|\mathrm{vac}\rangle, \tag*{(15)}

and one can derive that the corresponding amplitude in the original basis can be written as

⟨n∣ΨGHF⟩=det⁡[n⋆UGHF].(16)\langle\boldsymbol{n}|\Psi^{\mathrm{GHF}}\rangle= \det[\boldsymbol{n}\star\mathbf{U}^{\mathrm{GHF}}]. \tag*{(16)}

Here, the symbol n⋆\boldsymbol{n}\star represents the operation of selecting rows from the matrix on its right according to the occupied orbitals in n\boldsymbol{n}.

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 UGHF\mathbf{U}^{\mathrm{GHF}} into a set of configuration-dependent matrices UdNNBF(n)\mathbf{U}^{\mathrm{NNBF}}_d(\boldsymbol{n}), generated by a neural network with n\boldsymbol{n} as the input, which gives

⟨n∣ΨNNBF⟩=∑d=1Ddet⁡[n⋆UdNNBF(n)],(17)\langle\boldsymbol{n}|\Psi^{\mathrm{NNBF}}\rangle= \sum_{d=1}^{D}\det[\boldsymbol{n}\star\mathbf{U}^{\mathrm{NNBF}}_d(\boldsymbol{n})], \tag*{(17)}

with DD being the number of determinants.

In this work, we use a multilayer perceptron (MLP) to implement UdNNBF(n)\mathbf{U}^{\mathrm{NNBF}}_d(\boldsymbol{n}), which is specified by the number of layers nlayern_{\mathrm{layer}} and the number of hidden units nhiddenn_{\mathrm{hidden}}. Accordingly, we denote an NNBF model as NNBF(nlayern_{\mathrm{layer}}, nhiddenn_{\mathrm{hidden}}, DD). 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 NN electrons in KK spatial orbitals, with total spin SS and spin projection MSM_S. The electron spins can be coupled to form a spin function, denoted as Θσ1,…,σN\Theta_{\sigma_1,\ldots,\sigma_N}, which satisfies

S^2Θ=S(S+1)Θ,S^zΘ=MSΘ,(18)\hat{S}^{2}\Theta= S(S+1)\Theta,\qquad\hat{S}_{z}\Theta= M_{S}\Theta, \tag*{(18)}

and Θ\Theta can be expanded as a sum of products of one-electron spin functions,

Θσ1,…,σN=∑r=1RCr∏j=1N((sr)1jασj+(sr)2jβσj)=∑r=1RCrgr(σ1,…,σN).(19)\begin{aligned} \Theta_{\sigma_1,\ldots,\sigma_N} &= \sum_{r=1}^{R} C_r \prod_{j=1}^{N}\left((s_r)_{1j}\alpha_{\sigma_j}+(s_r)_{2j}\beta_{\sigma_j}\right)\\ &= \sum_{r=1}^{R} C_r g_r(\sigma_1,\ldots,\sigma_N). \tag*{(19)} \end{aligned}

Here, ασj\alpha_{\sigma_j} and βσj\beta_{\sigma_j} denote the spin-up and spin-down one-electron spin eigenfunctions for the spin variable σj\sigma_j, respectively.[53] The matrix sr∈F2×N\mathbf{s}_r \in\mathbb{F}^{2\times N} gives the coefficients of these one-electron spin functions in the rr-th product term, and RR denotes the number of product terms used to represent Θ\Theta. In this work, we give a detailed discussion on the alternative choices of Θ\Theta as well as the available implementations for Eq. (19) in § 3.

For a fixed spin function Θ\Theta, the SA-NNBF ansatz uses a neural network to produce spatial-orbital transformation matrices, denoted as U‾(n~)∈FK×N\overline{\mathbf{U}}(\tilde{\boldsymbol{n}}) \in\mathbb{F}^{K\times N}, with the spatial occupation number vector n~\tilde{\boldsymbol{n}} as the input. Here n~p=npα+npβ∈{0,1,2}\tilde{n}_p=n_{p\alpha}+n_{p\beta}\in\{0,1,2\}, where npαn_{p\alpha} and npβn_{p\beta} denote the occupation numbers of the spin-up and spin-down orbitals associated with spatial orbital pp, respectively. With the spatial matrix and the spin function specified, the coefficient matrix srs_r of the rr-th product is combined with Uˉ(nˉ)\bar{\mathbf{U}}(\bar{n}) to form a spin-orbital coefficient matrix Ur(nˉ)∈F2K×N\mathbf{U}_r(\bar{n}) \in\mathbb{F}^{2K \times N}, whose components are

[Ur](p,σ),j(nˉ)=Uˉpj(nˉ)×(sr)σj.(20)[\mathbf{U}_r]_{(p,\sigma),j}(\bar{n}) = \bar{U}_{pj}(\bar{n}) \times(s_r)_{\sigma j}. \tag*{(20)}

This matrix depends only on the spatial occupation nˉ\bar{n}, so configurations that differ only by a spin assignment share the same Ur(nˉ)\mathbf{U}_r(\bar{n}). The amplitude on a configuration nn is obtained by antisymmetrization: n⋆n\star selects the rows of the spin orbitals occupied in nn, and the determinant of those rows gives

ΨSA-NNBF(n)=∑r=1RCrdet⁡[n⋆Ur(nˉ)].(21)\Psi^{\mathrm{SA\text{-}NNBF}}(n) = \sum_{r=1}^{R} C_r \det[n\star\mathbf{U}_r(\bar{n})]. \tag*{(21)}

When Eq. (19) represents Θ\Theta exactly, the state in Eq. (21) is an eigenfunction of both S^2\hat{S}^{2} and S^z\hat{S}_{z} for any spatial matrix Uˉ(nˉ)\bar{\mathbf{U}}(\bar{n}). We denote this state by ∣ΨSA-NNBF(Uˉ,Θ)⟩|\Psi^{\mathrm{SA\text{-}NNBF}}(\bar{\mathbf{U}},\Theta)\rangle. As a function of the trainable matrix Uˉ\bar{\mathbf{U}}, the amplitude on configuration nn is written ΨSA-NNBF[Uˉ](n)\Psi^{\mathrm{SA\text{-}NNBF}}[\bar{\mathbf{U}}](n).

In analogy to NNBF and following Ref. 44, we simply choose an MLP to implement Uˉ(nˉ)\bar{\mathbf{U}}(\bar{n}) 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 Θ\Theta 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 SS and MsM_s, there exist many Θ\Theta’s that satisfy Eq. (18), [54] where in SA-NNBF, the choice of Θ\Theta 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 1/21/2. For given NN, SS and MsM_s, the spin functions generated by this method form a complete orthogonal basis for the corresponding NN-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 S=5/2S = 5/2. 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 S=5/2S = 5/2, and then couple the two local spins to the target total spin SS; see Figure 2(a).

Figure 2. Spin-coupling trees for (a) [2Fe-2S] in the ten-hole representation and (b,c) O2_2 in the 16-electron representation. In (c), four singlet pairs are omitted from the drawing. The RR 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]

ΘVB=12Npair/2[∏k=1Npair(ασ2k−1βσ2k−βσ2k−1ασ2k)]×[∏j=2Npair+1Nασj],(22)\Theta^{\mathrm{VB}} = \frac{1}{2^{N_{\mathrm{pair}}/2}} \left[\prod_{k=1}^{N_{\mathrm{pair}}} \left(\alpha_{\sigma_{2k-1}}\beta_{\sigma_{2k}}-\beta_{\sigma_{2k-1}}\alpha_{\sigma_{2k}}\right)\right] \times\left[\prod_{j=2N_{\mathrm{pair}}+1}^{N}\alpha_{\sigma_j}\right], \tag*{(22)}

where Npair≡(N−2S)/2N_{\mathrm{pair}} \equiv(N-2S)/2 is the number of electron pairs. In genealogical coupling scheme, Eq. (22) corresponds to the coupling path along which the first 2Npair2N_{\mathrm{pair}} 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 O2\mathrm{O}_2 shown in Figure 2(c), which couples each electron pair into an S=0S=0 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 RR 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

RSAAM=∏l(Nl+1),(23)R_{\mathrm{SAAM}} = \prod_l (N_l+1), \tag*{(23)}

in which the coefficient of a few terms might be coincidentally zero. Here NlN_l is the number of electrons in the ll-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), RSAAMR_{\mathrm{SAAM}} is nearly half of the electron number, while for the tree in Figure 2(c), RSAAMR_{\mathrm{SAAM}} 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 RR, such scheme is efficient enough for the proof-of-concept calculations in this work.

An Example of a Spin Function

For molecular O2\mathrm{O}_2, although the tree in Figure 2(b) gives a much smaller RR 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 O2\mathrm{O}_2 molecule. Moreover, we simplify the backflow spatial orbitals into fixed spatial orbitals, including two doubly occupied ones and two singly occupied ones:

Ω1=(φaφaφbφbφcφd),\Omega_1=(\varphi_a\varphi_a\varphi_b\varphi_b\varphi_c\varphi_d),

where φa\varphi_a–φd\varphi_d are spatial orbitals in the spatial product Ω1\Omega_1. One could verify that the SA-NNBF with Ω1\Omega_1 as the spatial part and the VB spin wavefunction ΘVB\Theta^{\mathrm{VB}} in Eq. (22) as the spin part, is identical to a restricted open-shell Hartree–Fock (ROHF) determinant up to normalization:

∣ΨSA−NNBF(Ω1,ΘVB)⟩∝∣ΨROHF⟩,(24)\left|\Psi^{\mathrm{SA-NNBF}}\left(\Omega_1,\Theta^{\mathrm{VB}}\right)\right\rangle\propto\left|\Psi^{\mathrm{ROHF}}\right\rangle, \tag*{(24)}

where

∣ΨROHF⟩≡∣φaφbφcφdφˉaφˉb⟩,\left|\Psi^{\mathrm{ROHF}}\right\rangle\equiv\left|\varphi_a\varphi_b\varphi_c\varphi_d\bar{\varphi}_a\bar{\varphi}_b\right\rangle,

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 Θ3∣3\Theta_{3|3} and Θ4∣2\Theta_{4|2}. The two trees give the following spin functions.

Figure 3. All the possible 2-leaf coupling tree with six electrons in the (S,MS)=(1,1)(S,M_S)=(1,1) sector, giving the spin functions Θ3∣3\Theta_{3|3} and Θ4∣2\Theta_{4|2}, respectively.

Θ3∣3=110(α1α2α3)(α4β5β6+β4α5β6+β4β5α6)−1325(α1α2β3+α1β2α3+β1α2α3)(α4α5β6+α4β5α6+β4α5α6)+110(α1β2β3+β1α2β3+β1β2α3)(α4α5α6),\begin{aligned} \Theta_{3|3}={}&\sqrt{\frac{1}{10}}(\alpha_1\alpha_2\alpha_3)(\alpha_4\beta_5\beta_6+\beta_4\alpha_5\beta_6+\beta_4\beta_5\alpha_6)\\ &-\frac{1}{3}\sqrt{\frac{2}{5}}(\alpha_1\alpha_2\beta_3+\alpha_1\beta_2\alpha_3+\beta_1\alpha_2\alpha_3)(\alpha_4\alpha_5\beta_6+\alpha_4\beta_5\alpha_6+\beta_4\alpha_5\alpha_6)\\ &+\sqrt{\frac{1}{10}}(\alpha_1\beta_2\beta_3+\beta_1\alpha_2\beta_3+\beta_1\beta_2\alpha_3)(\alpha_4\alpha_5\alpha_6), \end{aligned}
Θ4∣2=160(α1α2)(α3α4β5β6+α3β4α5β6+β3α4α5β6+β3α4β5α6+β3β4α5α6+α3β4β5α6)−1435(α1β2+β1α2)(α3α4α5β6+α3α4β5α6+α3β4α5α6+β3α4α5α6)+35β1β2α3α4α5α6.\begin{aligned} \Theta_{4|2}={}&\frac{1}{\sqrt{60}}(\alpha_1\alpha_2)(\alpha_3\alpha_4\beta_5\beta_6+\alpha_3\beta_4\alpha_5\beta_6+\beta_3\alpha_4\alpha_5\beta_6\\ &+\beta_3\alpha_4\beta_5\alpha_6+\beta_3\beta_4\alpha_5\alpha_6+\alpha_3\beta_4\beta_5\alpha_6)\\ &-\frac{1}{4}\sqrt{\frac{3}{5}}(\alpha_1\beta_2+\beta_1\alpha_2)(\alpha_3\alpha_4\alpha_5\beta_6+\alpha_3\alpha_4\beta_5\alpha_6\\ &+\alpha_3\beta_4\alpha_5\alpha_6+\beta_3\alpha_4\alpha_5\alpha_6)\\ &+\sqrt{\frac{3}{5}}\beta_1\beta_2\alpha_3\alpha_4\alpha_5\alpha_6. \end{aligned}

Here, αi\alpha_i and βi\beta_i denote the one-electron spin functions on electron ii. 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:

Ω2=(φaφbφcφdφaφb),\Omega_2 = (\varphi_a\varphi_b\varphi_c\varphi_d\varphi_a\varphi_b),

resulting in,

∣ΨSA-NNBF(Ω2,Θ3∣3)⟩=103∣ΨROHF⟩,∣ΨSA-NNBF(Ω2,Θ4∣2)⟩=53∣ΨROHF⟩.(25)\left|\Psi^{\mathrm{SA\text{-}NNBF}}(\Omega_2,\Theta_{3|3})\right\rangle= \frac{\sqrt{10}}{3}\left|\Psi^{\mathrm{ROHF}}\right\rangle, \\ \left|\Psi^{\mathrm{SA\text{-}NNBF}}(\Omega_2,\Theta_{4|2})\right\rangle= \sqrt{\frac{5}{3}}\left|\Psi^{\mathrm{ROHF}}\right\rangle. \tag*{(25)}

In practice, since U‾\overline{\mathbf{U}} 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 ΘVB\Theta^{\mathrm{VB}}, Θ3∣3\Theta_{3|3} or Θ4∣2\Theta_{4|2} 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 nhiddenn_{\mathrm{hidden}}, 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 nˉ0\bar{\mathbf{n}}^0, where no matter how complicated is the neural network architecture, the SA-NNBF on these configurations share a fixed same U‾(nˉ0)≡U‾0\overline{\mathbf{U}}(\bar{\mathbf{n}}^0) \equiv\overline{\mathbf{U}}^0. Therefore, the SA-NNBF amplitudes within such subspace are completely determined by n\mathbf{n} and matrix U‾0\overline{\mathbf{U}}^0 according to Eq. (20) and Eq. (21):

ΨSA-NNBF(n)=ΨSA-NNBF[U‾0](n),∀n→nˉ0.\Psi^{\mathrm{SA\text{-}NNBF}}(\mathbf{n}) = \Psi^{\mathrm{SA\text{-}NNBF}}[\overline{\mathbf{U}}^0](\mathbf{n}), \quad\forall\mathbf{n} \to\bar{\mathbf{n}}^0.

To reproduce the exact wavefunction ΨExact\Psi^{\mathrm{Exact}}, the SA-NNBF needs to satisfy the following set of equations:

ΨSA-NNBF[U‾0](n)=ΨExact(n),∀n→nˉ0.(26)\Psi^{\mathrm{SA\text{-}NNBF}}[\overline{\mathbf{U}}^0](\mathbf{n}) = \Psi^{\mathrm{Exact}}(\mathbf{n}), \quad\forall\mathbf{n} \to\bar{\mathbf{n}}^0. \tag*{(26)}

The equations in Eq. (26) are indexed by the configurations n\mathbf{n} whose spatial occupation is the chosen nˉ0\bar{\mathbf{n}}^0. Because ∣ΨExact⟩\left|\Psi^{\mathrm{Exact}}\right\rangle is a spin eigenstate, the right-hand sides ΨExact(n)\Psi^{\mathrm{Exact}}(\mathbf{n}) for this set of n\mathbf{n} 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 nˉ0\bar{\mathbf{n}}^0, 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, U‾0\overline{\mathbf{U}}^0. When nˉ0\bar{\mathbf{n}}^0 places NN electrons in NN distinct spatial orbitals, only the NN rows of those orbitals enter the determinants, so U‾0\overline{\mathbf{U}}^0 contributes at most N2N^2 unknown entries. For large NN, the number of independent equations exceeds N2N^2. 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 (N,S,MS)(N,S,M_S) sector.

To overcome the limitation above, in this work, we expand the SA-NNBF with multiple sets of spatial orbitals in analogy to the D>1D>1 form of Eq. (17):

ΨDSA-NNBF(n)=∑d=1D∑r=1RCrdet⁡[n⋆Udr(nˉ)],(27)\Psi_D^{\mathrm{SA\text{-}NNBF}}(\mathbf{n}) = \sum_{d=1}^{D}\sum_{r=1}^{R} C_r \det[\mathbf{n}\star\mathbf{U}_{dr}(\bar{\mathbf{n}})], \tag*{(27)}

where

[Udr(nˉ)](p,σ),j=[U‾d(nˉ)]pj(sr)σj,[\mathbf{U}_{dr}(\bar{\mathbf{n}})]_{(p,\sigma),j} = [\overline{\mathbf{U}}_d(\bar{\mathbf{n}})]_{pj}(s_r)_{\sigma j},

with {U‾d(nˉ)}\{\overline{\mathbf{U}}_d(\bar{\mathbf{n}})\} produced by the neural network. Following the notation for NNBF, we denote an SA-NNBF model configuration by SANNBF(nlayer,nhidden,Dn_{\mathrm{layer}}, n_{\mathrm{hidden}}, D).

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 ∣Ψ(N,S,MS)Exact⟩|\Psi^{\mathrm{Exact}}_{(N,S,M_S)}\rangle in a fixed (N,S,MS)(N,S,M_S) sector and any nonzero spin function Θ∈S(N,S,MS)\Theta\in S_{(N,S,M_S)}, an SA-NNBF state with the fixed spin function Θ\Theta can exactly represent ∣Ψ(N,S,MS)Exact⟩|\Psi^{\mathrm{Exact}}_{(N,S,M_S)}\rangle when DD 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 Θ\Theta, although in a practical calculation with a finite DD, 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 nlayer=1n_{\mathrm{layer}} = 1. For a given system, the number of trainable parameters npn_p is then approximately proportional to nhidden×Dn_{\mathrm{hidden}} \times D. At the same nhiddenn_{\mathrm{hidden}} and DD, NNBF has roughly twice as many trainable parameters as SA-NNBF since its backflow orbitals are spin orbitals rather than spatial orbitals. Increasing nlayern_{\mathrm{layer}} 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 npn_p between nhiddenn_{\mathrm{hidden}} and DD and scaling strategies for SA-NNBF and compare them with NNBF. We start with the triplet ground state of molecular O2_2, as a simple case to benchmark different strategies. Figure 4(a) examines how the accuracy is affected as DD increases while nhiddenn_{\mathrm{hidden}} decreases to keep npn_p 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 npn_p level, striking a balance between nhiddenn_{\mathrm{hidden}} and DD. However, the preferred strategies for the two ansätze differ substantially: the optimal DD values for SA-NNBF, typically D≥4D \ge4, 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 nhiddenn_{\mathrm{hidden}} or DD relative to FCI for the triplet ground state of O2_2 in the 6-31G basis with an (18o,16e)(18o,16e) active space. For each vertically aligned pair of data points with the same color, there is nhiddenSA−NNBF=2nhiddenNNBFn_{\mathrm{hidden}}^{\mathrm{SA-NNBF}} = 2n_{\mathrm{hidden}}^{\mathrm{NNBF}} 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 nhiddenNNBF=16/Dn_{\mathrm{hidden}}^{\mathrm{NNBF}} = 16/D, 32/D32/D and 64/D64/D at np≈10000n_p \approx10000, 20000 and 40000, respectively. For the same color, nhiddenSA−NNBFn_{\mathrm{hidden}}^{\mathrm{SA-NNBF}} is twice nhiddenNNBFn_{\mathrm{hidden}}^{\mathrm{NNBF}}. (b) Errors versus hidden width at different fixed DD’s where the lower and upper axes correspond to the hidden width for NNBF (nhiddenNNBFn_{\mathrm{hidden}}^{\mathrm{NNBF}}) and SA-NNBF (nhiddenSA−NNBFn_{\mathrm{hidden}}^{\mathrm{SA-NNBF}}), respectively. (c) Errors versus DD at different fixed NNBF hidden widths.

We further compare two different strategies of systematically scaling the model: increasing nhiddenn_{\mathrm{hidden}} at fixed DD in Figure 4(b), and increasing DD at fixed nhiddenn_{\mathrm{hidden}} 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 DD over nhiddenn_{\mathrm{hidden}} 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 S=0,1,…,5S = 0, 1, \ldots, 5 sectors respectively. Since the energies of these states increase with SS, 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 MSM_S to the target SS, 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 DD 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 nhiddenn_{\mathrm{hidden}} 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 O2_2.

Figure 5. Optimization curves of the energy error relative to the corresponding FCI energy for the lowest [2Fe-2S] states with S=0,…,5S = 0,\ldots,5 and MS=SM_S = S, obtained with vanilla NNBF(1,128,1) and SA-NNBF with different DD or nhiddenn_{\mathrm{hidden}}. The active space is (20o,30e). The curves are smoothed using a 500-step moving average. The shaded area represents the chemical accuracy (<1< 1 kcal/mol).

In summary, the O2_2 and [2Fe-2S] benchmarks show that, for the expanded SA-NNBF in this work, increasing DD is preferable to increasing nhiddenn_{\mathrm{hidden}} 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 nhiddenn_{\mathrm{hidden}} or a sufficiently large DD, whereas SA-NNBF can do so only by increasing DD sufficiently, highlighting the more important role of DD 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 O2_2, the lowest S=0S = 0 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 DD reduces the error to 0.46 mHa at D=5D = 5. By contrast, a vanilla NNBF calculation constrained only by Ms=0M_s = 0 also includes the lower-energy Ms=0M_s = 0 component of the triplet ground state and therefore cannot isolate this singlet by energy minimization alone.

modelnpn_{\mathrm{p}}energy (Ha)error (mHa)
(1,64,1)19 937−149.74522(4)-149.74522(4)1.46
(1,64,2)38 657−149.74574(3)-149.74574(3)0.94
(1,64,3)57 377−149.74595(2)-149.74595(2)0.72
(1,64,4)76 097−149.74611(2)-149.74611(2)0.56
(1,64,5)94 817−149.74621(2)-149.74621(2)0.46

Table 1. Energies and absolute errors relative to FCI for the lowest singlet state of O2_2 from single-state SA-NNBF calculations.

We further combine SA-NNBF with NES to calculate the lowest four triplets of O2_2, 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 D=3D = 3 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

stateenergy (Ha)error (mHa)
X 3Σg−^3\Sigma_g^-−149.78801(5)-149.78801(5)0.91
A′3Δu'{}^3\Delta_u−149.55769(3)-149.55769(3)0.55
A′3Δu'{}^3\Delta_u−149.55728(5)-149.55728(5)0.97
A 3Σu+^3\Sigma_u^+−149.55320(3)-149.55320(3)0.64

Table 2. Energies and absolute errors relative to FCI for the four lowest triplet states of O2_2 from a 4-state SA-NNBF+NES calculation.

Table 1: Energies and absolute errors relative to FCI for the lowest singlet state of O2_2 from single-state SA-NNBF calculations. The spin function is chosen as the one represented by Figure 2(b) with S=0S = 0.

4-state SA-NNBF calculation since constructing the NES wavefunction and local-energy matrices requires M2M^2 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 O2_2 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 S=1S = 1. The two A′3Δu'{}^3\Delta_u 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 S0\mathrm{S}_0, S1\mathrm{S}_1, and S2\mathrm{S}_2, there are over 30 higher-spin states lying between S1\mathrm{S}_1 and S2\mathrm{S}_2, all of which need to be included in a conventional NES calculation with NNBF to reach S2\mathrm{S}_2, 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 (S0\mathrm{S}_0, S1\mathrm{S}_1, S2\mathrm{S}_2) 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 S=0S = 0. 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 (<1< 1 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 DD sets of spatial orbitals and proved its universality for sufficiently large DD. Moreover, we performed a relatively systematic comparison of NNBF and SA-NNBF across the parameter allocation for the triplet ground state of O2\mathrm{O}_2 and the lowest states of [2Fe-2S] in different spin sectors. Within the tested range, increasing DD improved the accuracy of SA-NNBF more effectively than increasing nhiddenn_{\mathrm{hidden}} alone, highlighting the central role of DD 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 O2\mathrm{O}_2, which is its first excited state. Chemical accuracy was reached already at D=1D = 1, and increasing DD further improved the accuracy. We further combined SA-NNBF with NES to calculate low-energy triplets of O2\mathrm{O}_2 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 O2\mathrm{O}_2.

References

References

  1. [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
  2. [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
  3. [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
  4. [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
  5. [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
  6. [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
  7. [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
  8. [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
  9. [15]Carleo, G.; Troyer, M. Solving the Quantum Many-Body Problem with Artificial Neural Networks. Science 2017, 355, 602–606.DOI
  10. [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
  11. [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
  12. [18]Luo, D.; Clark, B. K. Backflow Transformations via Neural Networks for Quantum Many-Body Wave Functions. Phys. Rev. Lett. 2019, 122, 226401.DOI
  13. [19]Liu, A.-J.; Clark, B. K. Neural Network Backflow for Ab Initio Quantum Chemistry. Phys. Rev. B 2024, 110, 115137.DOI
  14. [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
  15. [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
  16. [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
  17. [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
  18. [24]Becca, F.; Sorella, S. Quantum Monte Carlo approaches for correlated systems; Cambridge University Press, 2017.DOI
  19. [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
  20. [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
  21. [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
  22. [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
  23. [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
  24. [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.
  25. [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
  26. [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
  27. [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
  28. [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
  29. [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
  30. [36]Sharma, S.; Chan, G. K.-L. Spin-Adapted Density Matrix Renormalization Group Algorithms for Quantum Chemistry. J. Chem. Phys. 2012, 136, 124121.DOI
  31. [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
  32. [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
  33. [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
  34. [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
  35. [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
  36. [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.
  37. [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
  38. [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
  39. [45]Kingma, D. P.; Ba, J. Adam: A Method for Stochastic Optimization. International Conference on Learning Representations. 2015.arxiv.org/abs/1412.6980
  40. [46]Loshchilov, I.; Hutter, F. Decoupled Weight Decay Regularization. International Conference on Learning Representations. 2019.
  41. [47]Sorella, S. Green Function Monte Carlo with Stochastic Reconfiguration. Phys. Rev. Lett. 1998, 80, 4558–4561.
  42. [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.
  43. [49]Chen, A.; Heyl, M. Empowering Deep Neural Quantum States through Efficient Optimization. Nat. Phys. 2024, 20, 1476–1481.DOI
  44. [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
  45. [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
  46. [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
  47. [53]Helgaker, T.; Jørgensen, P.; Olsen, J. Molecular Electronic-Structure Theory; John Wiley & Sons: Chichester, England, 2000.
  48. [54]Pauncz, R. Spin Eigenfunctions; Plenum Press: New York, 1979.
  49. [55]Salmon, W. I. Advances in Quantum Chemistry Volume 8; Advances in Quantum Chemistry; Elsevier, 1974; Vol. 8; pp 37–94.
  50. [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
  51. [57]Li, Z.; Chan, G. K.-L. Hilbert Space Renormalization for the Many-Electron Problem. J. Chem. Phys. 2016, 144, 084103.DOI
  52. [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
  53. [59]Jiménez-Hoyos, C. A.; Henderson, T. M.; Tsuchimochi, T.; Scuseria, G. E. Projected Hartree-Fock Theory. J. Chem. Phys. 2012, 136, 164109.DOI
  54. [60]Kolda, T. G.; Bader, B. W. Tensor decompositions and applications. SIAM Rev. 2009, 51, 455–500.DOI
  55. [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
  56. [62]Viteritti, L. L.; Rende, R.; Becca, F. Transformer Variational Wave Functions for Frustrated Quantum Spin Systems. Phys. Rev. Lett. 2023, 130, 236401.DOI
  57. [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
  58. [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

Paper details

vlnc:W_8b2g8tfu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents