betaFeedback

WorksPhysics and Astronomy

WF-Bench: A Benchmark for Neural Network WaveFunction Expressivity and Scaling Laws

Abstract

We present a comprehensive benchmarking dataset and empirical scaling law analysis for neural network wavefunctions by matching them to a wide spectrum of famous many body target wavefunctions. The dataset, WF-Bench, spans multiple distinct regimes of strongly correlated quantum matter, including topological states, Wigner crystals, and superconducting wavefunctions, providing a diverse and challenging test bed for neural network wavefunction expressivity. We introduce a systematic and reproducible benchmarking protocol for target wavefunction matching, enabling consistent performance evaluation across different neural network wavefunction architectures. By using wavefunction fidelity as the uniform metric, we discover empirical scaling laws that characterize how representability depends on system size and key model parameters, including number of determinant and model depth. By applying our benchmark protocol on Psiformer and Ferminet, we show that WF-Bench establishes a unified dataset driven framework for evaluating and comparing neural network wavefunctions and for guiding the design of future architectures.

Introduction

Recent years have seen rapid progress in neural network (NN) wavefunctions as flexible variational ansatze for quantum many body problems. Harnessing the representation power of deep neural networks, NN wavefunctions have demonstrated remarkable scalability and accuracy across a variety of quantum many body tasks, including ground-state optimization[42, 50], excited-state spectroscopy[16, 40], open system dynamics[13, 28], real-time evolutions[23, 9], and quantum state tomography[48, 41]. These capabilities have broad impact across multiple disciplines, including condensed matter physics[31, 20, 54], high energy physics[29, 27], quantum chemistry[1] (Hermann et al. 2023), quantum information science[19, 47], and quantum computations[34, 2]. As the field expands, an increasing number of powerful neural network architectures have been proposed to tackle these challenges. Notable examples include FermiNet[39], based on streaming permutation-equivariant one- and two body features; Psiformer[49], which leverages self-attention mechanisms; graph neural networks built on message passing[32, 38, 15]; and DeepSolid(Li et al., 2022), designed for periodic systems.

It has long been known that the exact solution of the many body Schrödinger equation requires computational resources that grow exponentially with the system size. However, approximate solutions can often be obtained with only polynomial computational cost. Therefore, a good NN wavefunction must have a careful balance between computational efficiency and representation accuracy. This is directly related to the expressive power of the underlying architecture. Despite rapid advances in NN wavefunctions, there is still no systematic understanding of how representation power varies across architectures and classes of physical systems, nor a unified benchmarking framework for evaluating NN wavefunction expressivity and scaling laws.

Recently, researchers have begun to investigate the expressivity and scaling behavior of neural wavefunctions in both theoretical and numerical studies [7, 4, 53, 37, 14, 36, 43] (Lu et al. 2026). ([36]) develops a density-current optimization objective together with a wavefunction overlap-based analysis of scaling in Laughlin, Moore–Read, and BCS wavefunctions. Inspired by the previous advancement, we introduce WF-Bench, a comprehensive dataset for benchmarking the representation power of NN wavefunctions. WF-Bench contains over 30 target wavefunctions spanning three major classes: topological states, superconducting states, and Wigner crystals. They cover a broad range of physically important systems with intrinsically different correlation structures. The topological class includes well-known trial wavefunctions with nontrivial topological order and complex phase structure, such as the Laughlin ([17]) and Moore Read ([35]) states of fractional quantum Hall systems. The superconducting class consists of Bardeen Cooper Schrieffer (BCS) type wavefunctions with diverse pairing symmetries and spin configurations. The Wigner crystal class comprises states exhibiting spontaneous translational symmetry breaking driven by strong interactions or external periodic potentials. Depending on the target state, fermionic antisymmetry is realized either through a Slater determinant form or a Pfaffian form.

Schematic of the WF-Bench workflow

Figure 1. Schematic of the WF-Bench workflow. The dataset consists three categories of wavefunctions: topological states, Wigner crystals and superconducting wavefunctions. By using −log⁡F-\log F as the loss function, NN wavefunctions are optimized to match both the phase and the amplitude of the target wavefunctions. This enables uniform representation power benchmarks across key system and model parameters, including number of electron, determinant number, and network depth.

In addition, by optimizing the fidelity between NN wavefunction and the target wavefunctions, we present a uniform benchmarking protocol under the framework of Monte-Carlo (MC) sampling. By applying our benchmarking protocol to two of the widely used NN wavefunction architectures, Psiformer and FermiNet, we systematically study the empirical scaling of the maximum achievable fidelity FF with respect to three important parameters: the number of electrons NeN_e, the number of determinant NdetN_{\mathrm{det}}, and the number of layers NlayerN_{\mathrm{layer}}. We found empirically that 1−F1-F follows a power-law scaling with NeN_e for all wavefunctions in our dataset. As NdetN_{\mathrm{det}} and NlayerN_{\mathrm{layer}} increase, the fidelity FF exhibits clear diminishing returns, with a sharp initial improvement followed by saturation at larger values.

Contributions of this work are summarized as followed:

  • A comprehensive dataset of physically important systems. We construct a comprehensive dataset spanning three major classes of quantum many body states—topological states, superconducting wavefunctions, and Wigner crystals—covering a wide range of correlation strength, pairing symmetry, and phase complexity.

  • A unified benchmarking protocol for NN wavefunction representability. We propose a unified training and evaluation protocol based on fidelity optimization, enabling fair and reproducible comparisons across different NN wavefunction architectures and expressiveness parameters, such as number of determinants and number of layers.

  • Empirical scaling laws for NN wavefunction representability. By systematically varying electron number and network capacity, we identify empirical scaling laws governing the achievable fidelity of NN wavefunction, and relate the scaling exponents to physical properties of the target wavefunctions.

Related Work

NN wavefunction architecture In the past few years, there has been a rapid growth of NN wavefunctions designed for different tasks. This includes early works for generic many body ground-state problems ([3]), as well as autoregressive formulations that enable efficient sampling and scalable optimization ([46, 12]). Subsequent work has extended NN wavefunctions to finite-temperature settings (Irikura & Saito, 2020), real-time dynamics [18], and geometrically motivated representations [10]. To incorporate fermionic statistics, several works introduce backflow transformations into NN wavefunctions [26] (Li et al. 2024). More expressive architectures have been proposed using fermionic hidden states [44], tensor-network based autoregressive networks [5], and message-passing networks for homogeneous fermionic gas [38, 15]. Recent works further explore Pfaffian and pairing-based neural wavefunctions [8, 24], applications to dense and strongly correlated quantum matter [52], and large-scale transformer-style architectures for physical systems (Zhdanov et al., 2025).

NN wavefunction expressivity There are ongoing efforts, both theoretical and numerical, to characterize the expressivity and scaling behavior of NN wavefunctions across different quantum settings. Theoretical analyses of the representational power and entanglement structure of neural network wavefunctions include [7], which shows that neural quantum states can efficiently represent highly entangled many body states, and [4], which demonstrates exact representations of quantum many body wavefunctions using deep neural networks. In [53], the conditions under which classical neural networks can represent quantum states are analyzed, while [37] derives rigorous bounds on the entanglement supported by neural network wavefunctions. Recently, (Lu et al., 2026) established an information-theoretic scaling law with numerical verifications for generic auto-regressive neural quantum states. Numerically, [36] introduces the density-current loss for optimization and studies scaling in Laughlin, Moore–Read, and BCS wavefunctions using fidelity as a key metric. [14] examined that the minimal energy error of NN wavefunctions for quantum chemistry scales as O(N−0.52)O(N^{-0.52}), outperforming traditional methods such as CCSD. [43] studied scaling laws of transformer neural quantum states for frustrated Heisenberg models.

Despite a number of studies focused on specific topics of models or regimes, there is still a lack of a uniform, consistent, and comprehensive study of how the representation power of NN wavefunctions scales across different physical systems and architectures. The WF-Bench dataset, together with the accompanying benchmarking protocol, is designed to fill this gap and to advance the field toward a systematic and reproducible framework for studying expressivity scaling in neural network wavefunction.

Algorithm 1 Benchmarking Protocol

Input: Electron numbers {Ne}\{N_e\}, target wavefunction ∣Φ⟩|\Phi\rangle, initial network parameters θ\theta

Output: Fidelity series {F(Ne)}\{F(N_e)\}

Initialize θlast=θ\theta_{\mathrm{last}} = \theta

for each Ne∈{Ne}N_e \in\{N_e\} do

Burn in MC walkers for 100Ne100N_e steps

Initialize parameters θ←θlast\theta\leftarrow\theta_{\mathrm{last}}

if ∣Φ⟩|\Phi\rangle is a topological state then

for t=1t = 1 to 20,00020{,}000 do

Burn in walkers for 5Ne5N_e steps

Sample configurations {R}∼pmix\{R\} \sim p_{\mathrm{mix}}

Estimate pretraining loss LpreL_{\mathrm{pre}} and update θ\theta

end for

end if

for t=1t = 1 to 100,000100{,}000 do

Burn in walkers for 5Ne5N_e steps

Sample configurations {R}∼pθ\{R\} \sim p_{\theta}

Estimate fidelity loss LFL_{\mathrm{F}} and update θ\theta

end for

for t=1t = 1 to 100100 do

Burn in walkers for 5Ne5N_e steps

Sample configurations {R}∼pθ\{R\} \sim p_{\theta}

Evaluate fidelity FF

end for

Record F(Ne)←mean⁡(F)F(N_e) \leftarrow\operatorname{mean}(F)

Store parameters θlast←θ\theta_{\mathrm{last}} \leftarrow\theta

end for

WF-Bench: A comprehensive dataset of important wavefunctions

We benchmark three classes of wavefunctions: topological states, superconducting wavefunctions, and Wigner crystal wavefunctions. Here, we introduce the form of wavefunctions.

Topological wavefunctions

We benchmark two famous trial wavefunctions for topological quantum states: Laughlin wavefunction Laughlin, [17] and Moore Read wavefunction.Moore & Read, [35] They corresponds to the ground states of the fractional quantum hall system at different filling factor ν=1/m\nu= 1/m. For odd mm, the Laughlin wavefunction is

ΨL({zj})=∏i<j(zi−zj)mG({zj}),(1)\Psi_{\mathrm{L}}(\{z_j\}) = \prod_{i<j}(z_i-z_j)^m \mathcal{G}(\{z_j\}), \tag*{(1)}

where zj=xj+iyjz_j=x_j+iy_j, and ℓB=ℏ/(eB)\ell_{\mathrm{B}} = \sqrt{\hbar/(eB)} is the magnetic length. G({zj})=exp⁡(−∑i∣zi∣2/4ℓB2)\mathcal{G}(\{z_j\}) = \exp\left(-\sum_i |z_i|^2/4\ell_{\mathrm{B}}^2\right) is the Gaussian factor that originates from lowest Landau level (LLL) projection. Similarly, for even mm, the Moore Read wavefunction is

ΨMR({zj})=Pf⁡(1zi−zj)∏i<j(zi−zj)mG({zj}).(2)\Psi_{\mathrm{MR}}(\{z_j\}) = \operatorname{Pf}\left(\frac{1}{z_i-z_j}\right)\prod_{i<j}(z_i-z_j)^m\mathcal{G}(\{z_j\}). \tag*{(2)}

where Pf⁡(⋅)\operatorname{Pf}(\cdot) denotes the Pfaffian. As trial wavefunctions for the excited states of the fractional quantum hall problem, quasiholes can be added to the Laughlin and Moore Read states. For NhN_h quasiholes at complex positions {ηa}a=1Nh\{\eta_a\}_{a=1}^{N_h}, we write

ΨXh({rj};{ηa})=[∏a=1Nh∏i=1Ne(zi−ηa)]ΨX({rj}),(3)\Psi_{\mathrm{Xh}}(\{r_j\};\{\eta_a\}) = \left[\prod_{a=1}^{N_h}\prod_{i=1}^{N_e}(z_i-\eta_a)\right]\Psi_{\mathrm{X}}(\{r_j\}), \tag*{(3)}

where X∈{L,MR}X \in\{\mathrm{L},\mathrm{MR}\}, and ηa=ηa,x+iηa,y\eta_a=\eta_{a,x}+i\eta_{a,y} denote the complex coordinates the quasiholes. Numerically, we place the quasiholes {ηa}\{\eta_a\} on a circle of radius 2mNeℓB/2\sqrt{2mN_e}\ell_{\mathrm{B}}/2, i.e., half of the droplet radius. In our dataset, 12 topological states with different mm and NhN_h are included. For example, laughlin_m3h3 denotes a laughlin m=3m=3 state with 3 quasiholes. We show the amplitude and phase structures of topological states in Figure 2a.

Feature plots of target wavefunctions

Figure 2. Feature plots of target wavefunctions. (a) Amplitude and phase of topological states obtained by scanning the position of one electron while fixing all others. Blue dots indicate fixed electrons, and red crosses denote quasiholes. (b) Charge density patterns of different Wigner crystals. (c) Real space structure of the pairing functions for different superconducting states.

Superconducting wavefunctions

We benchmark the BCS family of wavefunctions with different pairing functions De Gennes, [6]. The real space pairing function of the mean-field Cooper pair is given by

f(ri,rj)=∑kg(k)eik⋅(ri−rj),(4)f(r_i,r_j)=\sum_{\mathbf{k}}g(\mathbf{k})e^{i\mathbf{k}\cdot(r_i-r_j)}, \tag*{(4)}

where rir_i and rjr_j denote the positions of two electrons within the cooper pair, k\mathbf{k} denotes the quantized momentum of the periodic system, and g(k)g(\mathbf{k}) is the Cooper-pair orbital, which takes different forms for different superconducting states. We show the shape of different pairing functions in Figure 2c.

To impose particle-number conservation, we adopt the antisymmetrized geminal power (AGP) wavefunction,

ΨBCS({ri,σi})=A[ϕ(1,1′)ϕ(2,2′)⋯ϕ(Np,Np′)],(5)\Psi_{\mathrm{BCS}}(\{r_i,\sigma_i\})=\mathcal{A}\left[\phi(1,1')\phi(2,2')\cdots\phi(N_p,N'_p)\right], \tag*{(5)}

where A\mathcal{A} denotes the antisymmetrizer, implemented as either Pf⁡[⋅]\operatorname{Pf}[\cdot] or det⁡[⋅]\det[\cdot], and Np=Ne/2N_p=N_e/2 is the number of Cooper pairs. Here, ϕ(ri,rj,σi,σj)=f(ri,rj)⟨σi,σj∣χ⟩\phi(r_i,r_j,\sigma_i,\sigma_j)=f(r_i,r_j)\langle\sigma_i,\sigma_j|\chi\rangle is the “geminal”, constructed from the pairing function ff, with σi\sigma_i denoting the spin of the ii-th electron and ∣χ⟩|\chi\rangle the two-electron spin state, which is either singlet or triplet for BCS wavefunctions. In our dataset, we include 11 types of “geminal” functions, covering ss-, pp-, dd-, and ff-wave pairing with different spatial geometries and spin configurations. For example, bcs_pp_tp denotes a chiral px+ipyp_x+ip_y pairing with a triplet spin state ∣↑↑⟩|\uparrow\uparrow\rangle.

Wigner crystal wavefunctions

A Wigner crystal is a strongly correlated quantum state in which electrons spontaneously form crystalline patterns due to strong Coulomb interactions Wigner, [51]. It corresponds to spontaneous breaking of continuous translational symmetry into a discrete lattice structure. A trial wavefunction for the Wigner crystal is constructed by first choosing a set of lattice sites {Ri}\{R_i\}. For a system with single-particle potential V(r)V(r), the localized orbitals are taken as

ϕj(ri)=exp⁡[−αV(ri−Rj)],(6)\phi_j(r_i)=\exp[-\alpha V(r_i-R_j)], \tag*{(6)}

where α\alpha controls the degree of localization. We include three types of potentials: Gaussian orbital, squeezed Gaussian orbital, and a moiré orbital derived from perturbation theory Luo et al., [30]. The Wigner crystal wavefunction is then given by

ΨWC({ri})=det⁡[ϕj(ri)]i,j=1Ne.(7)\Psi_{\mathrm{WC}}(\{r_i\})=\det[\phi_j(r_i)]_{i,j=1}^{N_e}. \tag*{(7)}

The lattice sites {Ri}\{R_i\} are chosen according to the charge density patterns of the system. For the Wigner crystal wavefunctions included in our dataset, we consider {Ri}∈{A,B,C}\{R_i\}\in\{A,B,C\}, where A=[0,0]A=[0,0], B=[12,12]B=[\frac{1}{2},\frac{1}{2}], and C=[13,13]C=[\frac{1}{3},\frac{1}{3}]. Here [n,m][n,m] denotes the lattice coordinates within a hexagonal unit cell. For example, wc_sgau_v1AB denotes a ν=2\nu=2 integer filling Wigner crystal state with a squeezed Gaussian orbital, and the lattice sites are {Ri}={A,B}\{R_i\}=\{A,B\}. We show the charge density patterns of Wigner crystal wavefunctions in Figure 2b.

Benchmarking Protocol

To match a neural network wavefunction ∣Ψθ⟩\lvert\Psi_\theta\rangle with a target wavefunction ∣Φ⟩\lvert\Phi\rangle, a straightforward loss is based on the fidelity. We define the following:

LF=−log⁡[∣⟨Ψθ∣Φ⟩∣2⟨Ψθ∣Ψθ⟩⟨Φ∣Φ⟩](8)L_{\mathrm{F}}=-\log\left[\frac{\left|\langle\Psi_\theta|\Phi\rangle\right|^2}{\langle\Psi_\theta|\Psi_\theta\rangle\langle\Phi|\Phi\rangle}\right] \tag*{(8)}

Let Ψθ(R)=⟨R∣Ψθ⟩\Psi_\theta(\mathbf{R})=\langle\mathbf{R}|\Psi_\theta\rangle (similarly for Φ(R)\Phi(\mathbf{R})), the gradient of LFL_{\mathrm{F}} can be written in an expectation form by sampling from

pθ(R)∝∣Ψθ(R)∣2p_\theta(\mathbf{R}) \propto|\Psi_\theta(\mathbf{R})|^2,

∂θLF=2Re⁡{Epθ[(1−α(R)Epθ[α(R)])∂θlog⁡Ψθ∗(R)]}(9)\partial_\theta L_F = 2\operatorname{Re}\left\{\mathbb{E}_{p_\theta}\left[\left(1-\frac{\alpha(\mathbf{R})}{\mathbb{E}_{p_\theta}[\alpha(\mathbf{R})]}\right)\partial_\theta\log\Psi_\theta^*(\mathbf{R})\right]\right\} \tag*{(9)}

where α(R)=Φ(R)/Ψθ(R)\alpha(\mathbf{R}) = \Phi(\mathbf{R})/\Psi_\theta(\mathbf{R}). Despite its intuitive appeal, LFL_F becomes numerically challenging for large systems. In particular, α(R)\alpha(\mathbf{R}) demonstrates increasing variance as NeN_e grows, due to interference. As a result, LFL_F suffers from a vanishing effective signal, making direct optimization unstable at large NeN_e.

For wavefunctions with moderate phase structure, the behavior of α(R)\alpha(\mathbf{R}) at the early stage of training can be controlled via transfer learning. However, for topological wavefunctions with complex phase windings, even transfer learning fails. [36] has proposed a density-current loss which matches probability and currents instead of fidelity to tackle the issue in the form of Eq. (10), where L1L_1 is the probability-matching term and L2L_2 is the current-matching term.

Lpre=L1+αL2,(10)L_{\mathrm{pre}} = L_1 + \alpha L_2, \tag*{(10)}

where α\alpha is a hyper-parameter. However, the above protocol may sometimes suffer from self-trapping during early stage of training. Since ∣Ψθ⟩|\Psi_\theta\rangle and ∣Φ⟩|\Phi\rangle are not normalized, the network can reduce the loss by matching amplitudes on a small set of configurations without moving probability mass globally, creating concentrated high-probability regions that freeze the walkers. To resolve this issue, we develop the following loss: The probability-matching term minimizes the difference between the probability densities of the two wavefunctions,

L1=min⁡c∈REpmix[(2log⁡[α(R)]−c)2],(11)L_1 = \min_{c\in\mathbb{R}}\mathbb{E}_{p_{\mathrm{mix}}}\left[(2\log[\alpha(\mathbf{R})]-c)^2\right], \tag*{(11)}

where pmix∝pθ+pt2p_{\mathrm{mix}} \propto\frac{p_\theta+p_t}{2}. Our improvement of L1L_1 is explicitly designed to mitigate self-trapping. Importantly, the constant cc effectively estimates the relative normalization factor between the two wavefunctions. By choosing c=Epmix[log⁡α(R)]c = \mathbb{E}_{p_{\mathrm{mix}}}[\log\alpha(\mathbf{R})], L1L_1 penalizes mismatches in probability density while remaining insensitive to overall normalization. The mixed sampling distribution pmixp_{\mathrm{mix}} incorporates information from the target wavefunction into the training process, thereby preventing self-trapping. For L2L_2, the current-matching term in [36] is used to penalize discrepancies in the per-particle phase gradients,

L2=Epθ[∑i∥∇Riψθ(R)−∇Riϕ(R)∥2].(12)L_2 = \mathbb{E}_{p_\theta}\left[\sum_i\left\lVert\nabla_{\mathbf{R}_i}\psi_\theta(\mathbf{R})-\nabla_{\mathbf{R}_i}\phi(\mathbf{R})\right\rVert^2\right]. \tag*{(12)}

For all loss functions used in this work, we use the kronecker-factored approximate curvature (KFAC) optimizer(Martens & Grosse [33]) to obtain parameter update. A detailed pseudo-code for the benchmarking protocol can be found in Alg. 1. We also discuss the relationship between observable error and fidelity in Appendix § 14.

Neural Network Wavefunction Architectures

We apply our dataset to two widely used NN wavefunction architectures, Ferminets (Pfau et al. [39]) and Psiformer (von Glehn et al. [49]). For both of the architectures, the input many body positions R\mathbf{R} is first embedded as features. For Ferminet features includes both one electron and two electron part, where as only one electron feature is used for Psiformer.

The features are then passed into an expressive contextual encoder, which outputs a per electron tensor horb∈RNe×Nhidh_{\mathrm{orb}} \in\mathbb{R}^{N_e \times N_{\mathrm{hid}}}. For Ferminet, the contextual encoder includes a one body and two body streams, and NhidN_{\mathrm{hid}} is defined as the width of the one body stream. As the original Ferminet concatenate the one body and two body streams, we use the SchNet([45]) enhanced version, where inter-stream mixing is achieved via continuous-filter convolutions. For Psiformer, the features are passed through multiple layers of self attention blocks + multilayer perceptron (MLP). Consecutively, and NhidN_{\mathrm{hid}} refers to the attention dimension of the self attention block.

For both network, horbh_{\mathrm{orb}} is projected to real and imaginary part of ϕθk\phi_\theta^k via linear layers, where k=1,…,Ndetk = 1, \ldots, N_{\mathrm{det}}. ϕθk\phi_\theta^k is also known as the neural orbitals. Finally, Ψθ(R)\Psi_\theta(\mathbf{R}) is generated by taking the sum of determinant of ϕ\phi:

Ψθ(R)=∑kNdetdet⁡[ϕθk](13)\Psi_\theta(\mathbf{R}) = \sum_{k}^{N_{\mathrm{det}}} \det[\phi_\theta^k] \tag*{(13)}

where ϕθk∈CNe×Ne\phi_\theta^k \in\mathbb{C}^{N_e \times N_e} (We assume ϕθk\phi_\theta^k is complex for all target wavefunctions for consistency). To impose physics prior to the orbitals, an envelope factor Ωi,jk\Omega_{i,j}^k can be included: ϕi,jk←Ωi,jϕi,jk\phi_{i,j}^k \leftarrow\Omega_{i,j} \phi_{i,j}^k. In particular, we choose an exponential decay envelope Ωi,j=exp⁡(−σi∣rj∣)\Omega_{i,j} = \exp(-\sigma_i|\mathbf{r}_j|) for all topological wavefunctions to take account of strong decay in wavefunction imposed by the Gaussian factor GG. We refer readers to the original works cited above for further details on the architecture. The detailed parameters for both networks used in this work are included in the appendix.

Results and Discussions

By applying WF-Bench to Psiformer and Ferminet, we probe numerically the scaling behavior of the fidelity with respect to three key parameters: NeN_e, NlayerN_{\mathrm{layer}}, and NdetN_{\mathrm{det}}.

Electron Number Scaling

We first present the scaling of fidelity with respect to NeN_e. We observe that the fidelity scaling of all wavefunctions can be well approximated by a power law: F=1−α(Ne−2)βF = 1 - \alpha(N_e - 2)^\beta. The fitting parameter α\alpha sets an overall scale fidelity for small electron number, wile β\beta quantifies the rate at which fidelity decays with increasing NeN_e, reflecting the correlation strength and phase complexity of the target wavefunction. The constant shift by 2 is introduced assuming F(Ne=2)F(N_e = 2) approaches unity, which is a reasonable assumption given nontrivial electron-electron correlations only emerge for Ne≥2N_e \ge2 in all wavefunctions in our dataset. In Figure 3, we show the the F(8)F(8) values across all 31 wavefunctions included in our datasets, spanning over topological states, superconductors and Wigner crystal. We found that across all wavefunctions included in our dataset, the performance of Psiformer exceeds Ferminet, despite relatively the same number of parameters (for topological states with Ne=10N_e = 10, Ferminet has 7.3×1057.3 \times10^5, whereas Psiformer has 8.3×1058.3 \times10^5 parameters.)

Bar chart showing the value of $F(N_e = 8)$ for 31 wavefunctions

Figure 3. The value of F(Ne=8)F(N_e = 8) for all 31 wavefunctions included in the dataset.

We emphasize that representational difficulty is jointly determined by both the prefactor α\alpha and the exponent β\beta: even when β\beta is small, a large α\alpha can induce substantial fidelity loss already at small system sizes. We therefore report F(8)F(8) as a practical summary metric of representation power.

In Figure 4, we select 9 representative wavefunctions from the three datasets, and plot the scaling of wavefunction fidelity FF with respect to the number of electrons. We observe that the fidelity scaling of all wavefunctions can be well approximated by a power law. The fitting parameters as well as the root mean square error (RMSE) of the fit can be found in Table 1.

Fidelity scaling of representative wavefunctions

Figure 4. Fidelity scaling of 9 representative wavefunctions from superconductors (blue), topological states (red), and Wigner crystals (green). (a–c) Fidelity versus NeN_e for Ferminet with Ndet=8N_{\mathrm{det}} = 8 and Nlayer=2N_{\mathrm{layer}} = 2. (d–f) Corresponding results for Psiformer with Ndet=8N_{\mathrm{det}} = 8 and Nlayer=2N_{\mathrm{layer}} = 2. Dotted lines indicate empirical fits of the form F=1−α(Ne−2)βF = 1 - \alpha(N_e - 2)^\beta.

NAMEFERMlNET (Nlayer=2N_{\mathrm{layer}} = 2, Ndet=8N_{\mathrm{det}} = 8)PSIFORMER (Nlayer=2N_{\mathrm{layer}} = 2, Ndet=8N_{\mathrm{det}} = 8)
F(8) ↑α\alphaβ\betaRMSEF(8) ↑α\alphaβ\betaRMSE
LAUGHLIN_M50.87375.66E-043.028.24E-040.99663.30E-052.601.04E-03
MOORE_M20.97775.27E-042.175.90E-030.99343.39E-052.813.55E-03
LAUGHLIN_M3H20.97071.24E-042.991.80E-030.99681.49E-053.019.88E-04
BCS_S0.99255.38E-052.771.21E-030.99546.50E-041.173.62E-04
BCS_FP_TP0.89022.25E-032.212.77E-020.94989.14E-042.208.97E-03
BCS_DXY0.97654.31E-053.561.04E-020.98431.38E-031.411.73E-03
WC_GAU_V1A0.99982.67E-062.213.41E-050.99995.06E-072.564.44E-05
WC_MOIRE_V1A0.99933.79E-051.612.87E-040.99963.85E-062.561.10E-04
WC_SGAU_V2AC0.99814.03E-040.911.52E-040.99958.57E-051.151.40E-04

Table 1. Fidelity fittings of the 9 representative wavefunctions. For each wavefunction, we report the fidelity at Ne=8N_e = 8, the fitted parameters (α\alpha, β\beta), and the RMSE.

For superconductors, we include bcs_s, bcs_dxy, and bcs_fp_tp. The first two are time reversal symmetric singlet states, whereas the latter is a chiral triplet state with complex amplitudes and nontrivial phase structure. As shown in Figure 4(a,Figure 4d), the fidelity decays the fastest for the chiral triplet states in both Psiformer and Ferminet, which is an expected result of wavefunction complexity.

For topological states, we include laughlin_m3h2, laughlin_m5, and moore_m2. The first two are Laughlin wavefunctions with different mm, while the latter is the Moore Read state. The laughlin_m3h2 wavefunction includes a quasihole excitation, making it more challenging to learn than the ground-state Laughlin wavefunction at m=3m = 3. As shown in Figure 4(b,Figure 4e), the fidelity scaling of topological wavefunctions exhibits a large exponent in the power-law tendency: fidelity remain high at small NeN_e but decay rapidly as NeN_e increases. This behavior reflects the strong electron-electron correlations in topological states, whose complexity grows combinatorially with system size. Increasing mm further amplifies correlation effects, leading to a faster decay for laughlin_m5 compared with other topological targets. Although laughlin_m3h2 has a smaller mm than laughlin_m5, the Pfaffian structure of the Moore Read state makes moore_m2 more difficult to represent with determinant based neural ansätze, resulting in a lower F(8)F(8).

Finally, we observe that due to the highly nontrivial phase structure of topological states, network fidelity can collapse if the representation capacity is insufficient to capture enough phase information during pretraining, as illustrated in Figure 3.

Lastly, for Wigner crystal states, we include wc_gau_v1A, wc_moire_v1A, and wc_sgau_v2AC. The first two correspond to ν=1\nu= 1 Wigner crystals with Gaussian and moiré orbitals, respectively, while the last represents a ν=2\nu= 2 Wigner crystal in a squeezed Gaussian orbital with A–C hexagonal lattice stacking. Among the three classes of wavefunctions, Wigner crystal states exhibit the slowest fidelity decay with increasing electron number. This behavior arises because electrons are strongly localized by the Coulomb interactions, which suppresses complex phase winding compared with topological and superconducting states. This trend is reflected in the large F(8)F(8) values reported in Table 1. Despite the overall weak correlations, we observe a faster fidelity decay for wc_moire_v1A due to the increased structural complexity of the moiré orbital, as can be seen in Figure 4(c,Figure 4f).

Ablation study on NdetN_{\mathrm{det}} and NlayerN_{\mathrm{layer}}

In addition to fidelity scaling with respect to NeN_e, we also perform ablation study on the impact of key expressivity knobs, including the number of determinants NdetN_{\mathrm{det}} and the number of layers NlayerN_{\mathrm{layer}}. These are generic architectural parameters for NN wavefunctions based on neural network backflow formalism with NN orbitals.

In Figure 5, we show the fidelity scaling of bcs_s and moore_m2 as a function of the number of determinants NdetN_{\mathrm{det}}. For small NdetN_{\mathrm{det}}, the fidelity increases rapidly, indicating a significant gain in representation power. However, once NdetN_{\mathrm{det}} exceeds ∼32\sim32, further increases yield only marginal improvements in fidelity. Since the computational cost scales linearly with NdetN_{\mathrm{det}}, using excessively large determinant expansions results in substantial computational overhead with limited performance improvement.

Fidelity scaling of bcs_s and moore_m2 with respect to $N_{\mathrm{det}}$

Figure 5. Fidelity scaling of bcs_s (left) and moore_m2 (right) with respect to NdetN_{\mathrm{det}}. Ferminet (red squares) is evaluated at Ne=10N_{\mathrm{e}} = 10, and Psiformer (blue triangles) is evaluated at Ne=14N_{\mathrm{e}} = 14. Both Psiformer and Ferminet are set as Nlayer=2N_{\mathrm{layer}} = 2.

In Figure 6, we present the fidelity scaling of bcs_s and moore_m2 as a function of the number of layers NlayerN_{\mathrm{layer}}. When Nlayer=1N_{\mathrm{layer}} = 1, we observe the fidelity achieved by network is very low. In particular, for moore_m2, both Psiformer and Ferminet fail to represent the wavefunction. When NlayerN_{\mathrm{layer}} increases from 1 to 2, the fidelity improves markedly for both wavefunctions. Beyond Nlayer=2N_{\mathrm{layer}} = 2, further increases yield only modest gains in fidelity, indicating that deeper architectures do not substantially enhance representation power.

Fidelity scaling of bcs_s and moore_m2 as a function of the number of layers $N_{\mathrm{layer}}$

Figure 6. Fidelity scaling of bcs_s (left) and moore_m2 (right) as a function of the number of layers NlayerN_{\mathrm{layer}}. Ferminet (red squares) is evaluated at Ne=10N_{\mathrm{e}} = 10, and Psiformer (blue triangles) is evaluated at Ne=14N_{\mathrm{e}} = 14. Both Psiformer and Ferminet are set as: Ndet=8N_{\mathrm{det}} = 8.

Conclusion

We introduce a comprehensive dataset, WF-Bench, for evaluating the representational power of neural network wavefunctions across physically distinct many body systems. The dataset spans three classes: topological states, superconductors, and Wigner crystals. It comprises 31 target wavefunctions covering a broad range of correlation structures, phase complexities, and anti-symmetrization mechanisms.

Applying this benchmark to two widely used architectures, FermiNet and PsiFormer, we systematically characterize how fidelity scales with system size, wavefunction class, and key architectural knobs. Across all targets, the achievable fidelity follows an empirical power law decay with increasing electron number. We further observe diminishing returns from commonly used architectural parameters: fidelity improves rapidly with the number of determinants at small values but saturates beyond a moderate threshold, while increasing network depth yields limited gains beyond a small number of layers.

Our comparison highlights clear differences in representational capacity between architectures: at comparable parameter counts, Psiformer consistently achieves higher fidelity than Ferminet across all target wavefunctions. More broadly, the dataset and benchmarking protocol provide a quantitative, architecture agnostic framework for evaluating neural network wavefunctions and can be readily applied to emerging models. In the meantime, we note that the optimization protocols adopted in this work can still be further improved, which could potentially affect the scaling behavior established in this work. We leave such investigations to future works.

We expect this framework to serve as a reference point for the design, benchmarking and theoretical analysis for new NN wavefunction architectures. We also envision this dataset as a community driven resource that can be extended to additional classes of wavefunctions, including lattice models under second quantized formalism, chemical molecules, and other strongly correlated quantum systems.

Acknowledgement

DL acknowledges support from Beijing Municipal Science and Technology Commission and Zhongguancun Science Park Administrative Committee (No. 20251090054).

Data Availability

A JAX implementation of target wavefunctions and loss functions is available at https://github.com/L0bsterkun/WF-Bench.

Impact Statement

This work presents a comprehensive dataset and benchmark protocol for representability scaling of neural network wavefunctions. There are many potential societal consequences of our work, none of which we feel must be specifically highlighted here.

Details on the WF-Bench

The WF-Bench dataset includes a total of 31 wavefunctions. The F(8)F(8) value, as well as the fitting parameters and fitting RMSE of each wavefunctions is included in Table. Table 6.

NameFERMINET (NLAYER=2N_{\mathrm{LAYER}} = 2, NDET=8N_{\mathrm{DET}} = 8) F(8)F(8)α\alphaβ\betaRMSEPSIFORMER (NLAYER=2N_{\mathrm{LAYER}} = 2, NDET=8N_{\mathrm{DET}} = 8) F(8)F(8)α\alphaβ\betaRMSE
LAUGHLIN_M10.99941.11E-052.133.21E-040.99993.40E-073.161.25E-04
LAUGHLIN_M30.99474.68E-052.672.21E-040.99967.26E-073.315.21E-05
LAUGHLIN_M50.87375.66E-043.028.24E-040.99663.30E-052.601.04E-03
LAUGHLIN_M3H10.99543.50E-064.064.17E-040.99891.28E-052.433.13E-04
LAUGHLIN_M3H20.97071.24E-042.991.80E-030.99681.49E-053.019.88E-04
LAUGHLIN_M3H30.97221.92E-042.721.79E-030.99613.92E-041.292.54E-04
MOORE_M20.97775.27E-042.175.90E-030.99343.39E-052.813.55E-03
MOORE_M40.89379.84E-053.956.62E-030.98151.23E-042.757.97E-03
MOORE_M60.70682.00E-032.767.69E-030.96812.74E-042.646.97E-03
MOORE_M2H10.89931.70E-043.528.40E-030.99787.12E-074.671.40E-03
MOORE_M2H20.95904.22E-042.492.95E-030.99342.26E-053.111.45E-03
MOORE_M2H30.94937.10E-042.228.23E-030.99335.93E-052.621.75E-03
BCS_S0.99255.38E-052.771.21E-030.99546.50E-041.173.62E-04
BCS_DXY0.97654.31E-053.561.04E-020.98431.38E-031.411.73E-03
BCS_DX2Y20.97952.47E-053.688.89E-030.98592.89E-042.103.69E-03
BCS_PX_T00.97441.77E-042.653.83E-030.98909.23E-041.291.22E-03
BCS_PP_T00.97032.32E-042.671.43E-030.98803.99E-041.828.49E-04
BCS_PX_TP0.96024.77E-042.552.49E-030.98029.59E-041.691.26E-03
BCS_PP_TP0.94599.13E-042.368.59E-030.97285.04E-042.177.11E-04
BCS_FXX2Y2_TP0.84011.34E-021.582.55E-020.93402.54E-031.838.32E-03
BCS_FXX2Y2_T00.93943.72E-042.841.41E-020.98062.43E-031.181.65E-03
BCS_FP_TP0.89022.25E-032.212.77E-020.94989.14E-042.208.97E-03
BCS_FP_T00.96091.02E-043.312.76E-020.96611.19E-031.813.85E-03
WC_GAU_V1A0.99982.67E-062.213.41E-050.99995.06E-072.564.44E-05
WC_GAU_V2AB0.99981.49E-062.386.27E-050.99994.63E-072.557.21E-05
WCS_GAU_V2AC0.99814.03E-040.911.52E-040.99958.57E-051.151.40E-04
WCS_GAU_V2BC0.99692.34E-041.451.36E-030.99885.19E-051.718.60E-04
WC_MOIRE_V1A0.99933.79E-051.612.87E-040.99963.85E-062.561.10E-04
WC_MOIRE_V2AB0.99944.84E-062.606.26E-040.99981.13E-051.644.89E-04
WC_MOIRE_V2AC0.99941.75E-063.151.61E-030.99991.21E-051.484.86E-04
WC_MOIRE_V2BC0.99934.77E-040.611.01E-030.99982.46E-040.471.08E-03

Table 6. Fidelity fittings of the 31 wavefunctions in WF-Bench. For each wavefunction, we report the fidelity at Ne=8N_e = 8, the fitted parameters (α\alpha, β\beta), and the root-mean-square error (RMSE).

Topological wavefunctions

For topological wavefunctions, we include 12 wavefunctions. The setting of each wavefunctions can be found in Table 2. We note that during the training for topological wavefunctions, we apply a pre-determinant exponential envelope to take account of the Gaussian factor G({zj})\mathcal{G}(\{z_j\}) in the target wavefunction. The Gaussian factor originates from the lowest Landau level (LLL) projection, which ensures the normalizability of the wavefunctions. We also acknowledge that alternative boundary conditions have also been used, for examples a torus.

NamemmNhN_hAntisymmetrizer
laughlin_m110-
laughlin_m330-
laughlin_m550-
laughlin_m3h131-
laughlin_m3h232-
laughlin_m3h333-
moore_m220Pfaffian
moore_m440Pfaffian
moore_m660Pfaffian
moore_m2h121Pfaffian
moore_m2h222Pfaffian
moore_m2h323Pfaffian

Table 2. Parameters for the topological wavefunctions.

Superconducting wavefunctions

For superconducting wavefunctions, we start by choosing a single-particle dispersion ϵ(k)\epsilon(\mathbf{k}). The mean-field Cooper pair orbital is defined as: [6]

g(k)≡vkuk=Δ(k)ϵ(k)+E(k)(14)g(\mathbf{k}) \equiv\frac{v_{\mathbf{k}}}{u_{\mathbf{k}}} = \frac{\Delta(\mathbf{k})}{\epsilon(\mathbf{k}) + E(\mathbf{k})} \tag*{(14)}

where E(k)=ϵ(k)2+∣Δ(k)∣2E(\mathbf{k}) = \sqrt{\epsilon(\mathbf{k})^2 + |\Delta(\mathbf{k})|^2} is the Bogoliubov quasiparticle dispersion. We use a parabolic dispersion ϵ(k)=12(∣k∣2−kF2)\epsilon(\mathbf{k}) = \frac{1}{2}(|\mathbf{k}|^2 - k_F^2) and choose kF=2πNeAk_F = \sqrt{\frac{2\pi N_e}{A}} for area AA. We choose AA such that kFk_F is fixed at π2\sqrt{\frac{\pi}{2}} across different NeN_e. The k\mathbf{k}-space pairing orbitals are then Fourier transformed to the pairing function we used in the antisymmetrized geminimal power (AGP) function, as in Eq. (5). For a periodic square box with length LL, the momentum k=2πL(nx,ny)\mathbf{k} = \frac{2\pi}{L}(n_x,n_y) is quantized with integers nx,ny∈Zn_x,n_y \in\mathbb{Z}. Numerically, we include all momentum with ∣k∣<3kF|\mathbf{k}| < 3k_F to ensure convergence. The detailed setting of each superconducting wavefunctions can be found in Table. Table 3.

NameΔk\Delta_k∣χ⟩|\chi\rangleAntisymmetrizer
bcs_sΔ0\Delta_012(∣↑↓⟩−∣↓↑⟩)\frac{1}{\sqrt{2}}(|\uparrow\downarrow\rangle-|\downarrow\uparrow\rangle)Determinant
bcs_dxyΔ0sin⁡kxsin⁡ky\Delta_0\sin k_x\sin k_y12(∣↑↓⟩−∣↓↑⟩)\frac{1}{\sqrt{2}}(|\uparrow\downarrow\rangle-|\downarrow\uparrow\rangle)Determinant
bcs_dx2y2Δ0(cos⁡kx−cos⁡ky)\Delta_0(\cos k_x-\cos k_y)12(∣↑↓⟩−∣↓↑⟩)\frac{1}{\sqrt{2}}(|\uparrow\downarrow\rangle-|\downarrow\uparrow\rangle)Determinant
bcs_px_t0Δ0sin⁡kx\Delta_0\sin k_x12(∣↑↓⟩+∣↓↑⟩)\frac{1}{\sqrt{2}}(|\uparrow\downarrow\rangle+|\downarrow\uparrow\rangle)Pfaffian
bcs_pp_t0Δ0(sin⁡kx+isin⁡ky)\Delta_0(\sin k_x+i\sin k_y)12(∣↑↓⟩+∣↓↑⟩)\frac{1}{\sqrt{2}}(|\uparrow\downarrow\rangle+|\downarrow\uparrow\rangle)Pfaffian
bcs_px_tpΔ0sin⁡kx\Delta_0\sin k_x∣↑↑⟩|\uparrow\uparrow\ranglePfaffian
bcs_pp_tpΔ0(sin⁡kx+isin⁡ky)\Delta_0(\sin k_x+i\sin k_y)∣↑↑⟩|\uparrow\uparrow\ranglePfaffian
bcs_fxx2y2_tpΔ0sin⁡kx(cos⁡kx−cos⁡ky)\Delta_0\sin k_x(\cos k_x-\cos k_y)∣↑↑⟩|\uparrow\uparrow\ranglePfaffian
bcs_fxx2y2_t0Δ0sin⁡kx(cos⁡kx−cos⁡ky)\Delta_0\sin k_x(\cos k_x-\cos k_y)12(∣↑↓⟩+∣↓↑⟩)\frac{1}{\sqrt{2}}(|\uparrow\downarrow\rangle+|\downarrow\uparrow\rangle)Pfaffian
bcs_fp_tpΔ0(sin⁡kx+icos⁡ky)3\Delta_0(\sin k_x+i\cos k_y)^3∣↑↑⟩|\uparrow\uparrow\ranglePfaffian
bcs_fp_t0Δ0(sin⁡kx+icos⁡ky)3\Delta_0(\sin k_x+i\cos k_y)^312(∣↑↓⟩+∣↓↑⟩)\frac{1}{\sqrt{2}}(|\uparrow\downarrow\rangle+|\downarrow\uparrow\rangle)Pfaffian

Table 3. Parameters for the superconducting wavefunctions.

Wigner crystal wavefunctions

We implement the Wigner crystal wavefunctions by choosing a form of orbital function. The Gaussian and squeezed Gaussian orbital included in our dataset takes the form of:

ϕjgau(ri)=exp⁡[−αr~i,j2],ϕjsgau(ri)=exp⁡[−α((r~x;i,j)2s2+(r~y;i,j)2s2)](15)\phi_j^{\mathrm{gau}}(\mathbf{r}_i) = \exp[-\alpha\tilde{r}_{i,j}^{2}], \qquad\phi_j^{\mathrm{sgau}}(\mathbf{r}_i) = \exp\left[-\alpha\left(\frac{(\tilde{r}_{x;i,j})^2}{s^2}+(\tilde{r}_{y;i,j})^2s^2\right)\right] \tag*{(15)}

where r~i,j\tilde{r}_{i,j} is the minimal image distance between ri\mathbf{r}_i and Rj\mathbf{R}_j. We choose s=0.5s = 0.5 for ϕjsgau(ri)\phi_j^{\mathrm{sgau}}(\mathbf{r}_i).

For the moiré orbital, we choose:

ϕjmoireˊ(ri)=exp⁡[−α(V(rij)−V(0))](16)\phi_j^{\mathrm{moiré}}(\mathbf{r}_i) = \exp[-\alpha(V(r_{ij}) - V(0))] \tag*{(16)}

where VV is the moiré potential:

Vmoire(r,θ)=−6cos⁡ϕ+8π2cos⁡ϕ r2+16π333sin⁡ϕsin⁡(3θ) r3−8π43cos⁡ϕ r4−16π593sin⁡ϕsin⁡(3θ) r5+16π6405cos⁡ϕ(10−cos⁡(6θ))r6.(17)V_{\mathrm{moire}}(r,\theta)=-6\cos\phi+8\pi^{2}\cos\phi\,r^{2}+\frac{16\pi^{3}}{3\sqrt{3}}\sin\phi\sin(3\theta)\,r^{3}-\frac{8\pi^{4}}{3}\cos\phi\,r^{4}-\frac{16\pi^{5}}{9\sqrt{3}}\sin\phi\sin(3\theta)\,r^{5}+\frac{16\pi^{6}}{405}\cos\phi\left(10-\cos(6\theta)\right)r^{6}. \tag*{(17)}

We choose α=10\alpha=10 for all 8 wavefunctions included. The lattice vector is choosen as a1=nx(1,0)\mathbf{a}_1=n_x(1,0), a2=ny(12,32)\mathbf{a}_2=n_y\left(\frac{1}{2},\frac{\sqrt{3}}{2}\right), where nxn_x and nyn_y is the number of unit cell included in the super lattice.

Details on Network and Optimization Parameters

Here, we introduce details on the network as well as optimization parameter used during the training process. The network parameters used is listed in Table 4. The KFAC optimizer’s parameter used is listed in Table 5

PsiformerFerminet
attention dimension64–
number of attention head4–
MLP hidden dims (1e)256512
MLP hidden dims (2e)–32
use layer normTrue–
user last layer–True
separate spin channel–False
electron convolution dimension–16
complex outputTrueTrue

Table 4. Network architecture parameters.

NameKFAC
batch size4096
damping damping1×10−41 \times10^{-4}
minimal damping1×10−61 \times10^{-6}
norm constraint1×10−31 \times10^{-3}
Momentum0.0
Curvature EMA0.95
Estimation modefisher exact

Table 5. Optimizer parameters. The learn rate is chosen as lr=α(11+tτ)lr = \alpha\left(\frac{1}{1+\frac{t}{\tau}}\right), where τ=1×10−4\tau= 1 \times10^{-4}. α\alpha is set as 1×10−31 \times10^{-3} for the first wavefunction in the transfer series, and 1×10−41 \times10^{-4} for the rest of the wavefunctions.

Details on the Loss Functions

Here, we detail the loss functions used in the benchmarking protocol.

Fidelity loss

. By choosing LF=−log⁡F=−log⁡[∣⟨Ψθ∣Φ⟩∣2⟨Ψθ∣Ψθ⟩⟨Φ∣Φ⟩]L_F=-\log F=-\log\left[\frac{|\langle\Psi_\theta|\Phi\rangle|^2}{\langle\Psi_\theta|\Psi_\theta\rangle\langle\Phi|\Phi\rangle}\right], we have:

Lfid=log⁡[∑R∣Ψθ(R)∣2]+log⁡[∑R∣Φ(R)∣2]−log⁡[∣∑RΨθ∗(R)Φ(R)∣2](18)L_{\mathrm{fid}}=\log\left[\sum_{\mathbf{R}}|\Psi_\theta(\mathbf{R})|^2\right]+\log\left[\sum_{\mathbf{R}}|\Phi(\mathbf{R})|^2\right]-\log\left[\left|\sum_{\mathbf{R}}\Psi_\theta^*(\mathbf{R})\Phi(\mathbf{R})\right|^2\right] \tag*{(18)}

Its gradient is then :

∂θLfid=∂θ[log⁡[∑R∣Ψθ(R)∣2]]−∂θ[log⁡[∑RΨθ∗(R)Φ(R)]2]=∂θ[log⁡[∑R∣Ψθ(R)∣2]]−2Re⁡{∂θ[log⁡[∑RΨθ∗(R)Φ(R)]]}(19)\begin{aligned} \partial_{\theta}L_{\mathrm{fid}} &= \partial_{\theta}\left[\log\left[\sum_{\mathbf{R}}|\Psi_{\theta}(\mathbf{R})|^{2}\right]\right]-\partial_{\theta}\left[\log\left[\sum_{\mathbf{R}}\Psi_{\theta}^{*}(\mathbf{R})\Phi(\mathbf{R})\right]^{2}\right] \\ &= \partial_{\theta}\left[\log\left[\sum_{\mathbf{R}}|\Psi_{\theta}(\mathbf{R})|^{2}\right]\right]-2\operatorname{Re}\left\{\partial_{\theta}\left[\log\left[\sum_{\mathbf{R}}\Psi_{\theta}^{*}(\mathbf{R})\Phi(\mathbf{R})\right]\right]\right\} \tag*{(19)} \end{aligned}

For the first part, we have:

∂θ[log⁡[∑R∣Ψθ(R)∣2]]=2NθRe⁡{Epθ[∂θlog⁡[Ψθ∗(R)]]}Nθ=2Re⁡{Epθ[∂θ[log⁡[Ψθ∗(R)]]]}(20)\partial_{\theta}\left[\log\left[\sum_{\mathbf{R}}|\Psi_{\theta}(\mathbf{R})|^{2}\right]\right] = \frac{2N_{\theta}\operatorname{Re}\left\{\mathbb{E}_{p_{\theta}}\left[\partial_{\theta}\log[\Psi_{\theta}^{*}(\mathbf{R})]\right]\right\}}{N_{\theta}} = 2\operatorname{Re}\left\{\mathbb{E}_{p_{\theta}}\left[\partial_{\theta}\left[\log[\Psi_{\theta}^{*}(\mathbf{R})]\right]\right]\right\} \tag*{(20)}

For the second part, we have:

−2Re⁡{∂θ[log⁡[∑RΨθ∗(R)Φ(R)]]}=−2Re⁡{∂θ[∑RΨθ∗(R)Φ(R)]∑RΨθ∗(R)Φ(R)}=−2Re⁡{Epθ[Φ(R)Ψθ(R)⋅∂θlog⁡[Ψθ∗(R)]]Epθ[Φ(R)Ψθ(R)]}(21)-2\operatorname{Re}\left\{\partial_{\theta}\left[\log\left[\sum_{\mathbf{R}}\Psi_{\theta}^{*}(\mathbf{R})\Phi(\mathbf{R})\right]\right]\right\} = -2\operatorname{Re}\left\{\frac{\partial_{\theta}\left[\sum_{\mathbf{R}}\Psi_{\theta}^{*}(\mathbf{R})\Phi(\mathbf{R})\right]}{\sum_{\mathbf{R}}\Psi_{\theta}^{*}(\mathbf{R})\Phi(\mathbf{R})}\right\} = -2\operatorname{Re}\left\{\frac{\mathbb{E}_{p_{\theta}}\left[\frac{\Phi(\mathbf{R})}{\Psi_{\theta}(\mathbf{R})}\cdot\partial_{\theta}\log[\Psi_{\theta}^{*}(\mathbf{R})]\right]}{\mathbb{E}_{p_{\theta}}\left[\frac{\Phi(\mathbf{R})}{\Psi_{\theta}(\mathbf{R})}\right]}\right\} \tag*{(21)}

Therefore, we have:

∂θLfid=2Re⁡{Epθ[[1−α(R)Epθ[α(R)]]⋅∂θlog⁡[Ψθ∗(R)]]}(22)\partial_{\theta}L_{\mathrm{fid}} = 2\operatorname{Re}\left\{\mathbb{E}_{p_{\theta}}\left[\left[1-\frac{\alpha(\mathbf{R})}{\mathbb{E}_{p_{\theta}}[\alpha(\mathbf{R})]}\right]\cdot\partial_{\theta}\log[\Psi_{\theta}^{*}(\mathbf{R})]\right]\right\} \tag*{(22)}

where α(R)≡Φ(R)/Ψθ(R)\alpha(\mathbf{R}) \equiv\Phi(\mathbf{R})/\Psi_{\theta}(\mathbf{R}) is the complex ratio of the target and network wavefunction. We note that an alternative loss function L=1−FL = 1 - F is also widely used. Nevertheless, as −∂log⁡F∂θ=−1F∂F∂θ-\frac{\partial\log F}{\partial\theta} = -\frac{1}{F}\frac{\partial F}{\partial\theta}, the two loss functions converge to the same value as F→1F \to1 and the constant factor can be absorbed into the learning rate.

Mixed sampling Probability Loss

The pretraining loss is defined as

Lpre=L1+αL2,(23)L_{\mathrm{pre}} = L_{1} + \alpha L_{2}, \tag*{(23)}

with α=1−e−τ/t\alpha= 1 - e^{-\tau/t}. Numerically, we choose τ=2×104\tau= 2 \times10^{4} for the first wavefunction in the transfer series, and τ=1×103\tau= 1 \times10^{3} for the rest of the wavefunctions. As the form of L2L_{2} is already shown in the main-body of the paper, here, we detail the derivations for L1L_{1}. We start from the symmetric Kullback–Leibler (KL) divergence, and define our loss function as:

L=12[KL(pθ∥pt)+KL(pt∥pθ)]=12[Epθ(ln⁡pθ−ln⁡pt)+Ept(ln⁡pt−ln⁡pθ)](24)\begin{aligned} L &= \frac{1}{2}\left[\mathrm{KL}(p_{\theta}\Vert p_{t})+\mathrm{KL}(p_{t}\Vert p_{\theta})\right] \\ &= \frac{1}{2}\left[\mathbb{E}_{p_{\theta}}(\ln p_{\theta}-\ln p_{t})+\mathbb{E}_{p_{t}}(\ln p_{t}-\ln p_{\theta})\right] \tag*{(24)} \end{aligned}

However, as ∣Ψθ⟩|\Psi_{\theta}\rangle and ∣Φ⟩|\Phi\rangle are not normalized, pθp_{\theta} and ptp_{t} are unknown. We define c≡ln⁡(NtNθ)c \equiv\ln\left(\frac{N_{t}}{N_{\theta}}\right), where NtN_{t} and NθN_{\theta} is the normalization factor of ∣Φ⟩|\Phi\rangle and ∣Ψθ⟩|\Psi_{\theta}\rangle, respectively. The loss function can be rewritten as:

L=12[Epθ[(ln⁡∣Ψθ(R)∣2−ln⁡∣Φ(R)∣2−c)2]+Ept[(ln⁡∣Φ(R)∣2−ln⁡∣Ψθ(R)∣2+c)2]](25)L = \frac{1}{2}\left[\mathbb{E}_{p_{\theta}}\left[\left(\ln|\Psi_{\theta}(\mathbf{R})|^{2}-\ln|\Phi(\mathbf{R})|^{2}-c\right)^{2}\right]+\mathbb{E}_{p_{t}}\left[\left(\ln|\Phi(\mathbf{R})|^{2}-\ln|\Psi_{\theta}(\mathbf{R})|^{2}+c\right)^{2}\right]\right] \tag*{(25)}

Using quadratic KL loss instead of linear gives:

L1=12[Epθ[(2α(R)−c)2]+Ept[(2α(R)+c)2]](26)L_1 = \frac{1}{2}\left[\mathbb{E}_{p_\theta}\left[(2\alpha(\mathbf{R})-c)^2\right] + \mathbb{E}_{p_t}\left[(2\alpha(\mathbf{R})+c)^2\right]\right] \tag*{(26)}

Then we have:

∂L∂c=12[2(Epθ[2α(R)]+Ept[2α(R)])+4c](27)\frac{\partial L}{\partial c} = \frac{1}{2}\left[2\left(\mathbb{E}_{p_\theta}[2\alpha(\mathbf{R})] + \mathbb{E}_{p_t}[2\alpha(\mathbf{R})]\right) + 4c\right] \tag*{(27)}

Setting ∂L∂c=0\frac{\partial L}{\partial c}=0 gives c∗=−12(Epθ[2α(R)]+Ept[2α(R)])c^*=-\frac{1}{2}\left(\mathbb{E}_{p_\theta}[2\alpha(\mathbf{R})]+\mathbb{E}_{p_t}[2\alpha(\mathbf{R})]\right), which is a numerically stable expectation variable in logarithmic space. Plugging c∗c^* back to Eq. (26) gives :

L1=Epmix[(2α(R)−c∗)2](28)L_1 = \mathbb{E}_{p_{\mathrm{mix}}}\left[(2\alpha(\mathbf{R})-c^*)^2\right] \tag*{(28)}

Moreover, as ∂L∂c∣c=c∗=0\left.\frac{\partial L}{\partial c}\right|_{c=c^*}=0:

ddθL1=∂L1∂θ+∂L1∂cdcdθ=∂L1∂θ(29)\frac{d}{d\theta}L_1 = \frac{\partial L_1}{\partial\theta} + \frac{\partial L_1}{\partial c}\frac{dc}{d\theta} = \frac{\partial L_1}{\partial\theta} \tag*{(29)}

which means cc can be treated as a constant during the backpropagation.

Relation between Fidelity and Observable error

As fidelity is used as the benchmark metric for wavefunction expressivity, it is necessary to discuss the gap between fidelity and errors in observables, which are physically important and relevant in different training scenarios (for example, variational Monte Carlo methods). In general, the following bound exists for any bounded operator O^\hat{O} and F=∣⟨ψ∣ϕ⟩∣2F = \lvert\langle\psi|\phi\rangle\rvert^2:

∣⟨ψ∣O^∣ψ⟩−⟨ϕ∣O^∣ϕ⟩∣≤2∥O^∥1−F(30)\lvert\langle\psi|\hat{O}|\psi\rangle- \langle\phi|\hat{O}|\phi\rangle\rvert\le2\lVert\hat{O}\rVert\sqrt{1-F} \tag*{(30)}

Therefore, the error in the expectation value of any operator with bounded norm is bounded by infidelity. We would also like to point out the observable error scaling law could be an interesting direction for future works.

References

  1. [1]Barrett, T. D., Malyshev, A., and Lvovsky, A. Autoregres-sive neural-network wavefunctions for ab initio quantum chemistry. Nature Machine Intelligence, 4(4):351–358, 2022.
  2. [2]Benedetti, M., Lloyd, E., Sack, S., and Fiorentini, M. Parameterized quantum circuits as machine learning models. Quantum science and technology, 4(4):043001, 2019.DOI
  3. [3]Carleo, G. and Troyer, M. Solving the quantum many-body problem with artificial neural networks. Science, 355(6325):602–606, 2017.DOI
  4. [4]Carleo, G., Nomura, Y., and Imada, M. Constructing exact representations of quantum many-body systems with deep neural networks. Nature communications, 9(1):5322, 2018.DOI
  5. [5]Chen, Z., Newhouse, L., Chen, E., Luo, D., and Soljacic, M. Antn: Bridging autoregressive neural networks and tensor networks for quantum many-body simulation. Advances in neural information processing systems, 36:450–476, 2023.arxiv.org/abs/2304.01996
  6. [6]De Gennes, P.-G. Superconductivity of metals and alloys. CRC press, 2018.
  7. [7]Deng, D.-L., Li, X., and Das Sarma, S. Quantum entanglement in neural network states. Physical Review X, 7(2):021021, 2017.DOI
  8. [8]Gao, N. and Günnemann, S. Neural pfaffians: Solving many many-electron schrödinger equations. Advances in Neural Information Processing Systems, 37:125336–125369, 2024.arxiv.org/abs/2405.14762
  9. [9]Gutiérrez, I. L. and Mendl, C. B. Real time evolution with neural-network quantum states. Quantum, 6:627, 2022.
  10. [10]Han, X. and Hartnoll, S. A. Deep quantum geometry of matrices. Physical Review X, 10(1):011069, 2020.DOI
  11. [12]Hibat-Allah, M., Ganahl, M., Hayward, L. E., Melko, R. G., and Carrasquilla, J. Recurrent neural network wave functions. Physical Review Research, 2(2):023358, 2020.DOI
  12. [13]Irikura, N. and Saito, H. Neural-network quantum states at finite temperature. Physical Review Research, 2(1):013284, 2020.DOI
  13. [14]Jiang, D., Wen, X., Chen, Y., Li, R., Fu, W., Pham, H. Q., Chen, J., He, D., Goddard III, W. A., Wang, L., et al. Neural scaling laws surpass chemical accuracy for the many-electron schrödinger equation. arXiv preprint arXiv:2508.02570, 2025.arxiv.org/abs/2508.02570
  14. [15]Kim, J., Pescia, G., Fore, B., Nys, J., Carleo, G., Gandolfi, S., Hjorth-Jensen, M., and Lovato, A. Neural-network quantum states for ultra-cold fermi gases. Communications Physics, 7(1):148, 2024.arxiv.org/abs/2305.08831
  15. [16]Kiyohara, S., Tsubaki, M., and Mizoguchi, T. Learning excited states from ground states by using an artificial neural network. Npj Computational Materials, 6(1):68, 2020.DOI
  16. [17]Laughlin, R. B. Anomalous quantum hall effect: An incompressible quantum fluid with fractionally charged excitations. Physical Review Letters, 50(18):1395, 1983.
  17. [18]Lee, C. K., Patil, P., Zhang, S., and Hsieh, C. Y. Neural-network variational quantum algorithm for simulating many-body dynamics. Physical Review Research, 3(2):023095, 2021.DOI
  18. [19]Levine, Y., Yakira, D., Cohen, N., and Shashua, A. Deep learning and quantum entanglement: Fundamental connections with implications to network design. arXiv preprint arXiv:1704.01552, 2017.arxiv.org/abs/1704.01552
  19. [20]Levine, Y., Sharir, O., Cohen, N., and Shashua, A. Quantum entanglement in deep learning architectures. Physical review letters, 122(6):065301, 2019.DOI
  20. [23]Lin, J., Luo, D., Yao, X., and Shanahan, P. E. Real-time dynamics of the schwinger model as an open quantum system with neural density operators. Journal of High Energy Physics, 2024(6):1–23, 2024.
  21. [24]Lou, W. T., Sutterud, H., Cassella, G., Foulkes, W. M. C., Knolle, J., Pfau, D., and Spencer, J. S. Neural wave functions for superfluids. Physical Review X, 14(2):021030, 2024.DOI
  22. [26]Luo, D. and Clark, B. K. Backflow transformations via neural networks for quantum many-body wave functions. Physical review letters, 122(22):226401, 2019.DOI
  23. [27]Luo, D., Carleo, G., Clark, B. K., and Stokes, J. Gauge equivariant neural networks for quantum lattice gauge theories. Physical review letters, 127(27):276402, 2021.DOI
  24. [28]Luo, D., Chen, Z., Carrasquilla, J., and Clark, B. K. Autoregressive neural network for simulating open quantum systems via a probabilistic formulation. Physical review letters, 128(9):090501, 2022a.
  25. [29]Luo, D., Yuan, S., Stokes, J., and Clark, B. K. Gauge equivariant neural networks for 2+ 1d u (1) gauge theory simulations in hamiltonian formulation. arXiv preprint arXiv:2211.03198, 2022b.arxiv.org/abs/2211.03198
  26. [30]Luo, D., Reddy, A. P., Devakul, T., and Fu, L. Artificial intelligence for artificial materials: moir’e atom. arXiv preprint arXiv:2303.08162, 2023.arxiv.org/abs/2303.08162
  27. [31]Luo, D., Dai, D. D., and Fu, L. Simulating moir’e quantum matter with neural network. arXiv preprint arXiv:2406.17645, 2024.arxiv.org/abs/2406.17645
  28. [32]Luo, D., Zaklama, T., and Fu, L. Solving fractional electron states in twisted mote 2 with deep neural network. arXiv preprint arXiv:2503.13585, 2025.arxiv.org/abs/2503.13585
  29. [33]Martens, J. and Grosse, R. Optimizing neural networks with kronecker-factored approximate curvature. In International conference on machine learning, pp. 2408–2417. PMLR, 2015.arxiv.org/abs/1503.05671
  30. [34]Min-Gang, Z., Liu, Z., Yin, H.-L., Li, C., Tong-kai, X., and Chen, Z.-B. Quantum neural network for quantum neural computing. Research, 6, 2023.DOI
  31. [35]Moore, G. and Read, N. Nonabelions in the fractional quantum hall effect. Nuclear Physics B, 360(2-3):362–396, 1991.DOI
  32. [36]Nazaryan, K., Gaggioli, F., Teng, Y., and Fu, L. Artificial intelligence for quantum matter: Finding a needle in a haystack. arXiv preprint arXiv:2507.13322, 2025.arxiv.org/abs/2507.13322
  33. [37]Paul, N. Bound on entanglement in neural quantum states. arXiv preprint arXiv:2510.11797, 2025.DOI
  34. [38]Pescia, G., Nys, J., Kim, J., Lovato, A., and Carleo, G. Message-passing neural quantum states for the homogeneous electron gas. Physical Review B, 110(3):035108, 2024.DOI
  35. [39]Pfau, D., Spencer, J. S., Matthews, A. G., and Foulkes, W. M. C. Ab initio solution of the many-electron schrödinger equation with deep neural networks. Physical review research, 2(3):033429, 2020.DOI
  36. [40]Pfau, D., Axelrod, S., Sutterud, H., von Glehn, I., and Spencer, J. S. Accurate computation of quantum excited states with neural networks. Science, 385(6711):eadn0137, 2024.DOI
  37. [41]Quek, Y., Fort, S., and Ng, H. K. Adaptive quantum state tomography with neural networks. npj Quantum Information, 7(1):105, 2021.arxiv.org/abs/1812.06693
  38. [42]Ren, W., Fu, W., Wu, X., and Chen, J. Towards the ground state of molecules via diffusion monte carlo on neural networks. Nature Communications, 14(1):1860, 2023.DOI
  39. [43]Rende, R., Sinibaldi, A., Viteritti, L. L., Wiersema, R., Georges, A., and Carleo, G. Scaling laws for neural-network quantum states. arXiv preprint arXiv:2606.02794, 2026.arxiv.org/abs/2606.02794
  40. [44]Robledo Moreno, J., Carleo, G., Georges, A., and Stokes, J. Fermionic wave functions from neural-network constrained hidden states. Proceedings of the National Academy of Sciences, 119(32):e2122059119, 2022.DOI
  41. [45]Schütt, K. T., Sauceda, H. E., Kindermans, P.-J., Tkatсhenko, A., and Müller, K.-R. Schnet–a deep learning architecture for molecules and materials. The Journal of chemical physics, 148(24), 2018.DOI
  42. [46]Sharir, O., Levine, Y., Wies, N., Carleo, G., and Shashua, A. Deep autoregressive models for the efficient variational simulation of many-body quantum systems. Physical review letters, 124(2):020503, 2020.DOI
  43. [47]Shen, H., Zhang, P., You, Y.-Z., and Zhai, H. Information scrambling in quantum neural networks. Physical Review Letters, 124(20):200504, 2020.DOI
  44. [48]Torlai, G., Mazzola, G., Carrasquilla, J., Troyer, M., Melko, R., and Carleo, G. Neural-network quantum state tomography. Nature physics, 14(5):447–450, 2018.
  45. [49]von Glehn, I., Spencer, J. S., and Pfau, D. A self-attention ansatz for ab-initio quantum chemistry. arXiv preprint arXiv:2211.13672, 2022.arxiv.org/abs/2211.13672
  46. [50]Wang, J.-Q., Wu, H.-Q., He, R.-Q., and Lu, Z.-Y. Variational optimization of the amplitude of neural-network quantum many-body ground states. Physical Review B, 109(24):245120, 2024.DOI
  47. [51]Wigner, E. On the interaction of electrons in metals. Physical Review, 46(11):1002, 1934.
  48. [52]Xie, H., Li, Z.-H., Wang, H., Zhang, L., and Wang, L. Deep variational free energy approach to dense hydrogen. Physical Review Letters, 131(12):126501, 2023.DOI
  49. [53]Yang, T.-H., Soleimanifar, M., Bergamaschi, T., and Preskill, J. When can classical neural networks represent quantum states? arXiv preprint arXiv:2410.23152, 2024.arxiv.org/abs/2410.23152
  50. [54]Zhang, L. and Luo, D. Neural transformer backflow for solving momentum-resolved ground states of strongly correlated materials. arXiv preprint arXiv:2509.09275, 2025.arxiv.org/abs/2509.09275

Paper details

vlnc:W_268w3b9r
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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