Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2609.26544

Neural Network Backflow with Low-Rank Multi-Determinant Updates

Tianshu Huang, Xiaowei Ou, Vidvuds Ozolins

cond-mat.str-elcond-mat.supr-conquant-phSubmitted Sep 22, 2026• Announced Sep 23, 2026• 0 citations
arXivHTMLPDF

Abstract

Simulating strongly correlated fermions remains a long-standing challenge due to the exponential complexity of the Hilbert space and the intricate sign structure of many-body wavefunctions. We introduce a variational framework centered on a neural network backflow transformation that combines deep learning with variational Monte Carlo. The proposed ansatz employs a multi-determinant expansion with low-rank shifts to capture non-local correlations and complex sign structures. Applied to the two-dimensional Hubbard model at both half-filling and $1/8$ doping, the method achieves energies within $0.45\%$ of auxiliary-field quantum Monte Carlo at half-filling and captures intertwined charge- and spin-density stripe patterns at $1/8$ doping. These results demonstrate the potential of this framework as a scalable and interpretable approach to variational simulations of strongly correlated fermionic systems.

AI Summary

The paper addresses the computational cost of multi-determinant neural-network backflow wavefunctions for strongly correlated fermions. It constructs a variational ansatz with a shared deep backflow network, configuration-dependent softmax coefficients, and determinant-specific low-rank shifts to both orbital and backflow-modulation matrices. The matrix determinant lemma and reused LU factorizations reduce determinant-evaluation cost from $O(K_MK_FN_e^3)$ to $O(K_FN_e^3+K_MK_FN_e^2r)$, where $K_M,K_F$ count shifts, $N_e$ is electron number, and $r$ is update rank. Reported training-time gains reach approximately tenfold on an NVIDIA H200 GPU. Benchmarks use variational Monte Carlo with SPRING optimization for the periodic two-dimensional Hubbard model. At half-filling and $U/t=8$, energies per site are $-0.5268$, $-0.5245$, and $-0.5231$ on $6\times6$, $8\times8$, and $10\times10$ lattices; the abstract reports agreement within $0.45\%$ of AFQMC. At $1/8$ doping, a post-training Lanczos step gives energies $-0.7671$, $-0.7662$, and $-0.7622$ on $4\times8$, $4\times12$, and $4\times16$ lattices, respectively. The largest doped system exhibits coupled charge and spin stripes; pairing correlations beyond the first shell are statistically consistent with zero. Architectural comparisons, sampled determinant Gram matrices, singular-value spectra, and PCA examine basis redundancy and learned sign representations. The paper notes that PCA sign separation does not independently establish correct ground-state signs, and reduced pairwise similarity does not establish basis orthogonality. Reported energy error bars measure sampling uncertainty at fixed parameters, not training-run variability. Low-rank shifts trade expressivity for parameter and evaluation efficiency.

Likely largely AI-generatedIncrementalUnclear novelty

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Comments
16 pages, 13 figures

Related papers

Enhancing Neural Network Backflow

Kieran Loehr, Bryan K. Clark

Accurately describing the ground state of strongly correlated systems is essential for understanding their emergent properties. Neural Network Backflow (NNBF) is a powerful variational ansatz that enhances mean-field wave functions by introducing configuration-dependent modifications to single-particle orbitals. Although NNBF is theoretically universal in the limit of large networks, we find that practical gains saturate with increasing network size. Instead, significant improvements can be achieved by using a multi-determinant ansatz. We explore efficient ways to generate these multi-determinant expansions without increasing the number of variational parameters. In particular, we study single-step Lanczos and symmetry projection techniques, benchmarking their performance against diffusion Monte Carlo and NNBF applied to alternative mean fields. Benchmarking on a doped periodic square Hubbard model near optimal doping, we find that a Lanczos step, diffusion Monte Carlo, and projection onto a symmetry sector all give similar improvements achieving state-of-the-art energies at minimal cost. By further optimizing the projected symmetrized states directly, we gain significantly in energy. Using this technique we report the lowest variational energies for this Hamiltonian on $4\times 16$ and $4 \times 8$ lattices as well as accurate variance extrapolated energies. We also show the evolution of spin, charge, and pair correlation functions as the quality of the variational ansatz improves.

arXiv:2510.26906•Oct 30, 2025•10 citations

Neural network based nodal structure optimization for interacting fermionic systems

William Freitas, Bruno Ricardi de Abreu, S. A. Vitiello

Simulating strongly correlated fermionic systems remains a fundamental challenge in quantum physics, largely due to the sign problem in quantum Monte Carlo (QMC) methods. We present a neural network-based variational Monte Carlo (NN-VMC) approach, leveraging a flexible neural network ansatz to represent the many-body wavefunction. Focusing on quantum dots with up to 30 electrons, we demonstrate that NN-VMC significantly reduces variational bias and achieves ground-state energies surpassing those of fixed-node diffusion Monte Carlo (DMC). A key feature is that the neural network adaptively learns and optimizes nodal structures during energy minimization. We provide qualitative insights into the nodal structure of fermionic wavefunctions by comparing the nodal structures generated by NN-VMC with those obtained from traditional trial functions. Additionally, we reveal spin-resolved radial distributions and electron density profiles, highlighting the versatility and accuracy of NN-VMC. This work underscores the potential of machine learning to advance quantum simulations and deepen our understanding of strongly correlated systems.

arXiv:2411.02257•Physical review. B./Physical review. B•Sep 9, 2025•2 citations

Pareto Frontier of Neural Quantum States: Scalable, Affordable, and Accurate Convolutional Backflow for Strongly Correlated Lattice Fermions

Yuntian Gu, Zeyao Han, Wenrui Li +5 more

The recent development of Neural Quantum States (NQS) has established them as one of the most accurate methods for studying strongly correlated many-fermion systems, outperforming existing many-body approaches for large systems. However, NQS calculations are currently highly resource intensive. In this work, we introduce an improved Pareto frontier of efficiency and accuracy for NQS in the simulation of strongly correlated lattice fermions. This frontier is defined by two complementary backflow-related architectures: the Sparse Convolutional Ansatz for Lattice Electrons (SCALE), which achieves high computational efficiency, and the Accurate Convolutional ansatz for lattice Electrons (ACE), which delivers high variational accuracy, based on benchmark results on the Hubbard model and the t – J model on large lattice sizes. SCALE utilizes a tailored convolutional design to enable an efficient local update based on the low-rank update of determinants. This structure reduces the computational scaling from O ( N 4 ) to O ( N 3 ) in backflow methods, resulting in a practical speedup exceeding 40 × in our test cases while retaining accurate variational accuracy. As an application, we study the 1 / 8 -doped pure Hubbard model up to size 32 × 32 and find no significant energy difference between the horizontal and vertical unit-filled stripe states, in contrast to the half-filled stripe state when next-nearest-neighbor hoppings are included. ACE, conversely, employs a deep convolutional stack to maximize expressive power, achieving high accuracy on large systems. We perform extensive benchmark calculations on the Hubbard and t – J models with the two architectures. On the challenging repulsive Hubbard model, SCALE achieves variational energies competitive with leading methods at a fraction of the computational cost. Meanwhile, ACE achieves lower variational energies than previously reported results while requiring only one-sixth of the run-time for system size 16 × 4 . These NQS approaches provide , , and tools for exploring the rich physics of strongly correlated fermionic systems.

arXiv:2604.25775•PRX Intelligence•Oct 6, 2026•7 citations

Superconductivity in the $t$-$t'$ Hubbard Model from Symmetry-Preserving Neural-Network Quantum States

Riccardo Rende, Luciano Loris Viteritti, Antoine Georges

Despite its fundamental importance in the theory of strongly correlated electrons, the nature of the ground state of the two-dimensional doped Hubbard model remains intensely debated. Variational approaches provide a powerful route to this problem, but their conclusions can depend sensitively on the chosen wave-function parameterization, the mean-field initialization, or the pinning fields used to guide the optimization, as well as on boundary conditions. This can favor one type of symmetry breaking over another, making it difficult to distinguish the genuine interplay of intertwined or competing orders from biases induced by the variational parameterization. Here, we introduce the Symmetry-Preserving Backflow Pairing (SBP) ansatz, a neural-network wave function that respects translational symmetry by construction and thereby avoids these broken-symmetry minima. The SBP ansatz reaches state-of-the-art variational energies for the $t$-$t'$ Hubbard model on lattices up to $24\times24$ with $504$ electrons, below those of competing pure stripe solutions. By extrapolating to the thermodynamic limit, we find robust evidence for $d$-wave superconducting order, resolving a long-standing question about the $1/8$-doped model at $t'/t=-0.2$ and $U/t=8.0$. Built on general principles of symmetry and locality, the SBP wave function provides a broadly applicable variational representation for challenging interacting fermionic systems.

cond-mat.str-elCross-listedarXiv:2608.12465•Aug 14, 2026•9 citations

Optimization dynamics of Transformer backflow neural quantum states for the two-dimensional Hubbard model

Zong-Yu Liao, Jia-Qi Wang, Rong-Qiang He +1 more

Building on the multi-determinant Transformer backflow neural quantum state (NQS) ansatz and the associated multi-stage training workflow for the doped two-dimensional Hubbard model, we investigate how the optimization dynamics of the NQS depend on several key optimization and architectural hyperparameters. The workflow consists of neural-network backflow (NNB) initialization, supervised Transformer pre-training, and main energy optimization using the Moment-Adaptive ReConfiguration Heuristic (MARCH) within variational Monte Carlo. Using the doped $4\times4$ periodic Hubbard model at $U=8$ as a baseline, we examine how the update-norm threshold, Transformer width, number of determinant channels, and Monte Carlo batch size affect convergence. We find that a moderate update constraint improves the efficiency of MARCH optimization, larger Transformer width and more determinant channels improve the expressive capacity of the ansatz, and larger Monte Carlo batches reduce sampling noise in the update direction. We further test the same workflow at half filling, weaker interaction strength, open boundary conditions, and on a larger $8\times8$ doped lattice. These results identify practical optimization trends for Transformer backflow NQSs and highlight the balance between ansatz expressivity, MARCH update stability, and Monte Carlo sampling quality.

arXiv:2607.14875•Jul 16, 2026•0 citations

Backflow Transformations via Neural Networks for Quantum Many-Body Wave Functions

Di Luo, Bryan K. Clark

Obtaining an accurate ground state wave function is one of the great challenges in the quantum many-body problem. In this Letter, we propose a new class of wave functions, neural network backflow (NNB). The backflow approach, pioneered originally by Feynman and Cohen [Phys. Rev. 102, 1189 (1956)10.1103/PhysRev.102.1189], adds correlation to a mean-field ground state by transforming the single-particle orbitals in a configuration-dependent way. NNB uses a feed-forward neural network to learn the optimal transformation via variational Monte Carlo calculations. NNB directly dresses a mean-field state, can be systematically improved, and directly alters the sign structure of the wave function. It generalizes the standard backflow [L. F. Tocchio et al., Phys. Rev. B 78, 041101(R) (2008)10.1103/PhysRevB.78.041101], which we show how to explicitly represent as a NNB. We benchmark the NNB on Hubbard models at intermediate doping, finding that it significantly decreases the relative error, restores the symmetry of both observables and single-particle orbitals, and decreases the double-occupancy density. Finally, we illustrate interesting patterns in the weights and bias of the optimized neural network.

arXiv:1807.10770•Physical Review Letters•Jun 4, 2019•216 citations
Simulating strongly correlated fermions remains a long-standing challenge due to the exponential complexity of the Hilbert space and the intricate sign structure of many-body wavefunctions. We introduce a variational framework centered on a neural network backflow transformation that combines deep learning with variational Monte Carlo. The proposed ansatz employs a multi-determinant expansion with low-rank shifts to capture non-local correlations and complex sign structures. Applied to the two-dimensional Hubbard model at both half-filling and 1/81/81/8 doping, the method achieves energies within 0.45%0.45\%0.45% of auxiliary-field quantum Monte Carlo at half-filling and captures intertwined charge- and spin-density stripe patterns at 1/81/81/8 doping. These results demonstrate the potential of this framework as a scalable and interpretable approach to variational simulations of strongly correlated fermionic systems.
The paper addresses the computational cost of multi-determinant neural-network backflow wavefunctions for strongly correlated fermions. It constructs a variational ansatz with a shared deep backflow network, configuration-dependent softmax coefficients, and determinant-specific low-rank shifts to both orbital and backflow-modulation matrices. The matrix determinant lemma and reused LU factorizations reduce determinant-evaluation cost from O(KMKFNe3)O(K_MK_FN_e^3)O(KM​KF​Ne3​) to O(KFNe3+KMKFNe2r)O(K_FN_e^3+K_MK_FN_e^2r)O(KF​Ne3​+KM​KF​Ne2​r), where KM,KFK_M,K_FKM​,KF​ count shifts, NeN_eNe​ is electron number, and rrr is update rank. Reported training-time gains reach approximately tenfold on an NVIDIA H200 GPU.
Benchmarks use variational Monte Carlo with SPRING optimization for the periodic two-dimensional Hubbard model. At half-filling and U/t=8U/t=8U/t=8, energies per site are −0.5268-0.5268−0.5268, −0.5245-0.5245−0.5245, and −0.5231-0.5231−0.5231 on 6×66\times66×6, 8×88\times88×8, and 10×1010\times1010×10 lattices; the abstract reports agreement within 0.45%0.45\%0.45% of AFQMC. At 1/81/81/8 doping, a post-training Lanczos step gives energies −0.7671-0.7671−0.7671, −0.7662-0.7662−0.7662, and −0.7622-0.7622−0.7622 on 4×84\times84×8, 4×124\times124×12, and 4×164\times164×16 lattices, respectively. The largest doped system exhibits coupled charge and spin stripes; pairing correlations beyond the first shell are statistically consistent with zero.
Architectural comparisons, sampled determinant Gram matrices, singular-value spectra, and PCA examine basis redundancy and learned sign representations. The paper notes that PCA sign separation does not independently establish correct ground-state signs, and reduced pairwise similarity does not establish basis orthogonality. Reported energy error bars measure sampling uncertainty at fixed parameters, not training-run variability. Low-rank shifts trade expressivity for parameter and evaluation efficiency.
Accurately describing the ground state of strongly correlated systems is essential for understanding their emergent properties. Neural Network Backflow (NNBF) is a powerful variational ansatz that enhances mean-field wave functions by introducing configuration-dependent modifications to single-particle orbitals. Although NNBF is theoretically universal in the limit of large networks, we find that practical gains saturate with increasing network size. Instead, significant improvements can be achieved by using a multi-determinant ansatz. We explore efficient ways to generate these multi-determinant expansions without increasing the number of variational parameters. In particular, we study single-step Lanczos and symmetry projection techniques, benchmarking their performance against diffusion Monte Carlo and NNBF applied to alternative mean fields. Benchmarking on a doped periodic square Hubbard model near optimal doping, we find that a Lanczos step, diffusion Monte Carlo, and projection onto a symmetry sector all give similar improvements achieving state-of-the-art energies at minimal cost. By further optimizing the projected symmetrized states directly, we gain significantly in energy. Using this technique we report the lowest variational energies for this Hamiltonian on 4×164\times 164×16 and 4×84 \times 84×8 lattices as well as accurate variance extrapolated energies. We also show the evolution of spin, charge, and pair correlation functions as the quality of the variational ansatz improves.
Simulating strongly correlated fermionic systems remains a fundamental challenge in quantum physics, largely due to the sign problem in quantum Monte Carlo (QMC) methods. We present a neural network-based variational Monte Carlo (NN-VMC) approach, leveraging a flexible neural network ansatz to represent the many-body wavefunction. Focusing on quantum dots with up to 30 electrons, we demonstrate that NN-VMC significantly reduces variational bias and achieves ground-state energies surpassing those of fixed-node diffusion Monte Carlo (DMC). A key feature is that the neural network adaptively learns and optimizes nodal structures during energy minimization. We provide qualitative insights into the nodal structure of fermionic wavefunctions by comparing the nodal structures generated by NN-VMC with those obtained from traditional trial functions. Additionally, we reveal spin-resolved radial distributions and electron density profiles, highlighting the versatility and accuracy of NN-VMC. This work underscores the potential of machine learning to advance quantum simulations and deepen our understanding of strongly correlated systems.
The recent development of Neural Quantum States (NQS) has established them as one of the most accurate methods for studying strongly correlated many-fermion systems, outperforming existing many-body approaches for large systems. However, NQS calculations are currently highly resource intensive. In this work, we introduce an improved Pareto frontier of efficiency and accuracy for NQS in the simulation of strongly correlated lattice fermions. This frontier is defined by two complementary backflow-related architectures: the Sparse Convolutional Ansatz for Lattice Electrons (SCALE), which achieves high computational efficiency, and the Accurate Convolutional ansatz for lattice Electrons (ACE), which delivers high variational accuracy, based on benchmark results on the Hubbard model and the t – J model on large lattice sizes. SCALE utilizes a tailored convolutional design to enable an efficient local update based on the low-rank update of determinants. This structure reduces the computational scaling from O ( N 4 ) to O ( N 3 ) in backflow methods, resulting in a practical speedup exceeding 40 × in our test cases while retaining accurate variational accuracy. As an application, we study the 1 / 8 -doped pure Hubbard model up to size 32 × 32 and find no significant energy difference between the horizontal and vertical unit-filled stripe states, in contrast to the half-filled stripe state when next-nearest-neighbor hoppings are included. ACE, conversely, employs a deep convolutional stack to maximize expressive power, achieving high accuracy on large systems. We perform extensive benchmark calculations on the Hubbard and t – J models with the two architectures. On the challenging repulsive Hubbard model, SCALE achieves variational energies competitive with leading methods at a fraction of the computational cost. Meanwhile, ACE achieves lower variational energies than previously reported results while requiring only one-sixth of the run-time for system size 16 × 4 . These NQS approaches provide , , and tools for exploring the rich physics of strongly correlated fermionic systems.
Superconductivity in the ttt-t′t't′ Hubbard Model from Symmetry-Preserving Neural-Network Quantum States
Despite its fundamental importance in the theory of strongly correlated electrons, the nature of the ground state of the two-dimensional doped Hubbard model remains intensely debated. Variational approaches provide a powerful route to this problem, but their conclusions can depend sensitively on the chosen wave-function parameterization, the mean-field initialization, or the pinning fields used to guide the optimization, as well as on boundary conditions. This can favor one type of symmetry breaking over another, making it difficult to distinguish the genuine interplay of intertwined or competing orders from biases induced by the variational parameterization. Here, we introduce the Symmetry-Preserving Backflow Pairing (SBP) ansatz, a neural-network wave function that respects translational symmetry by construction and thereby avoids these broken-symmetry minima. The SBP ansatz reaches state-of-the-art variational energies for the ttt-t′t't′ Hubbard model on lattices up to 24×2424\times2424×24 with 504504504 electrons, below those of competing pure stripe solutions. By extrapolating to the thermodynamic limit, we find robust evidence for ddd-wave superconducting order, resolving a long-standing question about the 1/81/81/8-doped model at t′/t=−0.2t'/t=-0.2t′/t=−0.2 and U/t=8.0U/t=8.0U/t=8.0. Built on general principles of symmetry and locality, the SBP wave function provides a broadly applicable variational representation for challenging interacting fermionic systems.
Building on the multi-determinant Transformer backflow neural quantum state (NQS) ansatz and the associated multi-stage training workflow for the doped two-dimensional Hubbard model, we investigate how the optimization dynamics of the NQS depend on several key optimization and architectural hyperparameters. The workflow consists of neural-network backflow (NNB) initialization, supervised Transformer pre-training, and main energy optimization using the Moment-Adaptive ReConfiguration Heuristic (MARCH) within variational Monte Carlo. Using the doped 4×44\times44×4 periodic Hubbard model at U=8U=8U=8 as a baseline, we examine how the update-norm threshold, Transformer width, number of determinant channels, and Monte Carlo batch size affect convergence. We find that a moderate update constraint improves the efficiency of MARCH optimization, larger Transformer width and more determinant channels improve the expressive capacity of the ansatz, and larger Monte Carlo batches reduce sampling noise in the update direction. We further test the same workflow at half filling, weaker interaction strength, open boundary conditions, and on a larger 8×88\times88×8 doped lattice. These results identify practical optimization trends for Transformer backflow NQSs and highlight the balance between ansatz expressivity, MARCH update stability, and Monte Carlo sampling quality.
Obtaining an accurate ground state wave function is one of the great challenges in the quantum many-body problem. In this Letter, we propose a new class of wave functions, neural network backflow (NNB). The backflow approach, pioneered originally by Feynman and Cohen [Phys. Rev. 102, 1189 (1956)10.1103/PhysRev.102.1189], adds correlation to a mean-field ground state by transforming the single-particle orbitals in a configuration-dependent way. NNB uses a feed-forward neural network to learn the optimal transformation via variational Monte Carlo calculations. NNB directly dresses a mean-field state, can be systematically improved, and directly alters the sign structure of the wave function. It generalizes the standard backflow [L. F. Tocchio et al., Phys. Rev. B 78, 041101(R) (2008)10.1103/PhysRevB.78.041101], which we show how to explicitly represent as a NNB. We benchmark the NNB on Hubbard models at intermediate doping, finding that it significantly decreases the relative error, restores the symmetry of both observables and single-particle orbitals, and decreases the double-occupancy density. Finally, we illustrate interesting patterns in the weights and bias of the optimized neural network.