Numerically simulating quantum many-body systems is one of the central challenges in modern physics, as the exponential growth of the Hilbert space with system size renders large-scale exact solutions intractable. Over the past decade, generative artificial neural networks have emerged as powerful wave function ansätze that have powered advances in finding ground state representations, reconstructing quantum states from data, and modeling dynamics in quantum many-body systems. While recurrent neural networks (RNNs) are computationally efficient, the recurrent nature can limit their capacity to capture long-range correlations. Transformer models with their underlying self-attention mechanism excel at capturing these long-range correlations, but incur increased computational requirements, restricting applications to small physical systems. In this work, we address these limitations by exploring the multi-layer perceptron (MLP) mixer model as a wave function ansatz. By mixing information globally and locally through the system via a computationally lightweight mixing strategy, this architecture promises to combine the ability to capture long-range correlations with low computational costs. We benchmark the model on finding ground state representations of two-dimensional quantum many-body systems in various phases of matter and investigate the behavior of representational accuracy and computational efficiency for different physical system sizes. By comparing its performance to RNNs and transformer models with single-site and patched input definitions, we demonstrate that the MLP mixer architecture performs competitively with state-of-the-art network models. We further show that a hierarchical combination of this model with the transformer model outperforms other state-of-the-art architectures for large-scale quantum systems, opening new avenues for investigating quantum phases of matter.
AI Summary
The paper studies whether autoregressive MLP-mixer neural wave functions can combine long-range correlation modeling with lower computational cost than transformers. It adapts the mixer using causal masks, shifted inputs, channel-only normalization, and probability outputs, and constructs hierarchical architectures in which a coarse network conditions a second network that generates smaller patches.
Nine architectures—including sitewise and patched RNNs and transformers, mixers, and hierarchical combinations—are compared through variational ground-state optimization. Benchmarks cover the two-dimensional transverse-field Ising model on $4\times4$ and $8\times8$ lattices and Rydberg arrays up to $24\times24$ sites, with open boundaries. Each system is tested at three parameter values spanning ordered, disordered, and near-transition regimes, with five independent runs. Energy errors are measured against tensor-network or quantum Monte Carlo references; network parameter counts range from approximately $5{,}000$ to $30{,}000$.
Patched RNNs provide the reported best accuracy–cost tradeoff on the smallest systems. Hierarchical architectures become preferable at larger sizes; the transformer–mixer combination gives the lowest reported errors on the largest Rydberg arrays. At $24\times24$, it uses approximately $3.26\times10^6$ forward-pass FLOPs, versus $2.31\times10^6$ for the less accurate hierarchical mixer.
The study is restricted to positive real ground-state wave functions of stoquastic Hamiltonians; phase modeling is left for future work. Computational comparisons exclude sampling, observable evaluation, backpropagation, and memory costs. The hierarchical mixer's accuracy deteriorates at the largest size, and larger-system tests are acknowledged as necessary to establish scaling conclusions.
Little sign of AI-generated textIncrementalOverclaimsMissing comparisonPossible error
Central claim may not hold: Forward-pass FLOPs exclude architecture-dependent autoregressive sampling and training costs, so the reported benchmarks do not establish overall computational outperformance.
The 16×16 and 24×24 benchmarks compare only four hierarchical architectures, leaving the claimed large-system superiority over conventional single-site and patched architectures without direct accuracy–cost comparisons.
The claimed linear sequence-length scaling conflicts with the appendix's dense S×S token-mixing matrices, which cost O(CS²), while the separately listed hidden width suggests a potentially different implementation that is not consistently specified.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
Numerically simulating quantum many-body systems is one of the central challenges in modern physics, as the exponential growth of the Hilbert space with system size renders large-scale exact solutions intractable. Over the past decade, generative artificial neural networks have emerged as powerful wave function ansätze that have powered advances in finding ground state representations, reconstructing quantum states from data, and modeling dynamics in quantum many-body systems. While recurrent neural networks (RNNs) are computationally efficient, the recurrent nature can limit their capacity to capture long-range correlations. Transformer models with their underlying self-attention mechanism excel at capturing these long-range correlations, but incur increased computational requirements, restricting applications to small physical systems. In this work, we address these limitations by exploring the multi-layer perceptron (MLP) mixer model as a wave function ansatz. By mixing information globally and locally through the system via a computationally lightweight mixing strategy, this architecture promises to combine the ability to capture long-range correlations with low computational costs. We benchmark the model on finding ground state representations of two-dimensional quantum many-body systems in various phases of matter and investigate the behavior of representational accuracy and computational efficiency for different physical system sizes. By comparing its performance to RNNs and transformer models with single-site and patched input definitions, we demonstrate that the MLP mixer architecture performs competitively with state-of-the-art network models. We further show that a hierarchical combination of this model with the transformer model outperforms other state-of-the-art architectures for large-scale quantum systems, opening new avenues for investigating quantum phases of matter.
The paper studies whether autoregressive MLP-mixer neural wave functions can combine long-range correlation modeling with lower computational cost than transformers. It adapts the mixer using causal masks, shifted inputs, channel-only normalization, and probability outputs, and constructs hierarchical architectures in which a coarse network conditions a second network that generates smaller patches.
Nine architectures—including sitewise and patched RNNs and transformers, mixers, and hierarchical combinations—are compared through variational ground-state optimization. Benchmarks cover the two-dimensional transverse-field Ising model on 4×4 and 8×8 lattices and Rydberg arrays up to 24×24 sites, with open boundaries. Each system is tested at three parameter values spanning ordered, disordered, and near-transition regimes, with five independent runs. Energy errors are measured against tensor-network or quantum Monte Carlo references; network parameter counts range from approximately 5,000 to 30,000.
Patched RNNs provide the reported best accuracy–cost tradeoff on the smallest systems. Hierarchical architectures become preferable at larger sizes; the transformer–mixer combination gives the lowest reported errors on the largest Rydberg arrays. At 24×24, it uses approximately 3.26×106 forward-pass FLOPs, versus 2.31×106 for the less accurate hierarchical mixer.
The study is restricted to positive real ground-state wave functions of stoquastic Hamiltonians; phase modeling is left for future work. Computational comparisons exclude sampling, observable evaluation, backpropagation, and memory costs. The hierarchical mixer's accuracy deteriorates at the largest size, and larger-system tests are acknowledged as necessary to establish scaling conclusions.
Non-Hermitian (NH) quantum systems have emerged as a powerful framework for describing open quantum systems, nonequilibrium dynamics, and engineered quantum optical materials. However, solving the ground-state properties of NH systems is challenging due to the exponential scaling of the Hilbert space and exotic phenomena such as the emergence of exceptional points. Another challenge arises from the limitations of traditional methods like exact diagonalization (ED). For the past decade, neural networks (NNs) have shown promise in approximating many-body wavefunctions, yet their application to NH systems remains largely unexplored. In this paper, we explore different NN architectures to investigate the ground-state properties of a parity-time-symmetric, one-dimensional NH, transverse field Ising model with a complex spectrum by employing a recurrent neural network (RNN), a restricted Boltzmann machine (RBM), and a multilayer perceptron (MLP). We construct the NN-based many-body wavefunctions and validate our approach by recovering the ground-state properties of the model for small system sizes, finding excellent agreement with ED. Furthermore, for larger system sizes, we demonstrate that the RNN outperforms both the RBM and MLP. However, we show that the accuracy of the RBM and MLP can be significantly improved through transfer learning, allowing them to perform comparably to the RNN for larger system sizes. These results highlight the potential of neural-network-based approaches—particularly for accurately capturing the low-energy physics of NH quantum systems in case of both weak and strong non-Hermiticity.
The simulation of quantum matter with classical hardware plays a central role in the discovery and development of quantum many-body systems, with far-reaching implications in condensed matter physics and quantum technologies. In general, efficient and sophisticated algorithms are required to overcome the severe challenge posed by the exponential scaling of the Hilbert space of quantum systems. In contrast, hardware built with quantum bits of information are inherently capable of efficiently finding solutions of quantum many-body problems. While a universal and scalable quantum computer is still beyond the horizon, recent advances in qubit manufacturing and coherent control of synthetic quantum matter are leading to a new generation of intermediate scale quantum hardware. \n \nThe complexity underlying quantum many-body systems closely resembles the one encountered in many problems in the world of information and technology. In both contexts, the complexity stems from a large number of interacting degrees of freedom. A powerful strategy in the latter scenario is machine learning, a subfield of artificial intelligence where large amounts of data are used to extract relevant features and patterns. In particular, artificial neural networks have been demonstrated to be capable of discovering low-dimensional representations of complex objects from high-dimensional dataset, leading to the profound technological revolution we all witness in our daily life. \n \nIn this Thesis, we envision a new paradigm for scientific discovery in quantum physics. On the one hand, we have the essentially unlimited data generated with the increasing amount of highly controllable quantum hardware. On the other hand, we have a set of powerful algorithms that efficiently capture non-trivial correlations from high-dimensional data. Therefore, we fully embrace this data-driven approach to quantum mechanics, and anticipate new exciting possibilities in the field of quantum many-body physics and quantum information science. We revive a powerful stochastic neural network called a restricted Boltzmann machine, which slowly moved out of fashion after playing a central role in the machine learning revolution of the early 2010s. We introduce a neural-network representation of quantum states based on this generative model. We propose a set of algorithms to reconstruct unknown quantum states from measurement data and numerically demonstrate their potential, with important implications for current experiments. These include the reconstruction of experimentally inaccessible properties, such as entanglement, and diagnostics to determine sources of noise. Furthermore, we introduce a machine learning framework for quantum error correction, where a neural network learns the best decoding strategy directly from data. We expect that the full integration between quantum hardware and artificial intelligence will become the gold standard, and will drive the world into the era of fault-tolerant quantum computing and large-scale quantum simulations.
Due to their immense representative power, neural network quantum states (NQS) have gained significant interest in current research. In recent advances in the field of NQS, it has been demonstrated that this approach can compete with state-of-the-art numerical techniques, making NQS a compelling alternative, in particular for the simulation of large, two-dimensional (2D) quantum systems. In this study, we show that recurrent neural network (RNN) wave functions can be employed to study systems relevant to current research in quantum many-body physics. Specifically, we employ a 2D tensorized gated RNN to explore the bosonic Hofstadter model with a variable local Hilbert space cutoff and long-range interactions. At first, we benchmark the RNN-NQS for the Hofstadter-Bose-Hubbard (HBH) Hamiltonian on a square lattice. We find that this method, despite the complexity of the wave function, is capable of efficiently identifying and representing most ground-state properties. Afterwards, we apply the method to an even more challenging model for current methods, namely, the Hofstadter model with long-range interactions. This model describes Rydberg-dressed atoms on a lattice subject to a synthetic magnetic field. We study systems of size up to $12\ifmmode\times\else\texttimes\fi{}12$ sites and identify three different regimes by tuning the interaction range and the filling fraction $\ensuremath{\nu}$. In addition to phases known from the HBH model at short-range interaction, we observe bubble crystals and Wigner crystals for long-range interactions. Especially interesting is the evidence of a bubble crystal phase on a lattice, as this gives experiments a starting point for the search of clustered liquid phases, possibly hosting non-Abelian anyon excitations. In our work, we show that NQS are an efficient and reliable simulation method for quantum systems, which are the subject of current research. In particular, we demonstrate the ability of this method to simulate challenging systems with long-range interactions.
Neural-network quantum states have emerged as a powerful variational framework for quantum many-body systems, with recent progress often driven by massively parallel architectures such as transformers. Recurrent neural network quantum states, however, are frequently regarded as intrinsically sequential and therefore less scalable. Here we revisit this view by showing that modern recurrent architectures can support fast, accurate, and computationally accessible neural quantum state simulations. Using autoregressive recurrent wave functions together with recent advances in parallelizable recurrence, we develop variational ansätze, called parallel scan recurrent neural quantum states (PSR-NQS), which can be trained efficiently within variational Monte Carlo in one and two spatial dimensions. We demonstrate accurate benchmark results and show that, with iterative retraining, our approach reaches two-dimensional spin lattices as large as 52×52 while remaining in agreement with available quantum Monte Carlo data. Our results establish recurrent architectures as a practical and promising route toward scalable neural quantum state simulations with modest computational resources.
Neural network quantum states represent a powerful approach for solving electronic structures in strongly correlated molecular and material systems. For a neural network ansatz to be accurate, it must effectively learn the phase of a complex wave function. In this work, we demonstrate several different network structures as the phase network for a Transformer-based neural network quantum state implementation. We compare the accuracy of ground state energies, the number of parameters, and computational time across several small molecules. Furthermore, we propose a phase network setup that combines cross attention and multilayer perceptron structures, with the number of parameters remaining constant across different systems. Such an architecture may help reduce computational costs and enable transfer learning to larger quantum chemical systems.
ABSTRACT The field of quantum chemistry has long been defined by the central computational challenge of the quantum many‐body fermionic system problem—namely, the electronic Schrödinger equation. The exponential scaling of this equation makes it computationally intractable to solve exactly. As a result, various approximation methods have been developed, ranging from the widely used Density Functional Theory (DFT) to Coupled Cluster (CC) theory. While powerful, these conventional approaches often involve controlled truncations which cause a trade‐off between accuracy and computational cost, leaving many systems—particularly those with strong electron correlation—beyond the reach of ab initio simulation. Furthermore, the famous fermionic sign problem is particularly severe in stochastic methods like Quantum Monte Carlo (QMC). Traditional ansatz are also frequently engineered specifically for particular systems, which results in a fragmentation of methodologies and fundamentally limits their transferability. Recent research demonstrates that with carefully designed architectures, ANNs can not only represent ground states of model systems with high precision but also tackle the sign problem, offering a promising path forward for strongly correlated electronic structure calculations. This review charts the paradigm shift driven by Neural Network Quantum States (NNQS), which leverage the representational power of deep learning to overcome these barriers. We detail the architectural evolution from Restricted Boltzmann Machines to autoregressive models like RNNs and Transformers, which enable exact, uncorrelated sampling and bypass critical bottlenecks. The high‐efficiency optimization via neural networks was also explored, which decouples optimization cost from model complexity. Furthermore, the frontier of the field is marked by the integration of operator learning and hybrid quantum‐classical frameworks, such as Reinforcement Learning for contractive quantum eigen‐solvers to generate quantum circuits for simulating many‐body molecular systems, and Quantum‐Enhanced Neural Networks, which leverage quantum processors to enhance expressivity for hardware‐efficient ansatz. By framing the problem through the lens of computational resource allocation and quantum hardware integration—specifically, trading exponential memory demands for polynomial‐complexity optimization and sampling—this review elucidates how recent advances at the intersection of artificial neural networks and quantum computing have evolved. The emergence of these methods represents not merely progress toward powerful simulation tools, but also the creation of a hybrid computational interface. This convergence offers a versatile and systematically improvable framework for potentially addressing some of the most challenging problems in quantum physics and chemistry.