Putting machine learning to the test in a quantum many-body system

Quantum many-body systems pose a formidable computational challenge due to the exponential growth of their Hilbert space. While machine learning (ML) has shown promise as an alternative paradigm, most applications remain at the proof-of-concept stage, focusing narrowly on energy estimation at the lower end of the spectrum. Here, we push ML beyond this frontier by extensively testing HubbardNet, a deep neural network architecture for the Bose-Hubbard model. Pushing improvements in the optimizer and learning rates, and introducing physics-informed output activations that can resolve extremely small wave-function amplitudes, we achieve ground-state energy errors reduced by orders of magnitude and wave-function fidelities exceeding 99%. We further assess physical relevance by analysing generalized inverse participation ratios and multifractal dimensions for ground and excited states in one and two dimensions, demonstrating that optimized ML models reproduce localization, delocalization, and multifractality trends across the spectrum. Crucially, these qualitative predictions remain robust across four decades of the interaction strength, e.g. spanning across superfluid, Mott-insulating, as well as quantum chaotic regimes. Together, these results suggest ML as a viable qualitative predictor of many-body structure, complementing the quantitative strengths of exact diagonalization and tensor-network methods.

💡 Research Summary

This paper presents a thorough benchmark and substantial improvement of the deep‑learning architecture HubbardNet for the Bose‑Hubbard model, demonstrating that machine learning can move far beyond proof‑of‑concept energy estimates and become a reliable predictor of many‑body wave‑function structure across a wide range of interaction strengths. The authors first recall the exponential growth of the Hilbert space in quantum many‑body problems and the limitations of exact diagonalization, DMRG, and PEPS, especially for rapid parameter sweeps or higher‑dimensional systems. They then introduce the generalized fractal dimensions D_q (including D₁, D₂, and D_∞) and the inverse participation ratio as quantitative measures of localization, delocalization, and multifractality of eigenstates in the Fock basis.

HubbardNet, a fully‑connected multilayer perceptron with four hidden layers of 400 tanh neurons, takes as input the site occupations, the interaction strength U, and the particle number N, and outputs a single scalar that is transformed into a normalized Fock‑basis coefficient via an activation function. The original implementation used a simple cosine‑annealed learning‑rate schedule (max η = 0.01) and stochastic gradient descent. The present work replaces SGD with the Adam optimizer, reduces the maximum learning rate to 5 × 10⁻⁶, and retains the cosine‑annealing reset every 1 000 epochs. Crucially, the output activation is chosen to respect physical constraints: an exponential activation for the ground state (ensuring positivity) and a linear activation for excited states, combined with logarithmic scaling to capture amplitudes that span many orders of magnitude.

Training is performed on nine logarithmically spaced values of U ranging from 10⁻² to 10², thereby covering four decades and the full superfluid–Mott–chaotic crossover. After convergence (defined by a loss‑fluctuation below 10⁻⁷ over the last 50 steps), the network is tested on 69 out‑of‑sample U values. The resulting ground‑state energies have relative errors of order 10⁻⁴, while the wave‑function fidelities exceed 99 % across the entire range. The authors highlight three representative interaction strengths—U≈0.21 (deep superfluid), U≈2.05 (near the 1D critical point U_c≈3), and U≈20.5 (deep Mott)—and show that the predicted D₁, D₂, and D_∞ match exact diagonalization results, correctly reproducing the localization‑delocalization transition and the emergence of multifractal behavior.

For excited states, the original HubbardNet relied on a Gram‑Schmidt orthogonalization tower, which becomes computationally prohibitive for large Hilbert spaces. The authors introduce an observable‑based loss that directly minimizes a chosen fractal dimension (e.g., D₂) or the inverse participation ratio, thereby bypassing explicit orthogonalization. This approach yields excited‑state wave‑functions whose multifractal spectra agree with exact results both in the bulk of the spectrum (scaled energy ε≈0) and at high energies (ε≈1). The method works equally well in two dimensions (a 4 × 4 lattice with three bosons), confirming that the network can capture the same structural trends despite the lower particle density and different degeneracy patterns.

Overall, the paper demonstrates three key advances: (1) optimizer and learning‑rate refinements that improve convergence speed and accuracy by two orders of magnitude; (2) physics‑informed output activations that enable the network to resolve amplitudes spanning many decades, leading to wave‑function fidelities >99 %; and (3) a novel observable‑driven training scheme for excited states that eliminates the need for costly Gram‑Schmidt procedures while preserving multifractal characteristics. The authors discuss remaining challenges—scaling to larger lattices, handling strong disorder, and integrating with tensor‑network or quantum‑circuit variational methods—and outline future directions such as graph‑neural‑network extensions and hybrid ML‑Monte‑Carlo approaches. In sum, this work establishes machine learning as a viable, qualitatively accurate tool for probing the intricate structure of many‑body quantum states across diverse physical regimes, complementing traditional exact and tensor‑network techniques.