Neural network setups for a precise detection of the many-body localization transition: finite-size scaling and limitations

TL;DR

This paper investigates neural network methods for accurately detecting the many-body localization transition in quantum systems, highlighting challenges and limitations in finite-size scaling and interpretability.

Contribution

It introduces neural network architectures and dataset setups that minimize physical bias for analyzing phase transitions in quantum many-body systems.

Findings

Neural networks show variability in estimating critical points.

Uncertainties in critical parameters tend to be larger than conventional methods.

Minimal physical input leads to challenges in quantitative phase transition analysis.

Abstract

Determining phase diagrams and phase transitions semi-automatically using machine learning has received a lot of attention recently, with results in good agreement with more conventional approaches in most cases. When it comes to more quantitative predictions, such as the identification of universality class or precise determination of critical points, the task is more challenging. As an exacting test-bed, we study the Heisenberg spin-1/2 chain in a random external field that is known to display a transition from a many-body localized to a thermalizing regime, which nature is not entirely characterized. We introduce different neural network structures and dataset setups to achieve a finite-size scaling analysis with the least possible physical bias (no assumed knowledge on the phase transition and directly inputing wave-function coefficients), using state-of-the-art input data simulating chains of sizes up to L=24. In particular, we use domain adversarial techniques to ensure that the network learns scale-invariant features. We find a variability of the output results with respect to network and training parameters, resulting in relatively large uncertainties on final estimates of critical point and correlation length exponent which tend to be larger than the values obtained from conventional approaches. We put the emphasis on interpretability throughout the paper and discuss what the network appears to learn for the various used architectures. Our findings show that a it quantitative analysis of phase transitions of unknown nature remains a difficult task with neural networks when using the minimally engineered physical input.