Automated discovery of characteristic features of phase transitions in many-body localization

TL;DR

This paper introduces an AI-driven, nearly unsupervised method to identify a new order parameter for the many-body localization transition, significantly reducing computational effort and enhancing understanding of phase transitions.

Contribution

The study presents a novel AI-based approach using adversarial neural networks to discover characteristic features of phase transitions, improving efficiency and objectivity in identifying the transition point.

Findings

New order parameter for MBL transition identified

Reduction of numerical effort by approximately 100 times

Method applicable to poorly understood phase transitions

Abstract

We identify a new "order parameter" for the disorder driven many-body localization (MBL) transition by leveraging artificial intelligence. This allows us to pin down the transition, as the point at which the physics changes qualitatively, from vastly fewer disorder realizations and in an objective and cleaner way than is possible with the existing zoo of quantities. Contrary to previous studies, our method is almost entirely unsupervised. A game theoretic process between neural networks defines an adversarial setup with conflicting objectives to identify what characteristic features to base efficient predictions on. This reduces the numerical effort for mapping out the phase diagram by a factor of ~100x. This approach of automated discovery is applicable specifically to poorly understood phase transitions and exemplifies the potential of machine learning assisted research in physics.