|
Dynamical large deviations of two-dimensional kinetically constrained models using a neural-network state ansatz
C. Casert, T. Vieijra, S. Whitelam, I. Tamblyn Physical Review Letters, 127, 120602 (2021) Understanding rare dynamical events in glassy systems is a fundamental problem in statistical mechanics, but computing large-deviation statistics for such systems in two dimensions has remained computationally intractable. In this work, we employ recurrent neural networks as a variational ansatz to study the dynamical large deviations of the activity in two-dimensional kinetically constrained models, which serve as prototypical models of glasses. By adapting neural-network techniques originally developed for quantum many-body physics to classical stochastic systems, we perform the first finite-size scaling analysis of the large-deviation functions for the two-dimensional Fredrickson-Andersen model and characterize the spatial structure of high-activity configurations in the South-or-East model. This work establishes a new route to the study of dynamical large-deviation functions and demonstrates the broad applicability of neural-network ansatze across different domains of physics. |
