Exploring Thermodynamic Behavior of Spin Glasses with Machine Learning

Abstract

In this paper, we consider the regression problem of predicting thermodynamic quantities - specifically the average energy $\langle E \rangle$ - as a function of temperature $T$ for spin glasses on a square lattice. The spin glass is represented as a weighted graph, where exchange interactions define the edge weights. We investigate how the spatial distribution of these interactions relates to $\langle E \rangle$, leveraging several machine learning approaches that we specifically developed for this task. While $\langle E \rangle$ is used to demonstrate the approach, our framework is general and can be applicable to the prediction of other thermodynamic characteristics.