Convolutional Neural Networks Application for an Average Thermodynamic Characteristics Calculation of Spin Glasses

In this work, we introduce a novel methodology for studying the low-temperature phase of frustrated spin glass models using convolutional neural networks (CNNs). Our approach addresses the regression of thermodynamic properties, specifically the average energy ⟨E⟩, as a function of temperature T for spin glasses on a square lattice. By modelling the spin glass as a weighted graph, where exchange interaction values Jk are represented by the edges and mapped to lattice coordinates, we explore the functional relationship between ⟨E⟩ and the spatial distribution of J. We evaluate CNNs for their performance across various spin glass sizes and distributions of exchange integrals, demonstrating the potential of CNNs in capturing complex spin interactions and advancing the understanding of frustrated systems.