Data-driven assessment of thermodynamic stability and search for competing phases: Application to the 2D Material Defect dataset

Abstract

In theoretical materials science, combining traditionalmodeling methods like density functionaltheory (DFT) with data-driven approaches has becomeone of the main trends of the last decade.Machine learning models are now widely used forstructure-to-property predictions, modeling interatomicinteractions, and generating new moleculesand crystal structures. Among these, graph neuralnetworks (GNNs) stand out as a particularly popularchoice in many such applications [1]. The possibilityof DFT/GNN combinations is caused by data accumulationled to creation a number of general purpose(the Materials Project, AflowLib, etc.) databases.For specific problems, e.g., modeling catalytic processes[2], doping effects on phase stability [3], etc.,there are often no ready-made data collections, andit is necessary to create new ones to build customdata-driven solutions. Therefore, the developmentof DFT/GNN approaches and their implementationin a data-efficient manner are of particular interest.