Combined DFT/GNN analysis of compositional disorder and defects in MeX 2 (Me = Mo, W; X = S, Se) monolayers

Abstract

Abstract Chemical disorder and point defects critically influence the properties of two-dimensional (2D) materials, but their theoretical description is limited by the vast configurational space of possible atomic arrangements, which is intractable for conventional first-principles methods such as density functional theory (DFT). Here, we develop and validate a method for accelerated assessment of thermodynamic stability of alloyed transition metal dichalcogenide monolayers with and without vacancies. A combination of DFT and graph neural networks (GNNs) is applied across the composition/configuration space (CCS) representing structured collections of symmetry-unique atomic arrangements. First, using the method proposed and convex-hull-based analysis, we show that quaternary alloys are thermodynamically unstable at 0 K. This problem is further addressed by finite-temperature analysis. Second, we demonstrate that random sampling underestimates the configurational entropy contribution for highly defective quaternaries, while considering systematically sampled CCS delivers accurate assessments. Furthermore, structured training data are shown to improve GNN generalization and out-of-domain accuracy relative to random sampling. Third, reconstructing finite-temperature free-energy convex hulls for low-defect alloys, many quaternary compositions become thermodynamically stable. Chalcogen vacancies emerge as the dominant favored intrinsic defect in the studied systems, remaining only ≈20-30 meV/atom above the free-energy hull at temperatures up to 1200 K, whereas metal vacancies stay less favorable (≈70-90 meV/atom). Altogether these results establish a scalable route for finite-temperature thermodynamic screening of chemically disordered 2D materials with defects.