Sorption Molecular Dynamics of Point Defects in 2D Materials with a Machine Learning Potential

Abstract

Defect engineering is a crucial tool for tuning and tailoring the properties of two-dimensional (2D) materials. A key challenge is identifying which defects are kinetically stable at room temperature and sufficiently promising to warrant further investigation. However, identifying stable defects with suitable properties using first-principles calculations, such as density functional theory (DFT) is computationally demanding, effectively limiting the scope of high-throughput studies. In this work, we investigate whether the machine learning interatomic potential MACE can reproduce sorption trends (adsorption vs. absorption) previously obtained from high-throughput DFT calculations and extend the analysis to elevated temperatures. A total of 17,624 defect structures from the ADAQ database are simulated using MACE-accelerated molecular dynamics (MD) at both 0 K and 300 K. The results show that MACE successfully reproduces the overall sorption trends observed in DFT calculations at 0 K at a fraction of the computational cost. Simulations performed at 300 K further indicate that the qualitative sorption trends remain largely the same when including finite-temperature effects. These findings demonstrate that MACE-driven MD can be used as an effective high-throughput prescreening tool for defect-related studies in 2D materials, specially related to sorption. While conventional first-principles methods, such as DFT, remain necessary for accurate and quantitative validation, the use of MLIPs enables MD simulations of finite-temperature defect behavior at a fraction of the computational cost of DFT-based MD.