Neural Network-Based Interatomic Potential for the Study of Thermal and Mechanical Properties of Siliceous Zeolites

Abstract

Molecular dynamics simulations of zeolites are commonly employed for the characterization of their framework dynamics and response to the application of temperature and pressure. While classical interatomic potentials are commonly used for this task, they offer a description of the interactions in the system with limited accuracy. Density functional theory, meanwhile, is accurate, but its high computational expense limits its scalability for large systems or long dynamics. Recent advances in machine learning interatomic potentials, trained on computational data obtained at the quantum chemical level, offer a promising alternative combining high accuracy with computational efficiency. In this study, we developed an MLIP specifically for pure silica zeolites, trained on data from high-temperature ab initio MD simulations across various zeolitic topologies. This MLIP was then applied to predict structural properties, thermal expansion, and pressure response of different zeolites, demonstrating its potential for accuracy and generalizability in simulations of topologies beyond its initial training set.