Modelling silica using MACE-MP machine learnt interatomic potentials

Abstract

Silica polymorphs and zeolites are fundamental to a wide range of mineralogical and industrial applications owing to their diverse structural characteristics and thermodynamic and mechanical stability under varying conditions. Computational modelling has played a crucial role in understanding the relationship between the structure and functionality of silicas and silicates, including zeolites. In this study, we apply the MACE machine learnt interatomic potentials (MACE MP) to model the framework energies of siliceous zeolites and examine the phase transitions of silica and silicalite polymorphs under high-pressure conditions. MACE MP offers versatility by handling silicas with different coordination numbers, unlike earlier and successful IPs such as Sanders potentials (M. Sanders et al., J. Chem. Soc., Chem. Commun., 1984, 19, 1271-1273), which are restricted to four-coordinated Si environments and demand extensive re-parameterisation for higher coordination systems. The results reproduce the known metastability of siliceous zeolites relative to α-quartz, with energy differences between microporous and dense phases calculated by MACE-MP-0 medium and density functional theory (DFT) methods closely aligning with experimental calorimetric data. The high-pressure simulations reveal distinct compression behaviour in the quartz, coesite, and stishovite polymorphs of silica, with coesite and stishovite showing increased stability at elevated pressures in line with experimental data. The calculated phase transition pressures from quartz to coesite (∼3.5 GPa) and coesite to stishovite (∼9 GPa) are close to experimental findings, demonstrating the reliability of MACE-mp0 in modelling the structural and energetic properties of silica polymorphs. Furthermore, we examine the behaviour of fluoride ions in zeolite cages using MACE-MP, capturing known structural motifs such as pentacoordinated [SiO4F]− units and central cage-bound F− species, in agreement with prior DFT and experimental observations. Thus, we assess and demonstrate the suitability of off-the-shelf machine-learned foundation models, based on MACE-MP framework, for modelling silica materials of high importance from earth sciences to electronics and catalysis.