Machine learning potential for serpentines

TL;DR

This paper introduces a machine learning potential trained on DFT data to simulate serpentines, enabling detailed study of their high-temperature behavior and polymorphism under Earth's subduction conditions.

Contribution

The authors developed a novel ML potential for serpentine minerals based on DFT calculations, facilitating accurate molecular dynamics simulations of complex natural samples.

Findings

ML potential accurately reproduces high-temperature equations of state

Antigorite with larger periodicity is more stable at low temperatures

Potential enables investigation of complex antigorite superstructures

Abstract

Serpentines are layered hydrous magnesium silicates (MgO$\cdot$SiO$_2\cdot$H$_2$O) formed through serpentinization, a geochemical process that significantly alters the physical property of the mantle. They are hard to investigate experimentally and computationally due to the complexity of natural serpentine samples and the large number of atoms in the unit cell. We developed a machine learning (ML) potential for serpentine minerals based on density functional theory (DFT) calculation with the r$^2$SCAN meta-GGA functional for molecular dynamics simulation. We illustrate the success of this ML potential model in reproducing the high-temperature equation of states of several hydrous phases under the Earth's subduction zone conditions, including brucite, lizardite, and antigorite. In addition, we investigate the polymorphism of antigorite with periodicity $m$ = 13--24, which is believed to be all the naturally existent antigorite species. We found that antigorite with $m$ larger than 21 appears more stable than lizardite at low temperatures. This machine learning potential can be further applied to investigate more complex antigorite superstructures with multiple coexisting periodic waves.