Deep Learning Potential for Hydrous Stishovite and Post-Stishovite (Si-O-H System) based on r²SCAN meta-GGA

This repository contains a machine-learning interatomic potential (Deep Potential) for the hydrous silica (SiO₂-H₂O) system, specifically covering stishovite and CaCl₂-type post-stishovite phases. This potential excludes Aluminum (Al) and focuses on the Si-O-H ternary system. Methodology: Framework: DeepMD-kit v2 (Deep Potential Smooth Edition se_e2_a). Ab Initio Level: Density Functional Theory (DFT) using the r²SCAN meta-GGA exchange-correlation functional. Software: Quantum Espresso (DFT) and DeepMD-kit (Training). Training Domain: The potential was trained to cover pressure-temperature conditions up to 100 GPa and 3,000 K. Accuracy: The potential achieves a Root Mean Square Error (RMSE) of ~3.0 meV/atom in energy and ~0.14 eV/Å in force relative to the DFT reference. Atom Type Mapping: The potential uses the following type mapping Type 1: Silicon (Si) Type 2: Oxygen (O) Type 3: Hydrogen (H) Technical Parameters: Cut-off radius: 6.0 Å Embedding Network: (25, 50, 100) Fitting Network: (240, 240, 240) Descriptor: se_e2_a (two-body embedding with full radial and angular information) Applications: This potential is suitable for molecular dynamics (MD) simulations to investigate the thermodynamic properties, equation of state, and thermal conductivity of hydrous silica phases in the Earth's lower mantle. It supports hydrogen incorporation mechanisms, including hydrogarnet-type substitution (4H⁺ = Si⁴⁺) and interstitial H₂O channels.