Machine‑learning molecular dynamics (MLMD)

We train machine-learned potentials on high-quality DFT data to achieve first-principles accuracy at molecular dynamics speeds. This enables broad exploration of composition and pH–temperature–pressure conditions for mineral–water systems. By developing and applying machine-learned force fields (MLFFs), we simulate complex aqueous–mineral environments with near-DFT precision over nanosecond timescales and capture realistic dynamics efficiently.

Typical questions

• How to model system with explicit water molecules and mineral surface for more than 100 ps simulation time?
• How do rare events (nucleation steps) emerge over longer trajectories?
• What uncertainty arises from the ML potential and how do we bound it?
• Can we rapidly screen ligands/conditions for optimal leaching or metal recovery?

Approach

• Automated Workflow: Uses DeePMD-kit to combine iterative training of machine-learning potentials (MLPs) with free-energy calculations, accelerating computation of redox potentials, acidity constants, and solvation free energies.
• On-the-Fly Training: Machine-learned force fields (MLFFs) are trained dynamically during AIMD runs in VASP, reducing reliance on large pre-computed datasets.
• Rigorous Validation: All models are benchmarked against DFT/AIMD observables and experimental data to ensure accuracy and reliability.

Applications

• Surface and Solvent Interactions: Predict ion adsorption/desorption behaviour and reveal water structure and dynamics at mineral–water interfaces.
• Ligand Selectivity: Assess competitive binding of ligands under varying pH, temperature, and pressure conditions.
• Mineral Dissolution: Model dissolution pathways and identify rate-controlling steps.
• Chemical Properties: Compute pKa values, redox potentials, and stability constants for complex aqueous–mineral systems.

Tools

DeePMD‑kit/DP‑GEN, LAMMPS, CP2K/VASP

Related

#AIMD #DFT