Improving Aqueous Metal Salt Interactions Using Machine-Learned Interatomic Potentials.

Area of Science:

• Computational chemistry
• Environmental science
• Materials science

Background:

• Accurate simulation of aqueous metal salt solutions is crucial for environmental safety and energy applications.
• Trace metals like arsenic pose significant risks but are difficult to model accurately with classical force fields or ab initio methods.

Purpose of the Study:

• To develop and validate machine-learned interatomic potentials (MLIPs) for modeling aqueous arsenic (AsCl3) and magnesium (MgCl2) chloride solutions.
• To assess the performance of MLIPs against ab initio molecular dynamics (AIMD) and classical force fields (CFFs).

Main Methods:

• Utilized the NequIP/Allegro equivariant graph neural network architecture.
• Trained MLIPs on data from AIMD and density functional theory calculations.
• Compared MLIPs with AMBER and UFF classical force fields.

Main Results:

• MLIPs accurately reproduced ab initio energies and forces with low mean absolute error (< 1 meV/atom) and root-mean-square error (< 40 meV/Å).
• MLIPs effectively captured solvation structure, ion diffusion, and hydration dynamics.
• Achieved a speedup of approximately 10,000 times compared to AIMD simulations.

Conclusions:

• Developed MLIPs provide a reliable and efficient method for simulating trace metal solutions.
• These MLIPs can enhance modeling of trace metal speciation and transport for improved environmental and separation processes.