Accelerating explicit solvent models of heterogeneous catalysts with machine learning interatomic potentials.

Area of Science:

• Computational chemistry
• Materials science
• Chemical kinetics

Background:

• Modeling solvent effects on catalytic reactions is computationally expensive.
• Explicit solvent treatment requires molecular dynamics (MD) and enhanced sampling methods.

Purpose of the Study:

• To demonstrate machine learning interatomic potentials (MLIPs) for fast and accurate explicit solvent modeling of heterogeneous catalysts.
• To enable large-scale, realistic simulations of solvated catalysts.

Main Methods:

• Developed and utilized on-the-fly trained MLIPs coupled with active learning.
• Accelerated ab initio MD simulations by up to 4 orders of magnitude.
• Validated MLIPs against ab initio calculations for accuracy.

Main Results:

• MLIPs accurately reproduced water's structure in bulk and at metal-water interfaces.
• Predicted adsorption energies for key species (CO*, OH*, etc.) on Cu surfaces.
• Calculated free energy barriers for C-H scission of ethylene glycol on Cu and Pd surfaces.

Conclusions:

• MLIPs offer a computationally efficient approach for explicit solvent modeling in catalysis.
• This method enables accurate prediction of catalytic performance in realistic solvated environments.
• Paves the way for detailed studies of solvated catalysts at unprecedented scales.