Optimizing machine learning interatomic potentials for hydroxide transport: Surprising efficiency of...

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning interatomic potentials (MLIPs) are increasingly used in molecular simulations.
• Transferability of MLIPs across different chemical environments or concentrations is a key challenge.
• Aqueous potassium hydroxide (KOH) solutions present a chemically homogeneous system with varying concentrations.

Purpose of the Study:

• To investigate the transferability of MLIPs for aqueous KOH solutions across concentration variations.
• To identify strategies for improving MLIP transferability without extensive data.
• To enable accurate simulations of hydroxide transport dynamics under diverse electrolyte conditions.

Main Methods:

• Developing and fine-tuning MLIPs on specific KOH concentrations.
• Evaluating MLIP performance (force prediction errors) across a wide concentration range (0.56-17.89 mol L-1).
• Comparing models trained on single vs. multiple concentrations, and strategically selected intermediate concentrations.

Main Results:

• Models trained on specific concentrations showed poor transferability, with errors increasing from 30 to 90 meV Å-1.
• Fine-tuning on an intermediate concentration (6.26 mol L-1) yielded excellent transferability across all tested concentrations.
• The intermediate-concentration model accurately captured emergent high-concentration phenomena like hydroxide-hydroxide hydrogen bonding.

Conclusions:

• Strategic data selection, specifically intermediate concentrations, significantly enhances MLIP transferability in chemically similar systems.
• This approach offers a computationally efficient alternative to training on diverse datasets for robust MLIP performance.
• The findings provide practical guidelines for developing broadly applicable MLIPs for electrolyte simulations.