Self-Consistent Biased Fine-Tuning for Highly Accurate Reaction-Specific Machine-Learning Interatomic Potentials

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Accurate interatomic potentials are crucial for simulating chemical reactions.
• Developing system-specific potentials often requires extensive, high-level quantum mechanical calculations.
• Machine-learning interatomic potentials (MLIPs) offer a promising alternative for efficient simulations.

Purpose of the Study:

• To present a black-box algorithm for generating high-quality, system-specific MLIPs for gas-phase reactions.
• To enable direct comparison of reaction rate constants between MLIPs and reference methods.
• To optimize the trade-off between computational speed and accuracy in MLIP development.

Main Methods:

• Utilized a self-consistent fine-tuning approach for a Message-Passing Atomic Cluster Expansion (MACE) foundation model.
• Employed biased sampling along reaction coordinates using the Caracal program package.
• Iteratively refined the MLIP by resampling with a fine-tuned model until energy and force errors met predefined thresholds.
• Benchmarked the method by varying training data size and MACE architecture.
• Incorporated nuclear quantum effects via ring-polymer molecular dynamics (RPMD).

Main Results:

• Demonstrated the ability to generate high-quality reactive MLIPs with limited, expensive quantum mechanical reference calculations.
• Showcased the importance of including recrossing trajectories for accurate parametrization of reversible reaction mechanisms.
• Successfully benchmarked the algorithm on two gas-phase reactions: malonaldehyde internal proton transfer and methane-OH radical proton exchange.
• Achieved an optimal balance between speed and accuracy by systematically varying training data and MLIP architecture.

Conclusions:

• The presented algorithm enables automated, high-quality MLIP generation for reactive systems.
• This approach significantly reduces the computational cost associated with developing accurate interatomic potentials.
• The findings pave the way for more efficient and accurate simulations of chemical reactions using MLIPs.