From Pretrained to Precision: Fine-Tuning Universal Interatomic Potentials for Accurate Catalytic Reaction...

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Universal machine learning interatomic potentials (uMLIPs) offer high accuracy for various chemical systems.
• Current uMLIPs struggle with accurate catalytic reaction and barrier prediction.
• Improving uMLIPs for reaction modeling is crucial for catalysis research.

Purpose of the Study:

• To enhance the performance of established uMLIPs for catalytic reaction prediction using fine-tuning.
• To systematically compare fine-tuning against training from scratch for data efficiency and accuracy.
• To evaluate the impact of fine-tuning on model generalization across different tasks.

Main Methods:

• Evaluated two established uMLIPs.
• Applied fine-tuning strategies to improve reaction prediction accuracy.
• Compared fine-tuning with training from scratch across various tasks like MD simulations, adsorption energy, and transition state searches.
• Analyzed performance with varying training set sizes and assessed extrapolative generalization.

Main Results:

• Fine-tuning reduced the mean absolute error (MAE) for reaction energy predictions from 0.38 eV to 0.09 eV.
• Fine-tuned models required only 10%-30% of the data needed for training from scratch.
• Generalization capabilities of uMLIPs were maintained after fine-tuning.
• Improved accuracy was observed for both simple and complex tasks, including unseen elements.

Conclusions:

• Fine-tuning is an effective strategy to significantly boost uMLIP accuracy for catalytic reaction prediction.
• This approach enhances data efficiency and preserves model generalization.
• Fine-tuned uMLIPs show great potential for broader applications in diverse catalytic reaction systems.