Cost-Effective Strategy of Enhancing Machine Learning Potentials by Transfer Learning from a Multicomponent Data Set...

Area of Science:

• Computational Materials Science
• Catalysis
• Machine Learning

Background:

• Machine learning potentials (MLPs) enable efficient material simulations but require extensive ab initio data.
• Constructing large reference databases for catalyst-adsorbate systems is computationally expensive and challenging.
• Training MLPs with limited data can lead to overfitting and reduced practical applicability.

Purpose of the Study:

• To explore a cost-effective transfer learning strategy for developing accurate MLPs for catalyst-adsorbate systems.
• To investigate the use of limited ab initio references by leveraging pre-trained models from public databases.
• To assess the generalizability and stability of MLPs developed through transfer learning.

Main Methods:

• Utilized the Open Catalyst Project 2020 (OC20) database for pretraining MLP models using the ænet-PyTorch framework.
• Compared different strategies for selecting subsets of the OC20 database for transfer learning.
• Performed molecular dynamics simulations to evaluate the stability and accuracy of the developed MLPs.

Main Results:

• MLPs developed via transfer learning demonstrated superior generalizability and stability compared to those trained from scratch.
• Transfer learning significantly enhanced the accuracy and stability of MLPs for the CuAu/H2O system with ~600 data points.
• The transfer learning approach achieved stable and accurate predictions for up to 250 ps in molecular dynamics simulations of CuAu/6H2O, outperforming models without transfer learning.

Conclusions:

• Transfer learning offers a computationally cost-effective method for constructing accurate MLPs for catalyst-adsorbate systems with limited data.
• This strategy improves the stability and extrapolation capabilities of MLPs in molecular dynamics simulations.
• The proposed methodology facilitates broader applications in materials science and catalysis, enabling more efficient simulations.