A cost-effective strategy of enhancing machine learning potentials by transfer learning from a multicomponent dataset on ænet-PyTorch

Machine learning potentials (MLPs) offer efficient and accurate material simulations, but constructing the reference ab initio database remains a significant challenge, particularly for catalyst-adsorbate systems. Training an MLP with a small dataset can lead to overfitting, thus limiting its practical applications. This study explores the feasibility of developing computationally cost-effective and accurate MLPs for catalyst-adsorbate systems with a limited number of ab initio references by leveraging a transfer learning strategy from subsets of a comprehensive public database. Using the Open Catalyst Project 2020 (OC20) -- a dataset closely related to our system of interest -- we pre-trained MLP models on OC20 subsets using the {\ae}net-PyTorch framework. We compared several strategies for database subset selection. Our findings indicate that MLPs constructed via transfer learning exhibit better generalizability than those constructed from scratch, as demonstrated by the consistency in the dynamics simulations. Remarkably, transfer learning enhances the stability and accuracy of MLPs for the CuAu/H2O system with approximately 600 reference data points. This approach achieved excellent extrapolation performance in molecular dynamics (MD) simulations for the larger CuAu/6H2O system, sustaining up to 250 ps, whereas MLPs without transfer learning lasted less than 50 ps. We also examine the potential limitations of this strategy. This work proposes an alternative, cost-effective approach for constructing MLPs for the challenging simulation of catalytic systems. Finally, we anticipate that this methodology will pave the way for broader applications in material science and catalysis research, facilitating more efficient and accurate simulations across various systems.