Cost-Effective\nStrategy of Enhancing Machine Learning\nPotentials by Transfer Learning from a Multicomponent Data Set on\nænet-PyTorch

Machine learning potentials (MLPs) offer efficient and\naccurate\nmaterial simulations, but constructing the reference ab initio database\nremains a significant challenge, particularly for catalyst-adsorbate\nsystems. Training an MLP with a small data set can lead to overfitting,\nthus limiting its practical applications. This study explores the\nfeasibility of developing computationally cost-effective and accurate\nMLPs for catalyst-adsorbate systems with a limited number of ab initio\nreferences by leveraging a transfer learning strategy from subsets\nof a comprehensive public database. Using the Open Catalyst Project\n2020 (OC20)a data set closely related to our system of interestwe\npretrained MLP models on OC20 subsets using the ænet-PyTorch\nframework. We compared several strategies for database subset selection.\nOur findings indicate that MLPs constructed via transfer learning\nexhibit better generalizability than those constructed from scratch,\nas demonstrated by the consistency in the dynamics simulations. Remarkably,\ntransfer learning enhances the stability and accuracy of MLPs for\nthe CuAu/H<sub>2</sub>O system with approximately 600 reference data\npoints. This approach achieved excellent extrapolation performance\nin molecular dynamics simulations for the larger CuAu/6H<sub>2</sub>O system, maintaining stable and accurate predictions for up to 250\nps, whereas MLPs without transfer learning become unstable before\nreaching 50 ps. We also examine the potential limitations of this\nstrategy. This work proposes an alternative, cost-effective approach\nfor constructing MLPs for the challenging simulation of catalytic\nsystems. Finally, we anticipate that this methodology will pave the\nway for broader applications in materials science and catalysis research,\nfacilitating more efficient and accurate simulations across various\nsystems.