Fast and accurate prediction of adsorption energy of AgPd nanoalloys by deep learning potentials and neural networks

The accurate prediction of adsorption energy plays a crucial role in the design of efficient catalysts. However, the high-throughput prediction of adsorption energies based on first-principles methods presents significant challenges, hindering the rapid acceleration of catalyst development. In this study, a dataset of F atom adsorption energies was constructed, consisting of 1087 adsorption configurations on AgPd clusters, using the Ag-Pd-F deep learning potential. We employed four distinct deep neural network architectures, namely, Crystal Graph Convolutional Neural Networks (CGCNN), Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and Multilayer Perceptrons (MLP), to establish an accurate predictive model for F atom adsorption on 38-atom AgPd nanoclusters. Among the models tested, the CGCNN, based on deep learning potentials, demonstrated the highest accuracy and provided the best uncertainty estimates, achieving an R² of 0.8379 and an RMSE of 0.1036 eV on the test set. This approach not only outperformed the other models in terms of predictive precision but also offered superior computational speed and cost-efficiency. The findings highlight the potential of deep learning models, particularly deep learning potentials, in predicting adsorption energies of nanoclusters, paving the way for accelerated material design and performance prediction in catalysis and materials science.