SchNet_IIA: Potential\nEnergy Surface Fitting by Interatomic\nInteractions Attention Based on Transfer Learning Analysis

Machine learning methods for fitting\npotential energy\nsurfaces\nand molecular dynamics simulations are becoming increasingly popular\ndue to their potentially high accuracy and savings in computational\nresources. However, existing application models often rely on basic\narchitectures like artificial neural networks (ANNs) and multilayer\nperceptron (MLP), lagging behind cutting-edge technologies in the\nmachine learning domain. Furthermore, the complexity of current machine\nlearning frameworks leads to reduced interpretability and challenges\nfor improvement. Herein, we developed a model analysis method based\non the feature-representation-transfer approach to directly perform\ncausal analysis on the model. The internal action characteristics\nof the SchNet framework were successfully analyzed by constructing\ndifferent source tasks and we proposed interatomic interactions attention\nfor the characterization of doped clusters. The accuracy was enhanced\nby 0.015 eV/atom compared to the original model. The ability to capture\natomic environment characteristics was significantly improved. The\nactivation function was smoothed resulting in a 23.47% increase in\nthe convergence speed. Our SchNet_IIA model demonstrates superior\nperformance in capturing interatomic interactions. Our present work\nis of distinctive value as it presents a novel transfer learning analysis\nmethod with the potential to evolve into a generalized model analysis\napproach, providing new perspectives and solutions for the field.