Enhanced Encoding Module of AisNet with Charge Features: An Application to the S_N2 Data Set

This research introduces a charge-enhanced encoding module within the AisNet multifeature fusion framework, demonstrating superior force error accuracy and cross-model transferability, and achieving in state-of-the-art performance on the S_N2 dataset. Compared with the original AisNet, the addition of the Coulomb matrix enables AisNet-C with ternary features to maintain its leading performance in all feature dimensions. Through t-SNE and PCA analysis, long-range Coulomb interaction effectively enhances global environmental resolution, forming a synergistic relationship with other features. The performance of the AisNet-C is optimal at a dimension larger than 60, which is related to the accuracy of the ACSF fitting. Particularly, the PAINN model demonstrated the best compatibility with the enhanced module, resulting in the SOTA result of 0.00423 eV/Å and reducing the mean error over by 41.8% compared to PhysNet's benchmark. The key to PaiNN's success lies in the multimodule cooperation, which maintains intraclass atomic resolution while bringing different atom types closer together, thus forming a more compact feature space. Our plug-and-play encoding module can not only efficiently utilize multiple descriptors through feature decoupling, but also has the potential to be applied to a wider range of model architectures, providing new ideas for realizing physically informed machine learning force fields.