X2-GNN: A Physical\nMessage Passing Neural Network\nwith Natural Generalization Ability to Large and Complex Molecules

Neural network models excel in molecular property predictions\nbut\noften struggle with generalizing from smaller to larger molecules\ndue to increased structural diversity and complex interactions. To\naddress this, we introduce an E(3) invariant (and equivariant capable)\nmessage passing graph neural network (GNN), namely, X2-GNN, that integrates\nphysical insights via atomic orbital overlap integrals and core Hamiltonians.\nThese features provide essential information about bond strength,\nelectron delocalization, and many-body interactions, enhanced by an\nattention mechanism for improved learning efficiency. Benchmarked\nagainst mainstream GNNs on diverse data sets, X2-GNN trained solely\non the QM9 data set (up to nine heavy atoms) effectively generalizes\nto larger molecules with tens of heavy atoms, achieving credible per-atom\nerror rates. It also excels in potential energy surface modeling and\naccurately predicts the bond dissociation energy within subseconds.\nThese results highlight X2-GNN’s scalability and broad applicability,\nemphasizing the importance of integrating data-driven approaches with\nbasic knowledge from electronic structure theory.