Achieving both precise extrapolation and physical interpretability in machine learning techniques critically depend on the interatomic potentials (ML-IPs), but the formalism is still underdevelopment, particularly in data-scarce chemical reactions involving multicomponent systems or at extreme conditions. Here, we present a pairwise-decomposed physics-informed neural network (P2Net) that parameterizes an analytical bond-order potential (BOP) layer to decouple the energy contributions of atomic pairs. By leveraging fundamental physical principles, P2Net demonstrates excellency at extrapolating performance beyond its training regime and accurately capturing molecular geometries far from equilibrium. The pairwise energy decomposition further empowers the bond analyses for deprotonation and reactions, which is not easy with most ML-IPs. The atomic-pair energy offers how to elucidate the evolution of interatomic interactions as reactions proceed. Our methodology highlights remarkable efficiency for building ML-IPs and post-simulation analysis, thereby universal and consistent framework for ML-IPs to use complex and reactive systems.
