Two-atom interaction potential database supporting physics-informed features for machine learning and material property prediction.

Accurately predicting the energy and relevant properties of molecules and complex crystalline systems remains a cornerstone of computational materials science. While density functional theory (DFT) offers high accuracy, its computational cost scales prohibitively for large-scale materials. In this work, we present a comprehensive, first-principles-derived database of two-atom interaction potentials and interatomic forces to accelerate the property prediction via physics-informed machine learning. Computed using spin-polarized DFT via the GPAW package, the database encompasses all 2346 possible atomic pairs generated from 68 unique elements as a function of interatomic distance. The practical utility of this database was demonstrated within the many-body expansion framework by engineering highly descriptive features to predict three distinct types of energy for benchmarking. First, the framework was applied to predict the total energy of molecules within the G2/97 database. Second, for defect energetics on 2D materials (the IMP2D database), localized concentric shell potentials and spatial gradients were introduced to capture the formation energy of adsorbate defects. Finally, for molecular energetics within the QM9 database, we developed advanced descriptors, including a potential-based matrix, a two-body Bag of Bonds, and a three-body Ray's asymmetric parameter histogram to evaluate atomization energy. Evaluated using kernel ridge regression, the combined two-body and three-body feature set yielded an outstanding test result with the mean absolute error of 0.186 eV (R2 = 0.9994) on a 10 000-samples subset. This framework showcases the power of utilizing explicit, database-derived potentials as robust descriptors for high-fidelity predictions of energy and relevant properties.