Constructing Accurate\nand Efficient General-Purpose\nAtomistic Machine Learning Model with Transferable Accuracy for Quantum\nChemistry

Density\nfunctional theory (DFT) has been a cornerstone\nin computational\nscience, providing powerful insights into structure–property\nrelationships for molecules and materials through first-principles\nquantum-mechanical (QM) calculations. However, the advent of atomistic\nmachine learning (ML) is reshaping the landscape by enabling large-scale\ndynamics simulations and high-throughput screening at DFT-equivalent\naccuracy with drastically reduced computational cost. Yet, the development\nof general-purpose atomistic ML models as surrogates for QM calculations\nfaces several challenges, particularly in terms of model capacity,\ndata efficiency, and transferability across chemically diverse systems.\nThis work introduces a novel extension of the polarizable atom interaction\nneural network (namely, XPaiNN) to address these challenges. Two distinct\ntraining strategies have been employed, one direct-learning and the\nother Δ-ML on top of a semiempirical QM method. These methodologies\nhave been implemented within the same framework, allowing for a detailed\ncomparison of their results. The XPaiNN models, in particular the\none using Δ-ML, not only demonstrate competitive performance\non standard benchmarks, but also demonstrate the effectiveness against\nother ML models and QM methods on comprehensive downstream tasks,\nincluding noncovalent interactions, reaction energetics, barrier heights,\ngeometry optimization and reaction thermodynamics, etc. This work\nrepresents a significant step forward in the pursuit of accurate and\nefficient atomistic ML models of general-purpose, capable of handling\ncomplex chemical systems with transferable accuracy.