MLIP-Enhanced Thermochemistry Predictions across Organic Chemical Space

Abstract

This work presents thermoMLIP, a machine-learning-interatomic-potential-based workflow for efficient conformer optimization and thermochemistry prediction across the organic chemical space. By combining MLIP-driven geometry optimization with robust thermochemical corrections, thermoMLIP achieves near-DFT accuracy at a fraction of the computational cost. We further show that MLIP-based conformer searches systematically identify lower-energy structures than conventional approaches, improving thermochemical predictions across diverse data sets. Although thermoMLIP substantially accelerates thermochemistry calculations, its computational cost remains prohibitive for the exhaustive screening of very large chemical spaces. We therefore employ an uncertainty-aware active-learning strategy based on evidential graph neural networks to efficiently generate a large, uniform thermodynamic data set, which is subsequently used to train thermoGNN, a surrogate capable of near-instantaneous prediction of molecular thermodynamic properties. Together, thermoMLIP and thermoGNN establish a hierarchical framework that bridges accurate MLIP-based thermochemistry and ultralarge-scale molecular screening, enabling applications such as reaction-network exploration and high-throughput reaction discovery.