DeepTherm: A unified deep learning approach to thermochemistry prediction for gas-phase molecules

• Innovative architecture: Propose directed bond-centered message-passing process combined with global attention mechanisms to capture both local and long-range molecular interactions, enabling accurate property predictions for complex gas-phase molecules and radicals. • Integrated learning techniques: Transfer learning, ensemble methods, and molecular descriptors enhance accuracy and robustness. • Comprehensive thermochemistry dataset: 4,997 (closed/open-shell) species with diverse functional groups and molecular complexities. • Rigorous evaluation protocols: Introduce hierarchical evaluation (scaffolds, functional groups, molecular complexities) and outperforms 8 baseline models with 72 model-descriptor combinations. • Practical applications: Predict thermochemical properties for 11,458 species in 40 oxidation and pyrolysis reaction mechanisms of hydrocarbons. Validation against experimental ignition delay times of n-heptane showcases its reliability for kinetic modeling. Accurate thermochemical properties are pivotal in chemical mechanisms for modeling and optimizing complex reactive systems, as they directly impact reaction kinetics and heat release profiles. However, obtaining reliable data for a set of diverse, large, and complex molecules and radicals is challenging due to the limitations of traditional methods. In this work, we introduce DeepTherm, a unified deep learning framework that combines the directed message-passing process for local information and global attention mechanisms for long-range dependencies, effectively addressing the challenges posed by multiple functional groups and large species. DeepTherm integrates graph encoding, transfer learning, pre-defined molecular descriptors, and ensemble learning to achieve high accuracy and robust generalization. A curated dataset of 4,997 species, including closed-shell and open-shell species consisting of C, H, and O atoms, was derived from recent theoretical calculations for model training and testing. The model was rigorously evaluated using both random splits and hierarchical splitting strategies based on molecular scaffolds, functional groups, and molecular complexity. Comparative studies involving 72 model-descriptor combinations demonstrated the superior performance of DeepTherm. To showcase its applications, thermochemical properties for 11,458 species present in the oxidation and pyrolysis reaction mechanisms of 40 hydrocarbons are predicted. The resulting thermochemical data improved the fidelity of detailed gas-phase kinetic simulations compared with experimental measurements. The predictions also showed clear advantages over those based on group-additivity methods. Beyond thermochemical properties, the proposed architecture and unified framework establish a generalizable paradigm for molecular property prediction tasks, offering a scalable foundation for accelerating chemical kinetic modeling and materials discovery.