Geometry-Based Atomic Representations for Reactive Combustion Chemistry

Abstract

Reactive combustion trajectories contain atomic configurations in which bonds continuously stretch, break and form. This behavior challenges molecular representations based on fixed covalent graphs. Here, we present CombNetPack, a geometry-based deep learning framework based on local atomic environments. CombNetPack combines graph-atomic features, attention-enhanced continuous-filter convolutions and a Koopman autoencoder module to encode transient local interactions in reactive trajectories. We further constructed the HAC database, which contains 96,000 ab initio molecular dynamics (AIMD) snapshots from eight ammonia–hydrogen combustion reaction channels. Together with external hydrogen-combustion and reactive molecular-dynamics data sets, we evaluate CombNetPack against graph-based and selected neural-potential models. Feature-importance analysis, ablation studies and transfer-learning experiments reveal that the learned representation captures reaction-sensitive local interactions and transfers them to related, larger combustion-relevant systems. The framework also supports supervised prediction of fuel-related properties, including octane numbers, thermochemical quantities, heat of vaporization and laminar flame speed. The results suggest that geometry-based atomic representations provide a data-efficient route for precision modeling of nonequilibrium reactive systems.