A Data-Driven Parametric Reduced-Order Chemical Kinetics Model Derived from Atomistic Simulations

Abstract

Coarse-grained modeling in molecular simulations serves not only to extend accessible time and length scales beyond atomistic limits, but also to reduce high-dimensional chemical data to low-dimensional representations that expose the underlying latent structure. In the context of energetic materials, reduced-order chemical kinetics models are essential for describing thermally driven decomposition, deflagration, and detonation. Recent data-driven approaches based on machine learning and dimensionality reduction have shown promise for constructing such models directly from atomistic simulations; however, when reaction pathways vary strongly with thermodynamic conditions, these methods can produce latent representations that are difficult to interpret physically or extrapolate reliably. Here, we introduce a parametric, temperature-dependent autoencoder framework that learns a unified reduced-order description of chemical decomposition across a wide range of temperatures within a single model. Physical interpretability is enforced through non-negativity constraints and a softmax activation, enabling the latent variables to be directly associated with additive chemical components and their relative contributions. Reaction kinetics and heat-release parameters are optimized simultaneously within the neural-network architecture, providing a self-consistent coupling between chemical evolution and energetics. The proposed approach yields significantly improved reconstruction accuracy compared to a state-of-the-art dimensionality-reduction method, as quantified by reductions in mean-squared error, while preserving a physically meaningful latent representation. These results demonstrate that parametric, interpretable machine-learning models can provide robust reduced-order chemical kinetics suitable for multiscale modeling of complex reactive systems.