Data-driven neural surrogates for ReaxFF molecular dynamics simulations in engine-relevant combustion chemistry

Machine learning (ML) has gained significant traction in engine-related research, particularly because of its potential to improve predictive performance while reducing computational costs. However, most current applications rely on feedforward neural networks (FNNs; e.g., conventional artificial neural networks (ANNs)); these are well-suited for modeling static data and capturing nonlinear relationships, but do not explicitly encode temporal dependencies unless sequence context is introduced via feature engineering. Motivated by this limitation, we evaluate sequence-aware neural surrogates for engine-relevant combustion-chemistry time-series data. Specifically, the temporal evolution of an intermediate product group during polycyclic aromatic hydrocarbon (PAH) formation in C2H4/NH3 pyrolysis is modeled using reactive force field (ReaxFF) molecular dynamics (MD) trajectories, comparing an FNN baseline (with explicit time as an input) against a long short-term memory (LSTM)-based recurrent neural network (RNN). The results show that while the FNN baseline benefits from explicit temporal feature engineering, its predictive performance is inferior to that of the LSTM model, even when the network depth is increased. This behavior is consistent with the architectural limitations of feedforward models, which do not maintain an internal memory state, and therefore, tend to generalize poorly when the target dynamics are history dependent. In contrast, the LSTM model leverages gated memory to learn temporal dependencies and consequently improves the predictive accuracy of combustion-chemistry time-series modeling, providing an efficient surrogate once trained. Overall, our findings delineate the conditions under which sequence-aware recurrent architectures offer advantages over feedforward models for ReaxFF MD time-series surrogate modeling.