Rapid and accurate prediction of molecular dynamics simulations using physics-informed LSTM networks in engine emission analysis: A case study of C3H6/NH3 pyrolysis for PAH formation

Molecular dynamics (MD) simulations are essential tools for analyzing internal combustion engine emissions, particularly in the study of polycyclic aromatic hydrocarbon (PAH) and soot formation; however, these simulations are computationally intensive, requiring significant resources and time. Long Short-Term Memory (LSTM) networks offer an efficient alternative for modeling time-dependent, strongly coupled, and high-dimensional chemical processes, enabling faster predictions without sacrificing accuracy. This study explores the feasibility of using LSTM networks to predict MD simulation results in the context of engine emissions, an area where the application of time-series deep learning models remains limited, by simulating PAH formation through the pyrolysis of C 3 H 6 and NH 3 blends, a process characteristic of the localized oxygen-deficient environments in compression ignition engines . The results show that the LSTM model, trained on data from multiple C 3 H 6 /NH 3 blends, can predict species count variations for unseen blends, demonstrating strong potential for reducing computational costs. To improve model reliability and ensure adherence to conservation laws, physical constraints are incorporated into the loss function during training. Comparison of LSTM and physics-informed LSTM (PI-LSTM) performance reveals that integrating carbon balance constraints, a critical factor in internal combustion engine research, limits fluctuations in total carbon count, addressing the limitations of purely data-driven models. While such an innovative approach introduces a trade-off between prediction accuracy for individual species and physical consistency, it enhances the model overall reliability. Overall, this study demonstrates the potential of combining machine learning, particularly PI-LSTM, with MD simulations to reduce computational costs and maintain predictive accuracy in internal combustion engine emission research, offering the engine research community a powerful and transferable tool for tackling complex combustion challenges. • Time-series machine learning models capture dynamic chemical behaviors in MD simulations. • LSTM networks predict PAH formation in fuel pyrolysis reducing MD simulation costs. • Physics-informed LSTM enforces carbon balance reducing total carbon count fluctuations. • PI-LSTM model introduces physical constraints but may slightly reduce species accuracy. • The weighting factor balances physical constraints and data fitting for optimal performance.