State-specific dissociation dynamics on a global potential energy surface of N(4Su) + C2(a3Πu): Machine...

The dissociation dynamics of N(4Su) + C2 (a3Πu) under hypersonic conditions is critical for modeling radiative heating in aerospace thermal protection systems, yet it remains unexplored due to computational limitations. This work focuses on the systematic study of N + C2 collision-induced dissociation (CID) processes using molecular dynamics simulations with 50 000 quasi-classical trajectories (QCT) per rovibrational state on a global 12A″ potential energy surface to obtain cross sections (CS) and state-specific thermal rate coefficients (1000-20000 K). The results demonstrate that vibrational excitation dominates dissociation dynamics, with vibrational quantum states significantly lowering energy barriers and facilitating bond dissociation at reduced collision energies. In addition, we integrate quasi-classical trajectory simulations with neural networks, utilizing simulation datasets per rovibrational state to train our genetic-algorithm-optimized neural network. This framework achieves 99% prediction accuracy (R2 = 0.99) across all validation sets for state-specific dissociation CS and rate coefficients. We develop an effective CS model through machine learning (ML) to generate rate coefficient datasets covering all possible rovibrational levels. This machine learning approach reduces computational costs by three orders of magnitude compared to direct QCT calculations. Finally, this study further examines the proposition by Truhlar et al. regarding the influence of multi-electronic-state potential energy surfaces on the dissociation rate coefficients, including the application of modified rate-coefficient formulas to the current system. This study bridges the gap in molecular dynamics (MD) simulations for this system. Furthermore, it demonstrates that the combined MD and machine learning approach can generate comprehensive datasets from limited discrete data points. This creates a high-quality database for modeling non-equilibrium hypersonic flows and highlights the benefits of machine learning in reactive MD simulations.