Application of DeepOnet to model inelastic scattering probabilities in air mixtures

Hypersonic conditions during (re-)entry flight demand accurate representation of the nonequilibrium flow physics to enable predictive models. However, the large number of possible kinetic mechanisms for even small molecular systems makes enumerating all the possible reaction coefficients and quantifying their uncertainties a computationally challenging task. Addressing this concern, in this work we develop surrogate models for molecular scattering or quasi-classical trajectory calculations. Deep Operator network (DeepOnet), recently developed my Lu et al. [1], is employed to model the inelastic reaction rates. The state-to-state (StS) rates are obtained as a function of temperature and properties of the molecular diatomic potential. The training states used by the NN are randomly sampled from groups constructed using a novel diatomic potential-based grouping strategy. As an example, the surrogate QCT model is tested for the O₂ +O system due to its interest in hypersonic earth re-entry flight. We have demonstrated a reduction in the computational costs of QCT calculations by 95% while maintaining a good accuracy on the predicted StS rates. To further validate the rates, we carried out isothermal heat bath simulations at translational temperatures between 1,500 K and 20,000 K. The observables from the master equation analysis were reproduced within an accuracy of 5%.