A Machine Learning Framework for the Quantification of the Uncertainties Associated with Ab-Initio Based Modeling of Non-Equilibrium Flows

摘要

In this paper we presented a preliminary framework for quantifying the uncertainties in the ab-initio models predictions of non-equilibrium kinetic processes. The accuracy in estimating quantities at a flow macroscopic level has assumed to be mainly influenced by the fidelity of the PESs in capturing the interaction dynamics between particles, and by the legitimacy of the kinetic reduced order model in assuming local equilibrium between the quantum states. In order to model the former source of uncertainty, a Bayesian extension of the Permutation Invariant Polynomials Neural Networks has been developed, and stochastic PESs have been constructed. For improving and generalizing the grouping strategy at the base of coarse- grained models, instead, a diatomic potential based approach assisted by network analysis has been presented. Finally, the authors recommend the adoption of the level energy distance from the J-dcpcndent centrifugal barrier as guiding quantity for binning the high energy rovibrational levels, including the quasi-bound ones.