Development of a scattering model for diatomic gas-solid surface interactions by an unsupervised machine learning approach

Abstract

This work proposes a new stochastic gas-solid scattering model for diatomic gas molecules constructed based on the collisional data obtained from molecular dynamics (MD) simulations. The Gaussian mixture (GM) approach, which is an unsupervised machine learning approach, is applied to H2 and N2 gases interacting with Ni surfaces in a two-parallel wall system under rarefied conditions. The main advantage of this approach is that the entire translational and rotational velocity components of the gas molecules before and after colliding with the surface can be utilized for training the GM model. This creates the possibility to study also highly nonequilibrium systems and accurately capture the energy exchange between the different molecular modes that cannot be captured by the classical scattering kernels. Considering the MD results as the reference solutions, the performance of the GM-driven scattering model is assessed in comparison with the Cercignani-Lampis-Lord (CLL) scattering model in different benchmarking systems: the Fourier thermal problem, the Couette flow problem, and a combined Fourier-Couette flow problem. This assessment is performed in terms of the distribution of the velocity components and energy modes, as well as accommodation coefficients. It is shown that the predicted results by the GM model are in better agreement with the original MD data. Especially, for H2 gas the GM model outperforms the CLL model. The results for N2 molecules are relatively less affected by changing the thermal and flow properties of the system, which is caused by the presence of a stronger adsorption layer.