Prediction of thermal conductance and friction coefficients at a solid-gas interface from statistical learning of collisions

In this paper, we present the construction of statistical models of gas-wall collision based on data issued from molecular dynamics (MD) simulations and use them to predict the velocity slip and temperature jump coefficients at the gas-solid interface. The Gaussian mixture (GM) model, an unsupervised learning technique, is chosen for this purpose. The model shares some similarities with the well-known Cercignani-Lampis model in kinetic theory but it is more robust due to the unlimited number of Gaussian functions used and the ability to deal with correlated data of high dimensions. Applications to real gas-wall systems (argon-gold and helium-gold) confirm the good performance of the model. The trained GM model predicts physical and statistical properties including accommodation, friction, and thermal conductance coefficients in excellent agreement with the MD model.