A Machine Learning Framework for Scattering Kernel Derivation Using Molecular Dynamics Data in Very Low Earth Orbit

TL;DR

This paper introduces a machine learning framework that uses molecular dynamics data to derive more accurate scattering kernels for gas-surface interactions in very low Earth orbit, improving satellite aerodynamic modeling.

Contribution

The study develops a cVAE-based model trained on MD simulations to generate scattering kernels for any incident velocity, surpassing traditional models like Maxwell in accuracy.

Findings

cVAE accurately predicts the shift from diffuse to quasi-specular reflection.

Generated kernels lead to different aerodynamic coefficients compared to Maxwell model.

Application on a flat plate demonstrates improved modeling of gas-surface interactions.

Abstract

The free molecular flow regime in VLEO makes gas-surface interactions (GSIs) crucial for satellite aerodynamic modeling. The Direct Simulation Monte Carlo (DSMC) method is required to estimate aerodynamic forces due to the breakdown of the continuum assumption. In DSMC, the Maxwell model is the most widely used approach for GSI. It simplifies the process by treating it as a superposition of diffuse and specular reflections while assuming a constant accommodation coefficient. In reality, this coefficient is influenced by multiple factors, such as the angle and magnitude of the incident velocity. A high-precision GSI model could significantly improve satellite aerodynamics optimization and the design of efficient intakes for atmospheric breathing propulsion systems. This advancement would greatly refine mission planning and fuel requirement calculations, ultimately extending operational lifetimes and lowering costs. To gain a deep understanding of the GSI at the microscopic level, molecular dynamics (MD) simulations provide valuable insights into the physical processes involved. However, due to computational limitations, simulating an entire satellite is impractical. Instead, we use MD to analyze the impact of selected velocity vectors on a amorphous $\text{Al}_2\text{O}_3$ surface. The obtained scattering kernels for the respective velocity vectors are then used to train a conditional Variational Autoencoder (cVAE). This model is able to generate scattering kernels for any incident velocity vector and can be integrated into DSMC simulations, significantly enhancing their accuracy. Applications of this model on a flat plate have shown that the cVAE is able to predict the shift from diffuse to quasi-specular reflection with increasing polar angle. Additionally, the aerodynamic coefficients and molecular fluxes are considerably different from those obtained with the Maxwell model.