Machine-Learning-Driven Simulations of Hyperthermal Atomic Oxygen Impacts on (0001) Al2O3 for Low-Altitude Satellite Design.

In Very Low Earth Orbit (VLEO), spacecraft encounter residual atmosphere primarily composed of atomic oxygen (AO). The drag and operational lifetime of these spacecraft are determined by the interaction between AO and the surface, which for aluminum is a passivation layer of Al2O3. We present a blueprint for developing machine-learned interatomic potentials for reactive gas-surface systems. Gaussian-Moment Neural Networks, implemented in apax, are trained using an active learning workflow that incorporates a spin-state switching scheme. This approach accounts for the triplet-to-singlet transition that AO undergoes upon adsorption, a process usually neglected in machine-learned potential studies. This methodology enables the simulation of thousands of hyperthermal impacts on (0001) Al2O3 at orbital velocity with near-first-principles accuracy for systems four times larger and four orders of magnitude cheaper than ab initio molecular dynamics. Alumina is not an inert surface. Most impacting atoms adsorb rather than reflect, and with continuous exposure, the surface reaches a dynamic steady state where ∼14% of the binding sites are occupied. The surface remains intact. Atoms desorb primarily as O2 formed by recombination, indicating that surface chemistry, rather than simple scattering, governs the energy transfer during impact. Conventional spacecraft drag models assume that impacting atoms desorb thermally in equilibrium with the surface. Our findings demonstrate that this assumption is invalid for alumina. We determine an energy accommodation coefficient of 0.61, which is lower than the 0.82 to 1.0 range measured on actual spacecraft surfaces. This discrepancy quantifies the effect of technical surface roughness. When scaling the simulated momentum transfer to the European Space Agency's GOCE satellite, the measured reentry drag force is reproduced within a factor of 1.8. Since an atomically smooth surface still exhibits strong energy accommodation through chemical bond formation, effective drag reduction in VLEO requires materials that are both atomically smooth and chemically inert.