Deep potential modeling of oxygen adsorption and surface reconstruction on silica.

The binding energy of atomic oxygen on silica has remained uncertain for decades. This quantity directly governs catalytic heating during atmospheric re-entry. Classical potentials overestimate it by an order of magnitude (ReaxFFSiOGSI: ∼31 eV vs experiment: ∼3 eV), while first-principles methods cannot reach the scales required for adsorption equilibration. Although neural network potentials bridge this accuracy-efficiency gap, conventional sampling fails to provide training data spanning rare reactive events within collective phase transitions. Here, we train a neural-network potential (HyGSI) using a hybrid enhanced sampling-machine learning framework coupling on-the-fly probability enhanced sampling with deep potential molecular dynamics. HyGSI covers SiO2 melting, pressure-dependent surface reconstruction, and oxygen adsorption-recombination. The predicted adsorption energy is 2.96 ± 0.15 eV, consistent with the 1991 experimental estimate (3.00 ± 0.31 eV) and an order of magnitude lower than ReaxFFSiOGSI predictions. Furthermore, surface reconstruction reshapes available catalytic sites, coupling structural evolution to reaction chemistry. These findings provide corrected energetics for hypersonic re-entry heat-flux models and validated force fields for semiconductor plasma etching and fusion reactor wall modeling.