Conceptual

Machine-Learning Surrogate Models for Sticking Probabilities in Nonthermal Plasma Nanoparticle Growth

Reactive molecular dynamics simulations of colliding silane fragments are expensive, so this concept builds machine-learning surrogates that predict sticking (reaction) probabilities directly from cheap molecular features such as atom counts, per-atom unpaired-electron vectors, and translational temperature. It shows that choosing loss functions matched to the binomial nature of each collision outcome and imposing the correct permutation invariances (as in DeepSets) yields accurate predictions, and that only 15-25% of the energy and temperature sampling is needed for high accuracy.