Generalizable Prediction of Mixture Etching Rates Using Graph Neural Networks

Plasma-based etching and thin-film processing rely on high plasma densities and independent ion energy control to achieve high etching rates and anisotropy. However, accurately predicting spatially varying etching rates remains challenging. Physics-based models (PBMs) can capture the complex plasma dynamics, but they are computationally prohibitive due to the need to solve large systems of partial differential equations, especially when optimization tasks require repeated evaluations. Moreover, plasma processes depend strongly on gas chemistry, reactor configurations, and operating conditions, requiring separate PBMs for each scenario. The complexity is further amplified in the case of gas mixtures. To address this challenge, we propose an architecture that leverages pre-trained single-element neural network predictors, coupled through an inductively learned Graph Neural Network, GraphSAGE, to predict etching rates of mixtures. GraphSAGE enables inference on unseen graphs without retraining, making it possible to extend predictions to new mixtures using only the pool of pre-trained single-element models. We evaluate our approach on a two-gas argon-oxygen mixture, demonstrating promising accuracy and generalization capabilities.