Equipment-Informed Machine Learning-Assisted Feature-Scale Plasma Etching Model

We investigate means to merge feature-scale and reactor-scale models during plasma etching using Machine Learning (ML) and interpolative approaches. First, we test the SF6 plasma etching models based on a small dataset from literature. We find that Gaussian Process Regression (GPR) leads to significant over-fitting, resulting in waviness in the predicted values in the range where no data is available. A Neural Network (NN) model was likewise implemented with rectified linear unit activation. This model provides linear prediction between known values, resulting in a better qualitative fit to experimental observations. Finally, we perform 18 750 chamber simulations of a $\mathrm{C}1_{2}/\text{Ar}$ plasma while varying relevant input parameters. The data is used to build a six-dimensional spline-based inerpolative model of the chamber and provide a means to quickly extract relevant fluxes for the feature-scale model.