Development of a deep potential model for F and CF2 etching of Si and SiO2

An understanding of plasma-surface interactions at increasingly smaller scales is invaluable for the development of novel technologies and processing techniques. Molecular dynamics (MD) simulations can provide insights into atomic-scale interactions, though they are restricted by the availability of interatomic potentials. Machine learning methods, such as Deep Potential Molecular Dynamics (DeepMD), provide a systematic framework for the development of accurate and flexible ab initio-based models. In this work, we develop DeepMD models for the ion-enhanced etching of Si and SiO2 by F and CF2 radicals. We employ an active learning process to expand the data set on which the model is trained and demonstrate its effect on the model accuracy. The DeepMD results are compared to data from classical MD simulations and experiments. Physical sputtering yields of SiO2 by Ar+ ions show good agreement with previous simulation results using conventional interatomic potentials, though the predicted depth profiles are different. Etching yields are calculated as a function of ion energy and neutral to ion flux ratio for the Ar+ ion-enhanced etching of SiO2 and Si by F atoms, as well as for etching of SiO2 by CF2 radicals, showing reasonable agreement with experimental data. Finally, an ion-enhanced surface kinetic model is fitted to the DeepMD etch yields, and the fitted parameters are compared to quantities computed directly from DeepMD simulations. This study illustrates how molecular dynamics simulations using machine learning potentials can provide an accurate model of etching processes relevant to device manufacturing.