First Large-Scale (68 × 25 × 5 nm3) Atomistic Modeling for Accurate and Efficient Etching Process Based on Machine Learning Molecular Dynamics (MLMD)

Revealing the reaction mechanisms in manufacturing advanced devices is crucial for enhancing performance at nodes below 3 nm. In this study, we introduce the first large-scale atomistic model utilizing a machine learning molecular dynamics (MLMD) framework. This model simulates the lateral isotropic selective etching process of SiO2/Si stack structures in Gate-all-around field-effect transistors (GAAFETs) using F* radical plasma. The dataset for this model is generated using active learning strategies based on search for F* radical reacting with Si(100) substrate. Our simulation spans $68\times 25\times 5\ \text{nm}^{3}$, involving over 600,000 atoms, and demonstrates a $4\times$ improvement in model error for etching reaction trends compared to the classic molecular dynamics (CMD) method. Additionally, the model achieves a typical runtime of 16.7 hours, representing a $\sim 44.9\times$ speedup over CMD. Moreover, the simulation provides detailed reaction path and uncovers the mechanisms behind lateral etching. This framework paves the way for potential atom-to-device multi-scale process modeling, contributing to Design Technology Co-optimization (DTCO).