Atomistic Simulation of Reactive Ion Etching using Machine Learning Interatomic Potentials

Maintaining leadership in semiconductor manufacturing relies on plasma etching however capturing complex plasma-surface interactions remains difficult due to limited resolution of experiments. While molecular dynamics (MD) provides detailed insights, it lacks reliable interatomic potentials for diverse chemistries. Machine learning interatomic potentials (MLIPs) offer near-DFT accuracy at low cost, making them promising for etching simulations, but generating relevant DFT training data is still a major challenge.

This work presents a framework for generating and validating MLIPs to simulate reactive ion etching MD and bridge atomic-scale processes to larger scales. The methodology captures surface reactions during HF etching of Si₃N₄ by incorporating diverse training sets: baseline structures, reaction-specific data, and general-purpose datasets. The MLIPs are iteratively refined using density functional theory (DFT) results from MLIP-driven MD trajectories. Using the trained potentials, key processes such as surface modification and preferential sputtering are analyzed.

To expand chemical coverage, vapor-to-surface (VTS) MD simulations are conducted across a range of compositions and densities, capturing configurations representative of bulk, interfacial, and surface environments. These configurations, along with baseline structures and iterative training, are used to develop two MLIPs for the Si₃N₄/CH_xF_y and SiO₂/CH_xF_y systems. The resulting models show good agreement with experimental etch yields and reproduce surface height evolution.

To address the high computational cost of DFT data generation, a pretrained model incorporating a smooth transition to the ZBL potential at short ranges is developed. This model maintains ZBL repulsion and correlates well with both small- and large-cell MD results, allowing primary datasets to be bypassed and accelerating MLIP deployment.

Finally, a predictive 1-D continuum model, approximating Si₃N₄/HF etching as a first-order reaction, reproduces the transient dynamics of surface composition. By integrating MLIP development with multiscale modeling, this framework deepens understanding of plasma-surface interactions and contributes to advances in semiconductor manufacturing.