Molecular Dynamics Simulation of HF Etching of Amorphous Si₃N₄ Using Neural Network Potential

Dry etching plays a key role in the fabrication of semiconductor devices. Recently, the requirement of high aspect ratios (40–60) of deep-trench structures and the introduction of through-silicon vias for 3D circuit integration demand precise etching control. Traditional macroscale models, such as the level set and Monte Carlo-based methods, have been employed to predict the outcomes of process parameter optimization. However, phenomenological fitting of model parameters can result in inaccurate predictions of feature profiles due to the omission of detailed chemical processes. Molecular dynamics (MD) simulations can offer a close look into atomic processes in ion etching, overcoming the trial-and-error approaches for parameter determination in macroscopic interaction models. Classical potentials are an option but have limitations in both availability and accuracy. Density functional theory (DFT), although highly accurate, is computationally prohibitive for large-scale simulations. Recent advances in machine-learning potentials (MLPs) have demonstrated their efficacy in simulations involving surface catalysis, physical sputtering, and surface oxidation processes. However, application of MLPs for etching reactions between ions and solid-state materials is currently lacking. This is attributed to the complexity of predicting chemical reaction pathways, which creates challenges in constructing training sets for MLPs.<br/>In this study, we investigate the mechanisms underlying Si₃N₄ etching by HF ions using MD simulations driven by a neural network potential (NNP). The NNPs are trained using SIMPLE-NN package. The training set is constructed through manual curation of relevant atomic environments and an iterative training process. The basic constituents of atomic configurations are both crystalline and amorphous bulk structures, as well as their slab counterparts. Additionally, guided MD is employed to sample the chemical reactions between HF molecules and the Si₃N₄ substrate; here, hydrogen and fluorine are displaced toward nitrogen and silicon respectively based on the assumption of bond preference. Alongside guided MD, high-temperature MD of a slab is used to facilitate the sampling of diverse local atomic environments that may emerge during energetic ion collisions. Using manually curated training sets, NNPs are iteratively validated and refined. This refinement is achieved by incorporating structures from NNP trajectories that exhibit significant energy discrepancies when compared against DFT calculations. Utilizing this optimized NNP, we conduct etching simulations on 4 nm² amorphous Si₃N₄ substrates over a 5 ns timescale, and analyze characteristics such as etching yield and surface modification. Our results demonstrate a square-root dependence of the etching yield on the incident energy. Furthermore, we observe modifications at the topmost surface, where the ratio of Si to N is 1:1. This ratio is intriguingly similar to observations from experimental studies utilizing different reactive plasma species. The primary etch products at normal incidence are SiF₂, SiF₄, N₂, and NH₃. This behavior is attributed to chemical sputtering, in which the surface is significantly modified by incoming reactive species, becoming susceptible to chemical etching. In contrast, at high angles of incidence, the surface remains largely unmodified, with the primary etch products being lower-coordinated molecules of Si and N, along with N₂ molecules. This suggests that under these conditions, the etching is more affected by physical sputtering. Our methodology offers a cost-efficient approach to achieving DFT-level accuracy in MD simulations, thereby elucidating etching mechanisms and contributing to the optimization of the etching process.