This research aims to advance the atomistic understanding of dry-etching processes crucial for semiconductor manufacturing. Current molecular dynamics (MD) simulations are limited by the lack of reliable force fields, hindering their application in etching studies. To address this, a neural network potential (NNP) was developed to simulate the etching of amorphous Si₃N₄ with HF molecules. The NNP was trained using surface reactions in diverse local environments: baseline structures, reaction-specific data, and general-purpose training sets; and refined through iterative comparisons with density functional theory (DFT) results to ensure high accuracy. The trained NNP allowed for detailed simulations of key aspects such as preferential sputtering, surface modification, etching yield, threshold energy, and product distribution. A continuum model, derived from MD results, was also developed to replicate surface composition changes effectively. This study establishes a robust computational framework for atomistic etching simulations, offering a pathway to more accurate and efficient etching process designs for the semiconductor industry. This work was published in ACS applied Materials & Interfaces 2024, 16, 48457.

Solid-state electrolytes with argyrodite structures, have attracted considerable attention due to their superior safety and higher ionic conductivity than other solid electrolytes. Although experimental efforts have been made to enhance conductivity by disorder engineering, the underlying diffusion mechanism is not yet fully understood. Moreover, existing theoretical analyses based on ab initio molecular dynamics (MD) simulations have limitations in addressing various types of disorder at room temperature. In this study, we directly investigate Li-ion diffusion in Li₆PS₅Cl at 300 K using large-scale, long-term MD simulations empowered by machine-learning potentials (MLPs). The computed Li-ion conductivity, activation energies, and equilibrium site occupancies align well with experimental observations. Notably, Li-ion conductivity peaks when Cl ions occupy 25% of the 4c sites. In addition, Li-ion diffusion shows non-Arrhenius behavior, leading to different activation energies at high temperatures. These phenomena are explained by the interplay between inter- and intra-cage jumps. By elucidation of the key factors affecting Li-ion diffusion in Li₆PS₅Cl, this work paves the way for optimizing ionic conductivity in the argyrodite family. This study was published in ACS Applied Materials & Interfaces 2024, 16, 46442.

Message-passing graph neural network interatomic potentials (GNN-IPs), particularly those with equivariant representations, are attracting significant attention due to their high accuracy. However, parallelizing GNN-IPs in molecular dynamics simulations poses challenges due to their multiple message-passing layers. In this article, we propose an efficient parallelization GNN-IPs and develop a package, SevenNet (Scalable EquiVariance-Enabled Neural NETwork). Through benchmark tests, SevenNet achieves over 80% parallel efficiency in weak-scaling and exhibits nearly ideal strong-scaling performance as long as GPUs are fully utilized. We also develop a general-purpose GNN-IP, SevenNet-0, which is trained on a vast data set from the Materials Project. By parallelizing GNN-IPs, this work aims to bridge the gap between advanced machine-learning models and large-scale MD simulations. This study was published in Journal of Chemical Theory and Computation 2024, 20, 4857.

Developing a machine learning model capable of predicting the melting points of inorganic crystals allows for rapid melting point predictions, eliminating the need for experimental synthesis and complex theoretical calculations. We improve the accuracy of these predictions on a data-scarce melting point database by applying transfer learning to a crystal graph neural network. By pretraining the model with a large atomization energy database and then transferring this learning to the melting temperature database, we achieve a satisfactory root mean squared error of 218 K. Additionally, we find that the accuracy improvement for each crystal category is significantly influenced by the physical relationships between the pretraining database and the target property. This work was published at Computational Materials Science 2023, 234, 112783