Toward fast and accurate machine learning interatomic potentials for atomic layer deposition precursors

Abstract

Under thin film deposition, when used in conjunction with the semiconductor atomic layer deposition (ALD) method, the choice of precursor determines the properties and quality of the thin film. Organometallic precursors such as alkaline earth metals (Sr and Ba) and group 4 transition metals (Zr and Hf) with cyclopentadienyl and tetrakis (ethylmethylamino) ligands have recently gained attention for their stable deposition within high-temperature windows in the ALD. The design of organometallic precursors with an ab initio molecular dynamics (AIMD) simulations-based approach ensures high accuracy but comes with significant computational costs. In this study, we aim to develop a machine-learning interatomic potential (MLIP) through moment tensor potential (MTP) for fast and accurate potential development of Sr, Ba, Zr, and Hf precursors. To establish the reliable training database for MTP construction, we conducted AIMD simulations on each precursor across a range of temperature settings, resulting in a variety of atomic structures. Constructed MTPs enable efficient utilization of molecular dynamics (MD) simulations as well as calculations that achieve an accuracy that approximates density functional theory (DFT). MTP construction coupled with active learning ensures that the MTP for each precursor is reliable and that databases can be expanded. High prediction accuracy is demonstrated by a mean absolute error (MAE) of less than 0.04 eV/atom in all structures. In addition, generalization performance is confirmed for general structures (structures with the same chemical elements but different proportions) and is extended to cluster structures. The constructed MTP exhibits an MAE of less than 0.15 eV/atom, even for untrained cluster structures. These results demonstrate adequate representation and scalability as a basis for the development of MLIPs capable of atomic simulations of organometallic precursors under various thermodynamic conditions.

Keywords

• Organometallic precursors
• Machine learning interatomic potential
• Moment tensor potential
• Ab initio molecular dynamics