Evaluating Machine Learning Interatomic Potentials for Accurate and Scalable Modeling of Organometallic Precursors

Accurate modeling of organometallic precursors is essential for developing atomic layer deposition (ALD) processes that are required for fabricating high-performance thin films. The melting point of these precursors is often challenging to measure experimentally due to its sensitivity to environmental conditions, making computational methods a practical alternative. While density functional theory (DFT) is reliable, its high computational cost limits large-scale applications. In contrast, machine learning interatomic potentials (MLIPs) offer an attractive alternative by providing both high accuracy and computational efficiency. In this study, we developed and evaluated MLIPs for modeling cyclopentadienyl-based organometallic precursors, comparing the performance of moment tensor potential (MTP), crystal Hamiltonian graph neural network (CHGNet), and scalable equivariance-enabled neural network (SevenNet). The impact of active learning and fine-tuning with high-energy (HE)-state configurations was further investigated to improve generalization. CHGNet and SevenNet, both based on message-passing neural networks (MPNN), outperformed the polynomial-based MTP. Notably, SevenNet demonstrates superior performance with a mean absolute error (MAE) of 0.001 eV/atom for energy, R 2 of 1.000 for force, and a density MAE of 0.033 g/cm 3, enabled by incorporation of rotational symmetries via E(3)-equivariant convolutions. Melting point calculations were performed using the solid–liquid phase coexistence method to assess MLIP applicability. SevenNet showed the closest agreement with the experimental values, with a deviation of 0.65–2.05% for CUFBAG, while deviations for other precursors ranged from 1.87% to 7.58%. This study provides a methodological framework for constructing accurate and scalable MLIPs, enabling efficient and reliable simulations of organometallic precursors in ALD applications.