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

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Accurate modeling of organometallic precursors is crucial for atomic layer deposition (ALD) thin-film fabrication.
• Experimental melting point measurements are challenging; computational methods offer a practical alternative.
• Density functional theory (DFT) is accurate but computationally expensive; machine learning interatomic potentials (MLIPs) offer a balance of accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate MLIPs for modeling cyclopentadienyl-based organometallic precursors.
• To compare the performance of different MLIP architectures: moment tensor potential (MTP), crystal Hamiltonian graph neural network (CHGNet), and scalable equivariance-enabled neural network (SevenNet).
• To assess the impact of active learning and fine-tuning on MLIP generalization and accuracy.

Main Methods:

• Development and evaluation of MLIPs including MTP, CHGNet, and SevenNet.
• Application of active learning and fine-tuning strategies using high-energy state configurations.
• Melting point calculations using the solid-liquid phase coexistence method to evaluate MLIP applicability.

Main Results:

• Message-passing neural network-based MLIPs (CHGNet, SevenNet) outperformed polynomial-based MTP.
• SevenNet demonstrated superior performance with exceptional accuracy in energy, force, and density predictions.
• SevenNet-based melting point calculations showed close agreement with experimental values for organometallic precursors.

Conclusions:

• Accurate and scalable MLIPs, particularly SevenNet, can effectively model organometallic precursors for ALD.
• The developed MLIPs enable efficient and reliable simulations, advancing the design of high-performance thin films.
• This study provides a framework for constructing advanced MLIPs for materials modeling in chemical processes.