Benchmarking CHGNet Universal Machine Learning Interatomic Potential against DFT and EXAFS: The Case of Layered WS₂...

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Universal machine learning interatomic potentials (uMLIPs) offer efficient, accurate energy and force calculations for materials modeling.
• Rigorous validation of uMLIPs is crucial for their widespread adoption in scientific research.
• Thermal disorder significantly impacts material properties and requires accurate modeling.

Purpose of the Study:

• To assess the performance of the CHGNet uMLIP for modeling thermal disorder in layered 2H_c-WS₂ and 2H_c-MoS₂.
• To investigate the impact of fine-tuning uMLIPs with density functional theory (DFT) data on accuracy.
• To benchmark uMLIP predictions against experimental extended X-ray absorption fine structure (EXAFS) spectra.

Main Methods:

• Utilized the CHGNet universal machine learning interatomic potential (uMLIP).
• Modeled thermal disorder in 2H_c-WS₂ and 2H_c-MoS₂ using molecular dynamics simulations.
• Fine-tuned the uMLIP with varying amounts of compound-specific ab initio (DFT) data.
• Compared simulation results with ab initio calculations and experimental extended X-ray absorption fine structure (EXAFS) spectra.

Main Results:

• Fine-tuning CHGNet with DFT data reduced the systematic force underestimation (softening) inherent in uMLIPs.
• Optimized fine-tuning significantly improved the agreement between molecular dynamics-derived and experimental EXAFS spectra.
• Using approximately 100 DFT structures for fine-tuning is recommended for high accuracy and reliable EXAFS reproduction.

Conclusions:

• Fine-tuning universal machine learning interatomic potentials (uMLIPs) with targeted DFT data is essential for accurate materials modeling.
• The CHGNet uMLIP, when appropriately fine-tuned, can reliably model thermal disorder and thermal variations in bond lengths and angles.
• This study provides a framework for optimizing fine-tuning strategies to achieve ab initio-level accuracy in uMLIP applications.