Benchmarking structural evolution methods for training of machine learned interatomic potentials.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Training data generation is crucial for developing accurate machine-learned interatomic potentials (MLIPs).
• Molecular dynamics (MD) is a common method for evolving structures to sample configuration space for MLIP training data.
• The diversity and robustness of training data significantly impact MLIP performance.

Purpose of the Study:

• To benchmark contour exploration (CE) and dimer-method (DM) searches against MD for generating training data for MLIPs.
• To evaluate the ability of different structural evolution methods to produce diverse and robust density functional theory (DFT) training datasets.
• To formalize the process of generating initial structures for MLIP training data.

Main Methods:

• Benchmarking structural evolution methods: MD, CE, and DM.
• Utilizing Behler-Parrinello neural networks as MLIP models.
• Employing the polymorph-rich zirconium-oxygen system as a benchmark.
• Analyzing spatial descriptor diversity and statistical accuracy of generated datasets.

Main Results:

• CE and DM searches generally outperform MD in generating diverse and robust DFT training data.
• The choice of structural evolution method significantly impacts MLIP performance.
• The zirconium-oxygen system provides a rigorous testbed for evaluating these methods.

Conclusions:

• CE and DM are superior alternatives to MD for generating high-quality training data for MLIPs.
• Improved training data leads to more accurate and reliable MLIPs.
• Formalizing structure sourcing processes is essential for advancing MLIP development.