Crash testing machine learning force fields for molecules, materials, and interfaces: molecular dynamics in the TEA...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning force fields (MLFFs) are increasingly used in molecular simulations.
• Rigorous evaluation of MLFF performance is crucial for reliable scientific discovery.
• The TEA Challenge 2023 provides a platform for benchmarking MLFFs.

Purpose of the Study:

• To evaluate the performance of leading MLFFs (MACE, SO3krates, sGDML, SOAP/GAP, FCHL19*) in modeling molecules, interfaces, and materials.
• To compare simulation results from different MLFFs under identical conditions.
• To assess the impact of MLFF architecture versus training data quality on simulation accuracy.

Main Methods:

• Conducting molecular dynamics (MD) simulations using various MLFFs.
• Comparing MLFF results against density-functional theory (DFT) or experimental data where available.
• Performing comparative analysis across different MLFF architectures in the absence of DFT benchmarks.

Main Results:

• MLFF performance is largely dependent on the quality and representativeness of training datasets rather than specific architectures.
• Simulations show weak dependency on MLFF architecture when the problem is within the model's scope.
• Long-range noncovalent interactions pose a significant challenge for all current MLFFs.

Conclusions:

• Practitioners can choose MLFFs based on their specific needs, but dataset quality is paramount.
• Careful consideration is needed for systems dominated by long-range noncovalent interactions, like molecule-surface interfaces.
• The findings reflect the state of MLFF development as of October 2023.