Automated Machine Learning Pipeline: Large Language Models-Assisted Automated Data set Generation for Training...

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Machine learning interatomic potentials (MLIPs) offer near-quantum accuracy for molecular simulations but are challenging to develop.
• Current MLIP development requires extensive data generation, structure preprocessing, and model training/validation.

Purpose of the Study:

• To introduce an Automated Machine Learning Pipeline (AMLP) that integrates the entire MLIP workflow.
• To leverage large-language-model agents for automating code selection, input preparation, and output conversion.

Main Methods:

• The AMLP pipeline utilizes large-language-model agents and the MACE architecture.
• An analysis suite (AMLP-Analysis) based on ASE supports various molecular simulations.
• The pipeline was validated on acridine polymorphs.

Main Results:

• Fine-tuning a foundation model achieved mean absolute errors of 1.7 meV/atom (energies) and 7.0 meV/Å (forces).
• The resulting MLIP accurately reproduced DFT geometries (sub-Å accuracy).
• The MLIP demonstrated stability in microcanonical and canonical ensemble molecular dynamics simulations.

Conclusions:

• The AMLP significantly simplifies and automates the development of reliable MLIPs.
• The developed MLIPs provide accurate and stable simulations, extending the capabilities of computational methods.