ChemRefine: An Open-Source Automated and Interoperable Platform for Machine Learning and Quantum Chemistry...

Area of Science:

• Computational Chemistry
• Machine Learning
• Scientific Software Development

Background:

• Machine learning is increasingly used in computational chemistry to speed up tasks and improve predictions.
• Integrating new machine-learned interatomic potentials (MLIPs) into existing quantum mechanics (QM) workflows presents a significant challenge.

Purpose of the Study:

• Introduce ChemRefine, a modular platform for automating computational chemistry workflows.
• Enable seamless integration of advanced MLIPs into complex and high-throughput computational tasks.
• Simplify workflow creation using natural language processing.

Main Methods:

• Developed ChemRefine, a modular automated computational chemistry platform.
• Integrated MLIPs for automated training, fine-tuning, and deployment within workflows.
• Utilized ChemRefineGPT, a custom large language model, for generating input and configuration files from natural language descriptions.

Main Results:

• Successfully road-tested ChemRefine on diverse tasks: host-guest binding, property prediction, conformational sampling, and transition state finding.
• Demonstrated accelerated simulations with maintained prediction accuracy using MLIPs.
• Condensed multiple computational tasks into single, reproducible workflows.

Conclusions:

• ChemRefine provides a robust solution for integrating next-generation MLIPs into computational chemistry.
• The platform enhances efficiency and reproducibility for various scientific applications.
• ChemRefineGPT significantly lowers the barrier to entry for utilizing complex computational workflows.