The role of machine learning in computational chemistry is growing rapidly, promising expedited tasks while improving prediction accuracy. However, the interface of newly developed machine-learned interatomic potentials (MLIPs) with existing computational chemistry workflows based on quantum mechanics (QM) remains a challenge. Here, we introduce ChemRefine as a modular automated computational chemistry platform that enables researchers to seamlessly integrate the next generation of MLIPs into both complex and high-throughput computational workflows. We road-test ChemRefine on multiple real-world computational tasks, including host-guest binding for transmembrane anion transporter design, redox and photophysical property predictions for enzyme cofactors and photocatalysts, conformational sampling of organometallic complexes, and transition state finding for "click" reaction design. MLIPs can be automatically trained, fine-tuned, and deployed during workflows, accelerating simulations while maintaining prediction accuracy. In all cases, multiple computational tasks are condensed into a single reproducible workflow. To simplify the use of ChemRefine, we also introduce a custom large language model, ChemRefineGPT, that generates all required input and configuration files based on a natural language description of the desired workflow.
Migliaro et al. (Tue,) studied this question.