Agentic workflows for exploring the chemical space of lanthanide/actinide extractants

J. Schrier

Lanthanides and actinides have numerous applications in fields as diverse as nuclear energy, medicine, and magnets. Solvent extraction is one of the important and widely used separation methods, but the design of efficient and selective extractant molecules is an open challenge. In this talk, I will describe the use of a large language model (LLM) agentic workflow to explore the chemical space of extractant molecules. The LLM agent uses a human scientist’s written design suggestion to iteratively generate and optimized new molecules. Along the way, the agent tests its creations queries to a database of experimental data to confirm novelty and calling a several supervised cheminformatics models to predict molecular properties such as solubility, predicted binding, and synthetic accessibility. The results are then validated by using Architector to perform automated conformational searching of the metal-ligand-solvent binding possibilities, and density functional theory (DFT) calculations to predict binding energy. We demonstrate the capabilities by developing novel extractants for Am(III)/Eu(III) separation which exceed the current state of the art. This methodology can be applied to other f-block element separations, using our website at https://safe.lanl.gov/fsepamolgen/doc