Machine-Learning-Guided Ligand Optimization for Americium/Europium Coordination Discrimination

Efficient separation of trivalent minor actinides from lanthanides remains a central challenge in advanced nuclear fuel cycle strategies. Here, we develop a machine-learning-guided ligand design framework to optimize intrinsic Am³⁺/Eu³⁺ coordination discrimination under limited availability of experimental stability-constant data. A graph neural network model is trained via stepwise transfer learning and semi-supervised refinement to predict first-step metal-ligand stability constants (logK₁) for phenanthroline-derived ligands. The predicted differential descriptor (ΔlogK₁) is embedded as a reward to guide constrained molecular generation using a scaffold-preserving generative model. The resulting framework maintains chemical validity and structural diversity while enriching candidate ligands with enhanced predicted intrinsic Am³⁺/Eu³⁺ coordination preference relative to literature-reported references. This work demonstrates that coordination-chemistry-informed machine learning enables systematic exploration of ligand chemical space under data-scarce conditions and provides a practical computational screening route for candidate extractants targeting challenging An/Ln discrimination problems.