Surrogate-Guided Latent-Space Reinforcement Learning for Generative Ligand Design in Am(III)/Eu(III) Separation

Am(III)/Eu(III) separation is a stringent benchmark for actinide/lanthanide partitioning because the two trivalent ions exhibit closely related size and solution chemistry, making selectivity highly dependent on subtle ligand-controlled coordination differences. Herein, we report a surrogate-guided latent-space reinforcement learning workflow for generative ligand design under sparse and heterogeneous extraction-data conditions. A junction-tree variational autoencoder was pretrained on ligand structures extracted from CSD-derived coordination compounds and fine-tuned on literature-reported Am/Eu extractants to construct a chemically valid ligand latent space. A condition-aware XGBoost surrogate model predicted ln(SF) from ligand, solvent, and experimental-condition features and supplied the reward signal for proximal policy optimization (PPO). The surrogate retained useful ranking ability, including under a strict ligand-level split, while PPO enriched multiple regions associated with high surrogate-predicted performance. The generated candidates recapitulated established soft-donor and N-heterocyclic design patterns while suggesting new structural variants in the learned ligand space. This workflow demonstrates how surrogate-guided generative optimization can translate limited Am/Eu extraction data into chemically meaningful ligand-design hypotheses.