Electronic-Structure Informatics Guided Identification of Actinide/Lanthanide Selectivity Factors

We report an interpretable electronic-structure informatics strategy for identifying ligand features that govern actinide/lanthanide (An/Ln) selectivity at the metal-ligand interaction level, a domain where limited experimental data and open-shell electronic complexity have hindered broad adoption of machine learning (ML) approaches. A systematically benchmarked DFT database of terpyridine-type ligands coordinated to Am³⁺, Cm³⁺, Eu³⁺, and Gd³⁺ enabled the construction of compact electronic descriptors that predict complex-formation free energies and An/Ln selectivity differences via Gaussian process regression. The resulting model achieves quantitative accuracy (mean absolute error; MAE = 17.5 kJ mol^-1 for Δ<i>G</i>_bind; 2.6 kJ mol^-1 for ΔΔ<i>G</i>_An-Ln) while preserving chemical interpretability. Shapley Additive exPlanations (SHAP) analysis reveals that metal ionization potentials and inner-site charge patterns cooperatively govern selectivity, consistent with established coordination principles and hard-soft matching trends. These insights yield actionable design rules: electron-deficient central azines paired with moderately π-accepting terminal donors preferentially enhance An binding relative to Ln. By explicitly positioning the model as a controlled aqueous-phase thermodynamic screening layer, this work provides a reproducible and mechanistically grounded ML framework for predicting intrinsic An/Ln selectivity factors. The resulting electronic design variables are consistent with reported experimental trends and provide a practical basis for future extraction-oriented ligand design.