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

Area of Science:

• Nuclear Chemistry
• Computational Chemistry
• Materials Science

Background:

• Efficient separation of trivalent minor actinides from lanthanides is critical for advanced nuclear fuel cycles.
• Existing methods face challenges due to the similar chemical properties of actinides and lanthanides.
• Data scarcity for experimental stability constants hinders rational ligand design.

Purpose of the Study:

• To develop a machine-learning-guided framework for designing ligands with enhanced Am3+/Eu3+ coordination discrimination.
• To optimize ligand design under conditions of limited experimental stability constant data.
• To provide a computational screening route for novel extractants in actinide/lanthanide separation.

Main Methods:

• A graph neural network model was trained using stepwise transfer learning and semi-supervised refinement.
• The model predicted first-step metal-ligand stability constants (logK1) for phenanthroline-derived ligands.
• A scaffold-preserving generative model was guided by predicted differential stability constants (ΔlogK1) as a reward signal.

Main Results:

• The framework generated chemically valid and structurally diverse candidate ligands.
• Enriched candidate ligands showed enhanced predicted intrinsic Am3+/Eu3+ coordination preference compared to literature references.
• The approach demonstrated effective ligand chemical space exploration under data-scarce conditions.

Conclusions:

• Coordination-chemistry-informed machine learning enables systematic ligand design for challenging separation problems.
• The developed framework offers a practical computational screening method for nuclear fuel reprocessing.
• This approach facilitates the discovery of selective extractants for actinide/lanthanide separation.