Abstract: The implementation of partitioning and transmutation(P&T) strategies in advanced nuclear fuel cycles requires selective separation of minor actinides from lanthanides, for which solvent extraction is a key approach. The separation performance is governed not only by thermodynamic selectivity but also by the extraction process controllability, which is strongly affected by the coordination stoichiometry of the extracted complexes. Under common slope-analysis assumptions, the distribution ratio of a metal ion depends on extractant concentration with an exponent corresponding to the number of extractant molecules in the predominant extracted complex. Accordingly, the coordination stoichiometry difference(Δn) between Am(Ⅲ) and Eu(Ⅲ) for a given extractant provides a physically meaningful descriptor that determines how separation factors amplify with extractant concentration. In this work, |Δn| is used as a sign-invariant indicator of differentiated coordinative behavior and thus of controllability in binary Am/Eu separation. Here we develop an integrated AI-driven molecular design framework that combines property evaluation, molecular generation, and goal-directed optimization to enable de novo discovery of extractants targeting Am(Ⅲ)/Eu(Ⅲ) separation. The framework couples (1) a multimodal graph neural network(GNN) that predicts coordination stoichiometry for metal-ligand-solvent systems, (2) a variational autoencoder(VAE) that encodes discrete molecular structures into a continuous latent chemical space and decodes them into chemically valid and diverse SMILES representations, and (3) an actor-critic reinforcement learning(RL) strategy that performs optimization in the VAE latent space using rewards derived from GNN-predicted |Δn|. To ensure chemically realistic exploration, the reward function incorporates validity constraints implemented with RDKit and structural similarity control based on the Tanimoto coefficient, thereby balancing exploration and exploitation while maintaining optimization stability. Transfer learning is employed to address data scarcity: both the GNN and VAE are pretrained on molecular and coordination-complex structural data derived from the Cambridge Structural Database and subsequently fine-tuned using literature-derived Am(Ⅲ)/Eu(Ⅲ) solvent extraction data. A leakage-control splitting strategy is adopted by grouping samples that share the same extractant-solvent system but differ only in metal identity. The resulting framework enables rapid assessment of Am/Eu coordination stoichiometry differences and directs molecular generation toward regions of chemical space associated with larger |Δn|. Analysis of RL-generated molecules indicates that heteroatom-rich coordination environments and rigid aromatic scaffolds are preferentially enriched among candidates with higher predicted |Δn|, consistent with empirical design principles for An/Ln selectivity. Overall, this study provides an interpretable and systematic AI-based route for selectivity-driven extractant discovery and offers a practical complement to experimental and theoretical approaches for radiochemical separations relevant to advanced nuclear fuel cycles.