Bond-Level Prediction of U-Donor Bond Lengths in Uranyl Coordination Complexes Using Machine Learning for Ligand Recognition.

U-donor bond lengths in uranyl coordination complexes directly reflect metal-ligand interaction strength and provide important structural parameters for evaluating ligand recognition, selective complexation, and extraction/adsorption system design. To enable bond-level prediction, this study developed a DFT-derived machine learning framework for predicting U-donor bond lengths. A data set containing 15,177 coordination structures and 26,798 U-donor bond length data points was constructed from 3,330 unique ligands under a unified computational protocol. Five categories of models, including linear models, support-vector-based models, ensemble tree models, descriptor-based neural networks, and graph neural networks, were evaluated to compare representation strategies. Among descriptor-based models, XGBoost achieved the best performance, with R2, RMSE, and MAE values of 0.793, 0.057 Å, and 0.039 Å, respectively. Among graph neural networks, DimeNet++ performed best, with R2, RMSE, and MAE values of 0.821, 0.051 Å, and 0.036 Å. Functional-group-based evaluation and SHAP analysis indicated that donor atom type, neighboring atomic environment, local electronegativity difference, and local geometric features are key factors controlling U-donor bond length variations. External validation using unseen functional groups and uranyl structures demonstrated model generalization beyond the training space. This work provides a bond-level data set and a data-driven strategy for rapid U-donor bond length prediction and ligand screening in actinide coordination systems.