Prediction of Actinide–Ligand Complex Stability Constants by Machine Learning

Sustainable application of nuclear energy requires efficient sequestration of actinides, which relies on extensive understanding of actinide–ligand interactions to guide rational design of ligands. Currently, the design of novel ligands adopts mainly the time-consuming and labor-intensive trial-and-error strategy and is impeded by the heavy-metal toxicity and radioactivity of actinides. The advancement of machine learning techniques brings new opportunities given a sensible choice of appropriate descriptors. In this study, by using the binding equilibrium constant (log K1) to represent the binding affinity of ligand with metal ion, 14 typical algorithms were used to train machine learning models toward accurate predictions of log K1 between actinide ions and ligands, among which the Gradient Boosting model outperforms the others, and the most relevant 15 out of the 282 descriptors of ligands, metals, and solvents were identified, encompassing key physicochemical properties of ligands, solvents, and metals. The Gradient Boosting model achieved R2 values of 0.98 and 0.93 on the training and test sets, respectively, showing its ability to establish qualitative correlations between the features and log K1 for accurate prediction of log K1 values. The impact of these properties on log K1 values was discussed, and a quantitative correlation was derived using the SISSO model. The model was then applied to eight recently reported ligands for Am3+, Cm3+, and Th4+ outside of the training set, and the predicted values agreed with the experimental ones. This study enriches the understanding of the fundamental properties of actinide–ligand interactions and demonstrates the feasibility of machine-learning-assisted discovery and design of ligands for actinides.