Prediction of Actinide-Ligand Complex Stability Constants by Machine Learning.

Area of Science:

• Nuclear chemistry and materials science.
• Computational chemistry and machine learning.

Background:

• Efficient sequestration of actinides is crucial for sustainable nuclear energy.
• Current ligand design relies on slow, labor-intensive trial-and-error methods, hindered by actinide toxicity and radioactivity.
• Machine learning offers a promising alternative for accelerating ligand discovery.

Purpose of the Study:

• To develop accurate machine learning models for predicting actinide-ligand binding affinities (log K1).
• To identify key physicochemical descriptors influencing these interactions.
• To demonstrate the utility of machine learning in designing novel ligands for actinides.

Main Methods:

• Trained 14 machine learning algorithms using binding equilibrium constants (log K1) as the target property.
• Identified the 15 most relevant descriptors from a set of 282, covering ligands, metals, and solvents.
• Utilized the Gradient Boosting model for prediction and the SISSO model for quantitative correlation.

Main Results:

• The Gradient Boosting model achieved high accuracy, with R2 values of 0.98 on the training set and 0.93 on the test set.
• Identified key physicochemical properties influencing actinide-ligand interactions.
• Successfully predicted binding affinities for new ligands with experimental agreement.

Conclusions:

• Machine learning models, particularly Gradient Boosting, can accurately predict actinide-ligand binding affinities.
• This study provides fundamental insights into actinide-ligand interactions and their governing properties.
• Machine learning-assisted design accelerates the discovery of effective ligands for actinide sequestration.