Deep-Learning-Guided High-Throughput Evaluation of Ligands for Selective Sr/Cs Coordination

Abstract

From a coordination chemistry perspective, we aimed to advance the knowledge of Sr/Cs separation in the scheme of spent nuclear fuel reprocessing. Based on data mining of crystal structures and deep learning architecture, we summarized and analyzed coordination chemistry properties of Sr/Cs from complex structures(ca. 3.3×104 samples) of 8 alkaline and alkaline earth elements, especially focusing on coordination bond lengths as a representative figure of merit. Applying a Bayesian optimization approach, we were able to establish a high-performing transformer model which could predict the(differential) coordinative affinities toward Sr/Cs of ligand molecules, with exceptional accuracy. As a proof-of-concept, we systematically analyzed ca. 9.1×103 ligand molecules in terms of potential coordinative selectivity toward Sr/Cs and ranked them. In addition, we also determined different contribution of various functional groups for future molecular design of ligands with selectivity. The present study presented fundamental knowledge for coordination chemistry information in the context of radiochemistry and spent nuclear fuel reprocessing, provided guidance and reference for subsequent experiments regarding Sr/Cs separation.

Keywords

• deep learning
• bayesian optimization
• spent nuclear fuel reprocessing
• sr/cs separation