Russian team develops machine learning model to predict stability of rare earth and actinide complexes

Researchers from the Interdisciplinary Laboratory for Intelligent Chemical Design at the Department of Chemistry of Lomonosov Moscow State University, in collaboration with colleagues from the Department of Mechanics and Mathematics, have developed a new machine learning model for evaluating the stability of complexes of rare earth elements and trivalent actinides. The results have been published in the Journal of Chemical Physics.

The research team stated that machine learning can reduce time and material costs in chemical research, which is particularly important for rare, expensive, or hazardous lanthanides and actinides. Lanthanides, along with scandium and yttrium, are commonly classified as rare earth elements and are widely used in permanent magnets, batteries, electronic components, lasers, and other applications. Since rare earth elements are mostly present in trace amounts in mineral raw materials in nature and have similar chemical properties, their separation and purification processes are complex and costly.

Actinides are radioactive and have important applications in the nuclear industry. Improving the separation efficiency of actinides from spent nuclear fuel is considered conducive to advancing the closed nuclear fuel cycle and reducing the generation of radioactive waste. Currently, the selective binding of target elements using specially designed organic ligands is one of the important research directions for separating rare earth elements and actinides.

The researchers noted that traditional experimental screening of ligands faces considerable difficulties due to the similar chemical properties of f-block elements and the experimental limitations imposed by the radioactivity of actinides. In this study, the team established a new database of complexing agents for f-block elements using the NIST46 database and relevant literature data, and trained a graph convolutional neural network on this basis to predict complex stability constants.

Unlike previous models that relied primarily on ligand structure, the new model integrates information on multiple metal complexes into a single neural network architecture, thereby expanding the range of elements covered by rapid machine learning calculations of stability constants. The research team stated that one feature of the model is that even if a certain metal element is not included in the training set, it can still be used for relevant predictions. The project source code has been made publicly available.

The researchers stated that the new model performs well in predicting the properties of actinides and can support the search for more efficient and selective ligands. The current research focuses on complexes with a 1:1 stoichiometric ratio, and the team plans to continue exploring other types of complexes in the future and collaborate with researchers in the synthesis field to advance the application of the relevant models in areas such as spent nuclear fuel reprocessing.