MSU Trains Neural Network to Predict the Stability of Metal Coordination Complexes
Researchers at the Artificial Intelligence Center of Lomonosov Moscow State University (MSU), working with colleagues from the university's Institute of Artificial Intelligence, have combined experimental data on the complexation of various metals with organic ligands in a single computational model. This approach makes it possible to predict the stability of complexes involving metals for which no data were included in the training set.
In October 2026, a research team at Lomonosov Moscow State University unveiled a new tool: a graph convolutional neural network designed to predict the stability of coordination complexes formed by rare-earth elements and trivalent actinides.
The Scientific and Methodological Basis
The research addresses a particular challenge in working with rare-earth and radioactive metals: laboratory experiments involving these elements can be expensive and subject to stringent radiation-safety requirements. A key task is to identify organic molecules, known as ligands, that can selectively bind target ions in mixtures containing multiple components. The traditional approach of testing candidate compounds empirically requires extensive and costly experimental work.
The MSU team compiled and organized data from scientific publications and materials produced by the US National Institute of Standards and Technology (NIST), then used these data to train a graph neural network. A key feature of the model is its ability to predict the stability constants of metal complexes even when the metals in question were not represented in the training data. Rather than simply fitting known examples, the model identifies underlying patterns in chemical affinities. The developers have made the source code publicly available and provided a dedicated user interface for interacting with the model.
Strategic Importance for the Nuclear and Chemical Industries
The development could help strengthen Russia's technological sovereignty. Rare-earth elements are critical components in the electronics industry and the manufacture of permanent magnets, while actinides are fundamental to nuclear power and technologies for reprocessing spent nuclear fuel.
Using the neural network during the computational modeling stage makes it possible to screen promising ligands in advance, rule out options that are unlikely to work and reduce the amount of subsequent laboratory testing. For Russia's nuclear and chemical industries, this could significantly reduce the resources needed for research and development, while shifting the process from empirical trial and error toward the targeted, computational design of chemical compounds.
A Look Back at AI Methods in Chemistry
The MSU development builds on a sustained program of research. Looking back at earlier work reveals a steady progression in the development of Russian AI methods for chemistry.
In 2022, MSU researchers demonstrated that transfer-learning methods could be used to predict the physicochemical and biochemical properties of substances from a limited number of experimental examples. In 2024, a research team at the N. D. Zelinsky Institute of Organic Chemistry of the Russian Academy of Sciences developed neural network models capable of recognizing complex molecular structures from electron microscopy and optical microscopy data. In 2025, researchers developed an AI method for finding stable molecular geometries by combining quantum chemical calculations with machine learning. In the first half of 2026, MSU researchers introduced neural network architectures that improved the accuracy of molecular property predictions while also making the models' decisions more interpretable.
The latest development represents a logical next step in this research: moving beyond individual problems toward a more general-purpose tool for scientific prediction.
Model Limitations and Prospects for Adoption
It is important to emphasize that the new model is a research tool, not a ready-to-deploy industrial solution. It still needs to undergo rigorous experimental validation using independent datasets. Nevertheless, the work points to the development of a sustained research capability in Russia at the intersection of information technology, chemistry and materials science. AI does not replace experimental researchers, but it can substantially expand their analytical capabilities by enabling highly predictive modeling at the molecular level.
