Technetium-99 (Tc-99), a long-lived radioactive fission product of uranium and plutonium, accounts for nearly 6% of all fission products. Since technetium dissolves and migrates with groundwater in nuclear waste disposal sites, efficient methods are necessary to manage its behavior and mitigate environmental risks.
One possible approach is to securely store technetium-99 in a carbon matrix, where it forms stable carbides. These carbides could then be used as neutron beam targets to gradually convert technetium-99 into ruthenium-100, a stable isotope with potential applications in microelectronics and as a catalyst.
In a recent study, researchers from Skoltech, the AIRI Institute, Sberbank, the Mendeleev University of Chemical Technology, and the Frumkin Institute of Physical Chemistry and Electrochemistry of RAS developed a machine learning model that predicts the thermodynamic properties of technetium and carbon atomic configurations. The team then used the model’s predictions to construct a phase diagram of the Tc-C system in “temperature-carbon content” coordinates, identifying Tc phases that are thermodynamically favorable under specific conditions.
“From a practical standpoint, this diagram serves as a stability map of technetium carbides for experimenters and technologists. It helps select regimes in which technetium is firmly secured in the carbon matrix and determine the operating conditions for transmuting it into stable ruthenium,” Skoltech Materials Science PhD student Radion Zaripov commented.
Radioactivity poses a significant challenge for laboratory experiments involving Tc-99, resulting in limited data on solid technetium compounds. The team had to consider and determine the properties of hundreds of thousands of possible atomic configurations − a number far exceeding the few hundred unstable structures found in one of the largest open databases of calculated material properties for the Tc–C system. Calculating them all using standard quantum mechanical methods, such as density functional theory (DFT), would have taken years of processing time.
This is where AIRI’s hybrid approach proved useful. First, the researchers performed precise calculations on a small, representative subset of configurations (less than 0.2% of the total), and then trained a neural network model on this data. The trained algorithm could quickly scan the entire set with an error of a few millielectronvolts per atom. This helped reliably identify the most stable structures, including rare configurations with a one-in-ten-thousand or one-in-a-million probability of random discovery.
“We have previously used similar approaches to study functional materials and predict new compositions. In this study, we demonstrated that removing randomness from machine learning-based computational approaches accelerates property prediction and helps identify rare structures that could be overlooked in a random search,” Roman Eremin, the head of the AIRI New Materials Design group, said.
Credit: AIRI Institute
