AI Supplies Data For Modeling Safer And More Sustainable Nuclear Power Reactors

Moscow, Jan. 15, 2025

Researchers from Skoltech and the Institute of High Temperature Electrochemistry of UB RAS have developed and tested a model based on machine learning that predicts the properties of molten salts. These compounds are already used in metallurgy and hold promise for resolving the problem of mounting nuclear waste. Their industrially important properties are hard to measure in experiments. This makes models such as the one presented by the team in the Journal of Molecular Liquids crucial for making pure metal manufacture cheaper and nuclear power safer and more sustainable.

Molten salts are a very diverse class of compounds with a large number of physical properties relevant to the industry. Materials scientists are working on fine-tuning the composition and properties of molten salt mixtures to make the production of pure titanium, calcium, aluminum, and certain other metals more effective and to remove an important technological barrier hampering the development of next-generation nuclear reactors.

With much attention paid to solar and wind generation, nuclear power also has a critical role to play in the transition to a carbon-free future. While fusion reactors promise much but remain elusive, there is a nuclear power technology much closer to implementation that could also do a lot for the energy industry. And that technology relies on molten salts with optimized physical and chemical properties.

Molten-salt reactors would be safer, more sustainable, and would produce more power than those in use today. They are not prone to hydrogen explosions, such as the one involved in the Fukushima nuclear disaster, and generally operate at close to atmospheric pressure, whereas most present-day reactors require between 75 and 150 atmospheres, with implications for both safety and operational costs. Unlike most conventional systems, MSRs can be refueled while operating, without the need for a temporary shutdown. MSRs operate at roughly twice the temperature of conventional reactors, boosting power generation efficiency and the opportunities for capturing waste heat.

Among their other benefits, molten-salt reactors could ease the problem of mounting nuclear waste from conventional reactors. They produce highly radioactive minor actinides: neptunium-237, americium-241, etc. While this hazardous waste is hard to dispose of, it would be suitable fuel for a molten-salt reactor.

To tap into the potential of molten salts for both nuclear power engineering and metallurgy, engineers need to know their properties. Materials scientists are hard-pressed to supply that information, because of the sheer quantity of possible combinations of chemical elements and the number of technologically relevant properties. Going over every combination and doing an experiment would be incredibly expensive. Especially given the highly corrosive nature of molten salts and the high temperatures involved.

“Computationally guided search for melts with particular physico-chemical properties might substantially simplify and accelerate the development of next-generation nuclear reactors, since the number of real experiments will be minimized,” says the study’s lead author Nikita Rybin, a research scientist at Skoltech AI’s Laboratory of Artificial Intelligence for Materials Design. “In this study, we presented and tested a methodology that allows one to calculate thermophysical properties of molten salts at finite temperatures. Our findings for the salt known as FLiNaK (contains LiF, NaF, KF) coincide with the available experimental data, prompting us to continue that work with other salt compositions and expand the range of properties. This will eventually make computationally guided developments in next-generation nuclear reactors feasible.”

The solution used by the team to calculate molten salt properties is known as machine-learned interatomic potentials. These are trained on the output of smaller-scale models formulated with quantum mechanical accuracy. If it weren’t for machine learning, the fundamental calculations would have gotten way too demanding computationally by the time the researchers got to the scale large enough for the physical properties to emerge in the model.

The study reported in this story was supported by Russian Science Foundation Grant No. 23-13-00332.

© Scoop Media