An international team of researchers led by University of Wisconsin-Madison materials engineers has developed a machine learning-based tool called “SuperSalt” that accurately simulates and predicts the properties of molten salt systems. The tool will help researchers tailor molten salts for emerging energy storage applications and the harsh environments of next-generation nuclear reactors.
The research, led by UW-Madison materials science and engineering postdoctoral scholars Chen Shen and Siamak Attarian, appeared August 7, 2025, in the journal Nature Communications.
“Molten salt is very important for clean energy and nuclear reactors, but it’s very hard for experimentalists to get a deep understanding of its physical properties,” says Shen. “So we used machine learning to generate a semi-universal potential to understand multi-component salt systems.”
Chen Shen
Molten salt is just what it sounds like. Chloride, fluoride, and nitride-based salts, including sodium chloride or table salt, melt at around 1,000 degrees Fahrenheit or even hotter. In their molten state, these salts are particularly good at both transferring and storing heat. For instance, in concentrated solar power plants, mirrors direct sunlight to a receiver, which produces enough steam to run an electricity-generating turbine. Some of that heat can also be used to melt salt to store that energy for use when the sun isn’t shining.
In next-generation nuclear applications, fission-producing fuels, like uranium, are dissolved in molten salt, which eliminates the potential for meltdowns and provides other operating advantages. Molten salts are also important in producing and heat-treating certain metals and alloys.
But studying and developing new types of molten salt is difficult; the salts are often corrosive, sometimes radioactive, and, of course, incredibly hot. That means only very specialized labs and highly trained personnel are able to experimentally research them. Even then, exploring these materials is a long and difficult process.
In recent years, as computational power has increased, materials researchers have developed various ways to simulate molten salts, including highly accurate tools called force field simulations and machine learning interatomic potentials. However, a new individual “potential” (which maps atomic structures and potential energies) must be developed to study each variety of salt. Considering that salts can include half a dozen or more elements mixed in almost infinite ratios, creating bespoke potentials for each mixture is also inefficient.
The UW-Madison team, however, saw an opportunity to use emerging tools to conduct simulations of liquid chloride-based salts at scale. Over the course of many months, Chen and Attarian ran close to 80,000 quantum mechanics simulations and adapted physics-guided models and numerical methods to develop a set of constraints that apply to 11 chloride-based salts like sodium chloride, lithium chloride, and potassium chloride.
Siamak Attarian
The result is SuperSalt. The simulation tool uses a single machine learning potential that can investigate any combination of these salts with near quantum-mechanical accuracy, but 1,000 to 10,000 times faster than fully quantum simulations.
“This is really game-changing; with SuperSalt, you have this large compositional space and can explore the different mixtures of salts, different chemistries and really provide an understanding of the thermophysical properties of these salts,” says Izabela Szlufarska, a professor of materials science and engineering at UW-Madison who, with colleague Dane Morgan, collaborated on the research. “This really enables the design of salts because with quick calculations you can predict their properties.”
Besides allowing researchers to understand the properties of specific known salts, the tool will also allow them to search for salts with specific desirable properties. For instance, researchers may need a salt with a low melting temperature and high specific heat, which is good for thermal transfer. Or they may search for combinations that include inexpensive elements that also don’t activate under radiation. SuperSalt will guide them toward combinations that meet those criteria.
The team says it’s likely that over the next decade most of the data on salts will come from SuperSalt and similar simulation tools—rather than experiments. “SuperSalt and similar efforts represent an orders of magnitude speed-up in the ability to get basic data,” says Morgan. “This doesn’t build you a nuclear reactor or other salt-connected technology orders of magnitude times as fast, but it does remove a significant bottleneck in key materials development that could save millions and millions of dollars and years and years of work. So this could be very transformative for multiple fields.”
Other authors include Yixuan Zhang and Hongbin Zhang of the Technical University of Darmstadt, Germany; and Mark Asta of the University of California, Berkeley.
The authors acknowledge support from the Department of Energy (DOE) Office of Nuclear Energy’s (NE) Nuclear Energy University Programs (NEUP) under award # 21-24582; the Bridges-2 cluster at Pittsburgh Supercomputing Center (PSC) and Stampede3 cluster at Texas Advanced Computing Center (TACC) through allocations MAT240071 and MAT240075 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, by the National Science Foundation grants #2138259,#2138286, #2138307, #2137603, and #2138296; and Computational resources provided by the Center for High Throughput Computing (CHTC) at the University of Wisconsin–Madison.