AI accelerates search for safer, more durable materials for nuclear reactors

December 23, 2021, 9:30AMNuclear NewsJohn Spizzirri
A cutaway view of a nuclear reactor. Its construction consists of two essential material types: fuel, which comprises the rods and cores that hold the fuel (center vertical bands); and structural, those parts of the reactor that house the fuel materials. (Graphic: Shutterstock/petrov-k)

Researchers from the Department of Energy’s Argonne National Laboratory are developing a “tool kit” based on artificial intelligence that will help better determine the properties of materials used in building a nuclear reactor.

“What we’re trying to do is take the human factor out of the loop in the process of visualizing radiation damage in materials,” said Logan Ward, an assistant computational scientist. “We’re building AI tools that help quantify what we’re actually seeing in that material so we can build predictive models that will tell us what is happening in the reactors much faster.”

Most nuclear reactors rely on two essential material types: nuclear fuel, which comprises the rods and cores that hold the fuel and in which the reactions take place; and structural, which are those parts of the reactor that house the fuel materials.

The reactor’s fission process damages both types of materials over time, while other fission products damage mainly the fuel material. As a result, the integrity of the reactor materials is degraded.

“The challenge with these materials is that you have a lot of phenomena occurring when the nuclear reaction takes place,” said Argonne principal nuclear engineer Abdellatif Yacout. “There are the fission products from fuel materials that cause radiation damage, and there are fission gases that deform the material. We are trying to understand the impact of all of these things together.”

Choosing ideal materials

Different materials react differently to bombardment by neutrons and fission products, with some materials succumbing more quickly to the damage than others. For example, fuel materials deform faster, as they take the brunt of the radiation. It is this susceptibility that the researchers are trying to define faster and more precisely in an effort to choose materials that can withstand this punishment for many decades rather than a few years.

One way of determining a material’s properties during a reactor’s life cycle is to watch it in action over time. It could take years, however, for a material to reach critical radiation damage. In the end, the time and expense involved may turn up a product that is insufficient.

To shorten this development cycle, the researchers are using simulation irradiation tools housed at Argonne’s Intermediate Voltage Electron Microscopy-Tandem Facility (IVEM) and the Argonne Tandem Linac Accelerator System, a DOE Office of Science User Facility. By simulating the effects of radiation damage using an ion accelerator instead of neutrons or fission products, the researchers can create damage to a material in hours or days that is equivalent to years of damage in a nuclear reactor.

The accelerator is connected to a transmission electron microscope (TEM), which has a very high resolution that can magnify objects at the atomic scale. When a material is irradiated with ions, the researchers can witness the damage that occurs in real time and identify the specific types of defects.

“When we can understand the behavior of different defects and how they evolve under irradiation, we can understand a material’s behavior in a nuclear reactor environment,” noted Meimei Li, a principal materials scientist and project investigator. “But like any simulation tool, how do I know the ion radiation data it generates can represent the neutron damage in a reactor?”

For some time, researchers have used multi-scale physics-based models to compare the damage in two different types of radiation conditions. The method is accurate when applied to simple materials such as pure iron, for example, but it doesn’t do well with more complex materials.

Reactor materials are complex. Different alloys, such as uranium alloys, are used for the fuel material, and austenitic stainless steel—a mixture of many elements that resist corrosion—is just one of many types of structural material in the reactor environment.

The team is preparing to use a combination of TEM images and AI techniques to predict neutron damage in a reactor faster and more accurately.

The images, or micrographs, taken by the electron microscope at the IVEM facility show how the microstructure of a material evolves as the material is irradiated.

“We get these images that look like big spotted Dalmatian patterns, and a human has to go through and determine which of these spots are the defects we’re trying to model, measure how big they are, count how many of them there are, and determine what fraction of the image they take up,” explained Ward.

That data tells researchers how much radiation produces a given amount of damage. Combined with an understanding of physics, it provides a glimpse of the different processes and rates occurring in the material, which are then used to model what is actually happening in a reactor. But this still takes a long time, and it has to be done for many different radiation conditions and types of materials.

Where AI comes in

Because a lot of manual work has already gone into identifying defects, these data can be used to train machine learning models—which learn by data input and repetition—to go through these same micrographs and quickly pick out features that identify specific types of defects. This information then helps the team analyze experimental data from ion radiation in the IVEM.

Video from the IVEM will produce many different images of damage evolving in the material. Another pass through the machine learning models will provide information on the size, shape, and location of each defect, as well as the ability to track them over time.

“Once we have all of that data, we use another machine learning technique to fit physics models,” Ward said. “This gives us reaction rates, like how quickly the defects move and how quickly they dissolve or reappear. And that lets us build models that we can use to forecast the performance of a material. It will tell us how soon we’ll have to replace something.”

This updated toolbox, a combination of simulation, data, and AI, will provide information on both existing materials and those coming down the pipeline, helping save an enormous amount of time and expense in R&D. More important, it will ensure that the materials are stable and strong enough to withstand the intense radiation fields they receive, making for safer, more economical, longer-lasting nuclear reactors.

The research was funded through Argonne’s Laboratory Directed Research and Development program.

Click here for a video that shows images captured by Argonne’s transmission electron microscope.

John Spizzirri is a senior science writer in the Communications and Public Affairs Division at Argonne National Laboratory.

Related Articles