BYU professor says AI can shave years and millions off nuclear power development process

Nuclear power is an essential piece to the puzzle of meeting energy demands, but the design and licensing processes for modern nuclear reactors are lengthy and expensive.

Brigham Young University chemical engineering professor Matthew Memmott says he has discovered an unlikely tool to help shave critical years off the process: Artificial intelligence.

While this may seem dystopian, Memmott's strategy doesn't involve giving AI nuclear codes. Instead, he's harnessing the powers of AI to speed up the nuclear process and get more nuclear power online.

"Our demand for electricity is going to skyrocket in years to come and we need to figure out how to produce additional power quickly," Memmott said. "The only baseload power we can make in the gigawatt quantities needed that is completely emissions free is nuclear power. Being able to reduce the time and cost to produce and license nuclear reactors will make that power cheaper and a more viable option for environmentally friendly power to meet the future demand."

Usually, the time frame and cost to license a new nuclear reactor design in the U.S. is around 20 years and $1 billion. And that isn't even taking into account the time and money to build a nuclear reactor, which is another five years with a price tag anywhere from $5 billion to $30 billion.

But why is designing and building these reactors so complex and time consuming?

Memmott explained that doing so requires multi-scale efforts. Engineers deal with elements from neutrons on the quantum scale all the way up to coolant flow and heat transfer on the macro scale. He added that there are multiple layers of physics that are "tightly coupled" in that process: The movement of neutrons is tightly coupled to the heat transfer which is tightly coupled to materials which is tightly coupled to the corrosion which is coupled to the coolant flow.

The problems that arise when designing reactors are so massive and involve so much data it can take months for entire teams of experts to resolve the issues.

"When I was at Westinghouse it took the team of neutron guys six months just to run one of their complete-core multi-physics models. And if they made a mistake two months in, then they just wasted two months of the valuable computational time and they would have to start over," Memmott said.

But Memmott has found that AI can reduce the time burden, resulting in more power production to not only meet rising energy demands but also keep power costs lower for general consumers like homeowners and renters.

His research is predicated on replacing a portion of the required thermal hydraulic and neutronics simulation with a trained machine learning model to predict temperature profiles based on geometric reactor parameters that aren't consistent and then optimizing those parameters.

Doing so creates an optimal nuclear reactor design more efficiently than can be done through traditional design methods.

To come to this conclusion, Memmott and other BYU professors built a dozen machine learning algorithms to examine their ability to process the simulated data needed to design a nuclear reactor. They narrowed it down to the top three algorithms and refined the parameters until they discovered one that worked well enough to handle a preliminary data set as a proof of concept.

It worked and the resulting papers demonstrated that their refined model can optimize the design elements of a nuclear reactor far faster than the traditional method for doing so. So much faster that Memmott's AI algorithm took only two days to develop an optimal shield design for a nuclear reactor — a process that usually takes around six months to complete.

“When you look at nuclear reactor design, you have this huge design space of possibilities — it’s as if you have people combing through this mile-wide area looking for the right reactor design,” Memmott said. “Now, AI can help those people focus on that little quarter-sized sweet spot of design which will drastically reduce the search time. Of course, humans still ultimately make the final design decisions and carry out all the safety assessments, but it saves a significant amount of time at the front end.”