In a new study, scientists have shown how deep reinforcement learning can design more efficient nuclear reactors. The success of AI-inspired them in board games.
Nuclear power now provides more carbon-free electricity in the US than solar and wind power combined. This makes it a key player in the fight against climate change. However, the methods for its extraction are imperfect and outdated. The process needs to be optimized so that nuclear power can compete with coal and gas power plants in the market.
Reducing mining costs can be achieved by optimizing fuel rods deep in a nuclear reactor. They trigger reactions, and when ideally positioned, burn less fuel and require less maintenance. After decades of trial and error, nuclear engineers have learned to develop better layouts for expensive fuel rods to extend their life. Now artificial intelligence (AI) will help them.
Researchers at the Massachusetts Institute of Technology (MIT) and Exelon are confident that by turning the design process into a game, the AI system can be trained to generate dozens of optimal rod configurations that can extend each rod’s life by about 5%. This saves about $3 million per year for a typical power plant. An artificial intelligence system can find optimal solutions faster than humans and quickly change designs in a secure simulated environment.
“This technology can be applied to any nuclear reactor in the world,” explains senior study author Korish Shirvan, an assistant professor in the Department of Nuclear Science and Technology at the Massachusetts Institute of Technology. “By improving the economy of nuclear power, which supplies 20% of US electricity, we can help limit the growth of global carbon emissions and attract the best young talent to this important clean energy sector.”
In a typical reactor, the fuel rods are lined up or arrayed at levels of uranium and gadolinium oxide inside, like chess pieces on a board, with reactions that trigger radioactive uranium and rare earth gadolinium, slowing them down. In an ideal arrangement, these competing impulses are balanced to stimulate effective responses. Engineers have tried using traditional algorithms to improve human-designed layouts, but a standard 100-rod assembly can have an astronomical number of options to evaluate.
Researchers wondered if deep reinforcement learning – an artificial intelligence technique that has allowed superhuman proficiency in games like chess and goes – could speed up the verification process. Deep reinforcement learning combines deep neural networks that excel at picking patterns in datasets with reinforcement learning, which links learning to a reward signal like winning a game.
In a new experiment, the researchers trained their agents to place fuel rods according to a set of constraints, earning more points for each coup. Each constraint or rule chosen by researchers reflects decades of expert knowledge based on the laws of physics. The agent can score points, for example, by placing low-uranium rods at the edges of the assembly to slow down reactions there.
“Once you program the rules, the neural networks start to work really well,” said lead author Majdi Radaideh, a postdoc at Shirvan’s lab. – They don’t waste time on random processes. It was fun to watch them learn to play games, as a human does.”
Through Reinforcement Learning, the AI has learned to play ever more complex games and or better than humans. But its capabilities remain useless in the real world. Now researchers have proven that reinforcement learning has potential.
“This study is an exciting example of the use of artificial intelligence technology for the board and video games to help us solve practical problems in the world,” concludes study co-author Joshua Joseph, Research Fellow at MIT Quest for Intelligence.
Exelon is currently testing a beta version of an artificial intelligence system in a virtual environment. According to a company representative, the system may be ready for implementation in a year or two.