Automated control: The study examined a new ML approach for modeling the adjustment of nuclear microreactor power output to meet grid demand, based on the design of HolosGen’s Holos-Quad microreactor. According to the researchers, their ML approach, called multiagent reinforcement learning (RL), allows for more efficient training and reduced training time than previous approaches, helping researchers model reactors faster.
Holos-Quad microreactor: The high-temperature, gas-cooled micoreactor is designed for scalable, self-contained power generation. Its architecture was reportedly inspired by closed-loop turbo jet engines, replacing conventional combustion chambers with sealed nuclear fuel cartridges, which integrate fuel, moderation, heat exchange, and power conversion within individual pressure vessels.
Its compact design can fit inside a 40-foot International Organization for Standardization (ISO) container.
Load-following simulation: The research team focused on simulating load-following, which is the increase or decrease in electricity output to match the demand of the grid. The Holos-Quad system is designed to adjust power through the position of eight control drums that center around the reactor’s central core. One side of the control drum’s circumference is lined with a neutron-absorbing material that, when the drum is rotated inward, absorb neutrons from the core, causing the neutron population and the power to decrease. Meanwhile, rotating the cores outward keeps more neutrons in the core, increasing power output.
In the investigators’ multiagent RL approach, the eight control drums of the Holos-Quad reactor core were modeled as eight independent agents, with specific drums being controlled independently, while information about the core as a whole was also obtained.
Comparisons: Tests comparing the results of the team’s multiagent RL model approach with a single-agent approach (in which a single agent controls all eight drums) and the industry standard proportional-integral-derivative (PID) approach (with a feedback-based control loop) showed benefits for the RL model.
Main conclusions: The researchers noted that their ML model will need extensive validation in more complex and realistic conditions before it can be commercially applied by the nuclear power industry. Nevertheless, their findings “establish a more efficient path forward for reinforcement learning in autonomous nuclear microreactors.”
