Control of fast nuclear reactor for space propulsion using Reinforcement Learning

This project aims to develop and test an efficient regulation model for a fast fission reactor (FFR) applied to a space propulsion model. An FFR's reactivity level, held in check by a set of control drums surrounding the core, fluctuates around a safe subcritical value. These fluctuations have a typical characteristic time on the order of milliseconds, which is too quick for a technician or a feedback loop to operate the drums. This is why we must resort to intelligent systems instead. A well-trained Reinforcement Learning (RL) based Machine Learning (ML) model has the potential to predict possible deviations and regulate the reactivity levels by precisely setting the turning angle of the control drums and operating valves for the propellant flow across the reactor. We are developing a physics-based model of a nuclear thermal reactor for space propulsion, where we plan to incorporate RL algorithms to the control elements. The efficiency of the RL algorithm will be validated in a simulated environment for typical thrust and specific impulses needed through a standard Mars mission. Our project is thus fully aligned with NASA's plans for deep space missions using advanced propulsion technologies.