MIT research targets autonomous nuclear plant control without AI
The economics of nuclear power depend partly on whether operations can be automated sufficiently to support small plants in locations where large staffing models are not viable. A doctoral project at MIT is working on exactly this problem, and its chosen technical path reveals as much about regulatory reality as it does about engineering ambition.
Lauren Fortier, a second-year PhD student in MIT's Department of Nuclear Science and Engineering, spent several years operating nuclear reactors on a U.S. aircraft carrier before moving into research. Her transition from naval operations to academia was motivated by a specific observation: many plant procedures were heavily manual, and she saw potential to reduce that burden. The Navy supported her move to MIT for a master's degree, which she completed in 2025, and she is now continuing toward a doctorate under a supervisory control system framework.
The research question she is pursuing is how to build a supervisory control system for autonomous nuclear plant operations. Legacy plants run at full capacity with large human crews, so intensive manual operation is economically justified. Distributed microreactors in remote areas cannot support that staffing model. Fortier's goal is an integrated supervisory system that can manage plant objectives without relying on pre-scripted procedures for every contingency.
Her approach uses finite state automata rather than machine learning. The choice is deliberate. Fortier states that the team is not using a data-driven statistical approach because the tools to validate such systems for nuclear operations do not yet exist. Finite state automata offer a discrete event structure where each transition is explicit and traceable: if a condition occurs, a defined response follows. The framework she is developing incorporates objective-oriented operations, meaning the control system determines the sequence needed to reach a stated goal rather than executing a predetermined checklist. This is a meaningful distinction because it shifts the system from procedure-following to reasoning about plant state, while remaining within an automation paradigm that nuclear regulators can audit.
The human-machine interface question runs alongside the control architecture. Fortier collaborated with Katya Le Blanc, a senior human factors scientist at Idaho National Laboratory, and worked with INL's Human System Simulation Laboratory to address how operators take over from automated systems when conditions require it. She also interned at Westinghouse in 2025 to test autonomous operation concepts against commercial reactor design assumptions. Her academic advisors include Sacit Cetiner (joint appointment between MIT NSE and INL), Anuranda Annaswamy (Active-Adaptive Control Laboratory), and Curtis Smith (KEPCO Professor of Practice at MIT NSE and former director of INL's Nuclear Safety and Regulatory Research Division).
Fortier frames the deployment pathway as a gradual progression. Automated procedures that guide operators through familiar steps are intended to build trust before introducing higher levels of autonomy. This is a pragmatic stance toward regulatory acceptance, and it aligns with the finite state automata choice: a system whose every decision path is auditable is easier to certify than one whose behavior emerges from training data.
Fortier's work received recognition from the Department of Energy's Nuclear Energy University Program, winning the 2025 Innovations in Nuclear Energy Research and Development Student Competition. The practical goal is to apply the supervisory control framework to next-generation equipment as a step toward commercial microreactor deployment.
The project illustrates a specific constraint in nuclear automation that does not apply to most other domains. The validation gap for machine learning in safety-critical infrastructure is real, and the finite state automata approach is a direct engineering response to it. What remains unaddressed in the source is how the framework scales to plants with different thermal hydraulic characteristics or operational profiles, and whether the objective-oriented architecture can maintain its guarantees when plant objectives become complex or interdependent.