Bridging AI and nuclear power for enhanced reactor safety

Nuclear reactors generate reliable, low-carbon electricity by using heat from nuclear fission to turn turbines. These steady energy producers are a crucial component of clean power generation. Nuclear engineers are responsible not only for understanding reactor dynamics, ensuring proper maintenance and recognizing unusual behavior, but also for determining appropriate corrective actions when necessary.   

The Advanced Reactor Operation and Monitoring Assistant using Generative Pre-trained Transformer (AROMA-GPT) is an innovative tool integrated into a digital twin framework that could play a crucial role in enhancing advanced reactor monitoring and advisory control. 

Developed by graduate research assistant Zavier Ndum Ndum under the supervision of Dr. Yang Liu, assistant professor in Texas A&M University’s nuclear engineering department, the tool leverages human-in-the-loop generative artificial intelligence to support nuclear engineers in ensuring safe and efficient reactor operations. Their research was recently published in Progress in Nuclear Energy.  

“This research is impactful because it points to a practical and responsible way to use artificial intelligence in nuclear engineering,” Ndum said. “If we can build trustworthy AI systems that help engineers automate workflows, retrieve the right technical knowledge, monitor digital twins and support operator training, then we can reduce friction across the entire research-and-development pipeline. That could help accelerate design studies, safety analysis, workforce training and, eventually, deployment.”   

This new technology could help nuclear engineers and operators understand what a reactor is doing and suggest informed steps in real time. The key to this development is that AI is not acting alone, nor does it replace the human operator. Instead, it is AI working within a human-in-the-loop framework, grounded in reactor physics, supported by domain knowledge and connected to specialized tools.  

According to Ndum, this design helps address three important concerns around AI use in high-consequence, knowledge-intensive systems: safety, trust and faithfulness. Safety comes from keeping people involved in decisions. Trust comes from grounding the system in physics and engineering logic. Faithfulness means the system stays close to verified reactor behavior.   

A digital twin is a virtual representation of a system that is updated in real time. This idea can apply to both existing nuclear facilities and new advanced designs. It focuses on accuracy and timeliness, which help improve safety and efficiency in nuclear operations.  

Using the digital twin, the testbed stays aligned with how a nuclear reactor would evolve under changing conditions. The twin reflects the system’s evolving condition with enough realism to support analysis, control studies and decision support. For advanced reactors, this is especially important because many of these systems do not yet physically exist. However, they are well-developed engineering concepts with active developers and strong scientific foundations.   

“Using the digital twin means we can build a rigorous virtual testbed and use it to explore design optimization, control strategies and difficult edge cases — including AI-assisted control — that cannot be easily tested on a physical plant today,” Ndum said. “That is the key tweak I made in defining the digital twin for this work.”   

An additional benefit to these findings is that the underlying framework is broad. The physics-informed backbone can be adapted to other reactor concepts, and the AI layer can be updated as needed. This adaptability means that the twin is not tied to a specific AI medium or reactor but can be adapted to the user’s needs.   

“What matters is not one specific reactor type or one specific large language model,” Ndum said. “What really matters is the structure: a physics-based digital twin, trusted domain knowledge through retrieval, specialized tools or subroutines and human oversight. That is what makes the system portable, model-agnostic and better suited for long-term use in nuclear engineering.”   

AROMA-GPT is part of a larger research program Ndum is developing focused on physics-informed, human-in-the-loop generative AI for nuclear engineering. The larger goal is to create a suite of trustworthy AI-assisted applications that can help accelerate the development, analysis, monitoring, training and eventual deployment of advanced reactor technologies by connecting established physics codes, digital twins and curated nuclear knowledge bases to specialized AI agents. In that broader framework, AROMA-GPT represents the real-time supervisory and monitoring extension of the same research vision.  

This research was conducted in the Scientific Machine Learning for Advanced Reactor Technologies (SMART) Lab, directed by Liu.  

By Michelle Revels, Texas A&M University College of Engineering

###