Large language model-assisted digital twin for remote monitoring and control of advanced reactors

In this paper, we present a novel framework for real-time autonomous monitoring and control of advanced small-scale nuclear reactors by integrating a simulator-based digital twin (DT) with a domain-knowledge-enhanced large language model (LLM). The framework’s architecture, featuring robust bi-directional connectivity via the OPC-UA industrial protocol, was demonstrated on a lead-cooled fast reactor (LFR) case study. The LFR demonstration leverages a customized, MATLAB-Simulink-based simulator with natural circulation and burnup modeling capabilities, which has been validated against the System Analysis Module (SAM) code. The system employs a hybrid supervisory methodology that incorporates both a traditional rate-limiting Reference Governor (rl-RG) and a novel, tool-centric LLM agent. The LLM agent’s primary capability involves processing incoming sensor data streams, invoking the simulator and related customized tools to obtain physics-based results, and interpreting this data to understand the current system status. Augmented with custom-built retrieval and diagnostic subroutines, the LLM agent then explains the reactor’s state and generates context-aware, quantitative recommendations for management actions, such as reactivity control and power maneuvers. We found that the control signals recommended by the LLM agent successfully matched those generated by the traditional rl-RG. Furthermore, the framework includes an intuitive, multi-tab graphical user interface that enhances human-machine interaction by enabling real-time transient visualization, interactive LLM queries, and detailed transient analysis. This research highlights the transformative potential of AI-enhanced, DT-based control frameworks for advanced reactors and demonstrates its utility as a robust educational and training tool.