Development of Heat Pipe-Cooled Microreactor Condition Monitoring Platform using Deep Learning for Digital Twin Implementation

Abstract

Heat pipe-cooled microreactors have been proposed to enhance the safety, reliability, and deployment of nuclear systems in remote and off-grid environments. The heat pipe-cooled microreactor utilizes passive heat transfer without the need for complex coolant circulation systems, compared to conventional NPPs that rely on active cooling mechanisms. Considering the purpose of microreactor development, the autonomous operation of these systems requires advanced condition monitoring techniques capable of real-time fault detection and predictive diagnostics. Conventional monitoring methods primarily depend on in-core instrumentation and threshold-based anomaly detection, which have limited adaptability, increased maintenance complexity, and the generation of additional radioactive waste. Although advanced condition monitoring technology has been proposed to address the limitation of the conventional method, the large-scale NPPs are usually considered. For regulatory approval, the development should be accompanied by research on advanced condition monitoring technology. In this study, a deep learning-based condition monitoring platform was developed to support the implementation of a digital twin for heat pipe-cooled microreactors. An experimental investigation was conducted to evaluate the thermal performance of heat pipes under various operating conditions, including steady-state operation, startup behavior, inclination changes, loss of cooling accidents, and shutdown sequences. The experimental results were analyzed to investigate heat transfer behavior on system performance. A data-driven monitoring approach was employed to infer fuel temperature and conditions using only indirect temperature measurements from adiabatic and condenser sections of the heat pipe, reducing reliance on in-core instrumentation. Deep learning models were trained using experimental datasets to enable real-time prediction of reactor conditions and anomaly detection. In addition, the discussion model was developed to help the decision making of the operator, using language model. The integration of explainable AI techniques and uncertainty quantification methods enhances the reliability and interpretability of model predictions, ensuring reliability in safety-critical nuclear systems. The proposed condition monitoring system demonstrated that it could detect deviations in heat pipe performance from transient thermal behavior, providing robust diagnostic capabilities for transient scenarios with outstanding performance. The proposed methodology enables early fault detection with a reduction of the dependency on manual inspections, enabling remote reactor operation. This study could contribute to the advancement of digital twin implementation for heat pipe-cooled microreactors by providing a framework for AI- assisted predictive maintenance and autonomous monitoring.