Application of a Physics-Informed Convolutional Neural Network for Temperature Field Monitoring in Advanced Reactors

Abstract

In this work, the capabilities of a physics-informed Convolutional Neural Network (CNN) to evaluate the temperature distribution in advanced reactors are explored. This algorithm was demonstrated to be capable of reconstructing the temperature field in both solid and fluid domains from a limited set of measurements taken at the boundaries of the domain. This method opens new perspectives on critical limitations of current sensor technologies for deploying the next generation of nuclear reactors. We showcase how the CNN capabilities could benefit two promising candidates among these systems. First, the CNN was tested to reconstruct the temperature fields within the solid region of a High-Temperature Gas Reactor (HTGR) fuel assembly. Industry experience has shown this material is prone to large thermal-mechanical loads close to the allowable limits during operation. Developing an indirect measurement technique is a current demand in this community. In addition, predictions are made for the temperature distribution of the circulating fuel in a Molten Salt Fast Reactor (MSFR). In this case, the use of the CNN is justified by the fact that traditional direct measurements with thermocouples can be unreliable in the salt mixture due to harsh environmental conditions, e.g., oxidation effects, high neutron flux, elevated temperatures, etc. Both the test cases demonstrate the potential of the CNN-based field reconstruction method to fill existing technological gaps and meet the demands of the industry for accurate and non-invasive monitoring techniques.