A hybrid machine learning approach for improving fuel temperature prediction of research reactors under mix convection regime

Benchmarking results from experiments on research reactors showed that power reactors’ mathematical model produced conservative results in predicting the maximum cladding temperature on hot channels, with the mean difference showing overestimation. This overestimation is an accuracy issue arising from: the rigid demand and requirement for highly specialized expertise needed for input preparation of large and complex mathematical models in a computer program, the simplification and assumptions when translating physical phenomena into mathematical models, the complexity of the cooling regime, and the specific characteristics of research reactors. This paper aims to overcome the accuracy issue using a hybrid method that combines mathematical models, machine learning, and experimental results. Machine learning compensates for the bias between experimental results and mathematical models and discovers the factors affecting the mismatch. Our experimental results indicated that the proposed hybrid method has significantly better accuracy than the power reactors’ mathematical model and can discover the affecting factors.