Combination of Non-Destructive Methods in Estimating Irradiation-Induced Reactor Pressure Vessel Steel Alloy Embrittlement with Machine Learning

Embrittlement of the reactor pressure vessel is one of the most crucial factors that limit the lifetime of a nuclear reactor, and it is traditionally determined using destructive methods. We demonstrate that machine learning algorithms can build models that estimate the embrittlement, which is measured as the ductile-brittle transition temperature, accurately based on a combination of non-destructive methods. A data set consisting of 29 non-destructively measured parameters and ductile-brittle transition temperature data for 157 samples has been used to train and test three machine learning models. The samples were standard Charpy V-notch samples made out of six different steel alloys and had been treated in different irradiation conditions. Three algorithms were used: linear regression, support vector machine and artificial neural network. The test mean absolute error of approximately 17 °C of our novel non-destructive method is of the same order as the accuracy of a destructive method