Prediction of the Transition-Temperature Shift Using Machine Learning Algorithms and the Plotter Database

Abstract

The long-term operating strategy of nuclear plants must ensure the integrity of the vessel, which is subjected to neutron irradiation, causing its embrittlement over time. Embrittlement trend curves used to predict the dependence of the Charpy transition-temperature shift, ΔT41J, with neutron fluence, such as the one adopted in ASTM E900-15, are empirical or semi-empirical formulas based on parameters that characterize irradiation conditions (neutron fluence, flux and temperature), the chemical composition of the steel (copper, nickel, phosphorus and manganese), and the product type (plates, forgings, welds, or so-called standard reference materials (SRMs)). The ASTM (American Society for Testing and Materials) E900-15 trend curve was obtained as a combination of physical and phenomenological models with free parameters fitted using the available surveillance data from nuclear power plants. These data, collected to support ASTM’s E900 effort, open the way to an alternative, purely data-driven approach using machine learning algorithms. In this study, the ASTM PLOTTER database that was used to inform the ASTM E900-15 fit has been employed to train and validate a number of machine learning regression models (multilinear, k-nearest neighbors, decision trees, support vector machines, random forest, AdaBoost, gradient boosting, XGB, and multi-layer perceptron). Optimal results were obtained with gradient boosting, which provided a value of R2 = 0.91 and a root mean squared error ≈10.5 °C for the test dataset. These results outperform the prediction ability of existing trend curves, including ASTM E900-15, reducing the prediction uncertainty by ≈20%. In addition, impurity-based and permutation-based feature importance algorithms were used to identify the variables that most influence ΔT41J (copper, fluence, nickel and temperature, in this order), and individual conditional expectation and interaction plots were used to estimate the specific influence of each of the features.

Keywords

• machine learning
• neutron embrittlement
• gradient boosting