A Critical Review of Machine Learning Analyses of Reactor Pressure Vessel Embrittlement and Hardening

Predicting the irradiation-induced embrittlement or hardening of nuclear reactor pressure vessels (RPVs) is key for safe long-term operation. Embrittlement trend curves (ETCs) are used to predict RPV embrittlement as a function of chemical and irradiation environment variables. There is increasing interest in applying machine learning (ML) to available mechanical testing data to produce ETCs, as opposed to conventional regression analysis. This work provides a focused review of the current state of ML applications to RPV embrittlement and identifies successes, failures and opportunities for the future.

The review shows that a wide range of ML techniques and approaches have been applied to the analysis of RPV steel embrittlement data. ML models are capable of predicting irradiation-induced embrittlement even more accurately than the leading analytical ETCs, however, this is only true when the models are interpolating within parameter spaces containing sufficient training data. When extrapolating – e.g., to a higher fluence – ML model accuracy can drastically diminish and, in some cases, fail to make physically reasonable predictions. ML investigations of the contributors to RPV embrittlement can reproduce the dominant known effects but are yet to provide insights not known from decades of coordinated research, mainly due to a lack of studies considering the effects of variables not already known to dominate embrittlement. Future work is recommended for ML to help advance the field of RPV embrittlement and perhaps improve predictions of RPV structural integrity to support long-term operation goals.