Reactor pressure vessel embrittlement: Insights from neural network modelling

Materials for Sustainable Nuclear Energy: A European Strategic Research and Innovation Agenda for All Reactor Generations

“…These sophisticated algorithms turn out to be often very powerful. The specific example in the nuclear materials field where this approach is being applied with some degree of success concerns correlations for RPV steel embrittlement versus neutron fluence and other variables [174][175][176].…”

“…One of the main problems with data-driven modelling procedures is that they are too often blind: the AI produces in most cases a sort of "black box" transfer function between input and output, a priori devoid of any physics (even though sometimes this procedure manages to improve also our physical understanding [174,175]). The more data are available, the higher are the chances that the procedure provides probative results, although it remains dangerous and unwarranted to rely on extrapolations [176].…”

“…The more data are available, the higher are the chances that the procedure provides probative results, although it remains dangerous and unwarranted to rely on extrapolations [176]. In the case of RPV steels, a large amount of data is available from surveillance and MTR experiments, thus this approach is especially promising [174][175][176]. However, this situation is not necessarily common in the nuclear materials field.…”

Machine learning in nuclear materials research

“…Distinct from the semi-empirical models just discussed, there have been five models of RPV embrittlement using standard ML approaches. 94,[110][111][112][113] The earliest work came from Obraztsov et al 94 in 2006, who used a surveillance database of DBTT shifts DTx (we denote this temperature shift DTx as it was not clear from our available references how it was measured) for 41 main metal and weld-seam materials in the VVER-440 vessels. Features included a dozen alloy elements, fluence, power plant number, and a binary coding of main metal or weld seam.…”

“…These results are extremely encouraging and show that ML training on RPV databases can identify essential features and provide quantitative extrapolative predictions. However, it is not clear how well such a model would perform on surveillance alloys, which are irradiated at lower flux and under less controlled conditions, nor does the analysis provide us a clear guideline on what is needed for a model that can be applied for quantitative prediction.A recent study in 2018 from Mathew et al111 explored the use of ensemble Bayesian NNs to model both combined surveillance and test reactor data (the U.S. NRC Embrittlement Data Base (EDB) database114 , which Mathew et al simply refer to as the Nuclear Regulatory (NUREG) database and test reactor data (part of the Irradiation Variables (IVAR) database). Similar features to the previous modeling were included, with a focus on the elements generally acknowledged to be most If one assumed the values are uniformly distributed from 0 to 170 MPa and one simply guessed the mean of 85MPa for all data points, the MAE would be 42.5 MPa, which is only moderately larger than that obtained from the model.…”