Neural network analysis of Charpy transition temperature of irradiated low-activation martensitic steels

Abstract

We have constructed a Bayesian neural network model that predicts the change, due to neutron irradiation, of the Charpy ductile-brittle transition temperature (ΔDBTT) of low activation martensitic steels given a 40-dimensional set of data from the literature. The irradiation doses were < 100 displacements per atom (dpa). Results show the high significance of irradiation temperature and (dpa) 1/2 in determining ΔDBTT. Sparse data regions were identified by the size of the modelling uncertainties, indicating areas where further experimental data are needed. The method has promise for selecting and ranking experiments on future irradiation materials test facilities.

Elongation of Irradiated Steels

“…There have been previous attempts in the context of fusion materials, to model and extrapolate irradiation behaviour using a neural network [5,6]. There are two fundamental differences with the work presented here.…”

“…A great deal of research is underway to study the consequences of such an environment on the properties of the steels, but because appropriate experimental facilities do not exist, much of the work is theoretical. Back-propagation artificial neural networks were recently been used to model the yield strength and the ductile-to-brittle transition temperature of irradiated steels [5,6]. Here we extend the work to estimating the elongation of irradiated steels.…”

Designing optimised experiments for the international fusion materials irradiation facility

“…We also intend to use the models to suggest a composition and heat-treatment for an optimised LAM steel for consideration as a power plant material, and hence for IFMIF experimentation. [4] and [5] for full explanations of these figures).…”

“…We have modelled tensile and Charpy properties for irradiated low activation martensitic (LAM) steels using an artificial neutral network (ANN) approach [4,5]. The modelling method has been described previously; it suffices to say here that it is based on a Bayesian inference method which gives as outputs not only a numerical prediction, but also the important modelling uncertainty in that prediction which includes experimental noise, the ability of the Journal of Nuclear Materials, Volumes 367-370, 2007, 1586-1589 network to "fit" the data, and how far outside the limits of the input database predictions are being made.…”

Neural Networks and Information in Materials Science

“…There are therefore efforts to estimate the influence of large doses of energetic neutrons on the properties of candidate steels [81][82][83][84]. Most of these efforts include as many variables in the analysis as the availability of the data will permit, on the grounds that Bayesian methods are robust enough to minimize the influence of irrelevant variables, thus avoiding the risk of preconceived bias.…”