KliC. Windsor, G. Cottrell, R. Kemp

In designing a fusion power plant it is important to be able to predict the metallurgical properties of steels at the irradiation levels from 20 to 100 displacements per atom (dpa) experienced due to high-energy neutron bombardment. In particular it is necessary to assess the accuracy of extrapolating data collected from low irradiation dose specimens to those expected from high-dose specimens. Databases are available for the yield stress and Charpy transition temperature shift of low activation ferritic alloys as a function of irradiation level. Published studies have been made of the prediction accuracy of a test dataset containing the yield stress and Charpy ductile to brittle transition temperature shift with irradiation of alloys at irradiations 20 to 100 dpa using neural networks trained below this irradiation level [1,2]. This work considers the reliability of neural networks in this “extrapolation” mode compared to the conventional interpolation mode. The conclusion is that in order to obtain good results it is necessary to impose dimensionality reduction techniques which improve prediction results at the expense of fit to the training data.
This work was supported by the United Kingdom Engineering and Sciences Research Council and the European Communities under the contract of Association between UKAEA and EURATOM. The views and opinions expressed herein do not necessarily reflect those of the European Commission.[1] Windsor et al, Modelling Simul. Mater. Sci. Eng. 16 (2008) 025005[2] Windsor et al, Modelling Simul. Mater. Sci. Eng. 16  (2008) 075008.

Extrapolating sparse data

Download complete paper