Prediction of mechanical properties of PWR vessel steel heads containing residual carbon macrosegregation using Artificial Neural Networks

Forged Reactor Pressure Vessel (RPV) heads are manufactured from large ingots that are more than a hundred tonnes in weight, although the forging industry has implemented important measures to reduce carbon macrosegregation in these components, residual chemical inhomogeneities may still be present in the final product. Such remaining chemical inhomogeneities, along with other manufacturing variables, will translate into mechanical properties differences within the part. The present paper describes the development of predictive models that can be used to assess the effect of carbon macrosegregation on some of the most conventional mechanical properties of nuclear RPV heads. The particularity of these models is the use of a Machine Learning method to create them, in this case the neural networks technique built within a Bayesian framework was used to develop tensile property models (UTS, YS and Elongation). A first attempt was also performed to create models for more complex properties such as the Charpy impact toughness and the fracture toughness. The results show that the tensile property models, and to a less extent the Charpy impact and fracture toughness models seem to have captured reasonably well the variable interactions describing the problem involved, this was observed in spite of the reduced number of input variables used to create the models. This is because neural network models can cope with complex input variable interactions more efficiently than conventional linear relationships. The models can therefore be used to explore the effect of specific individual manufacturing variables, as well as variable interactions on the mechanical properties of nuclear RPV heads of steel 16MND5 (equivalent to SA-508 cl 3 steel).