Strength of Ferritic Steels: Neural Networks and Genetic Programming

An analysis is presented of a complex set of data on the strength of steels as a function of chemical composition, heat treatment, and test temperature. The steels represent a special class designed to resist deformation at elevated temperatures (750–950 K) over time periods in excess of 30 years, whilst serving in hostile environments. The aim was to compare two methods, a neural network based on a Bayesian formulation, and genetic programming in which the data are formulated in an evolutionary procedure. It is found that in the present context, the neural network is able more readily to capture greater complexity in the data whereas a genetic program seems to require greater intervention to achieve an accurate representation.