The prediction of the hot strength in steels with an integrated phenomenological and artificial neural network model

Abstract The hot torsion data of a commercial 304 stainless steel has been analysed with an integrated phenomenological-artificial neural network model (IPANN), developed from the Estrin–Mecking (EM) phenomenological model and a back-propagation artificial neural network (ANN) model. In order to predict the flow stress in this model, the work-hardening coefficient and its product with the stress were used as inputs, along with strain, temperature and strain rate. The Pearson correlation coefficient was used to evaluate the performance and terminate the simulation of the IPANN model, whilst the standard errors were employed to quantitatively compare the accuracy of different models. The IPANN model is able to predict the distribution of flow stress more accurately in the work-hardening and dynamic recrystallisation regimes in comparison with the original EM and ANN models. The training speed is significantly improved and the test of the model is satisfactory, if reasonable training data is provided. In addition, by using the phenomenological model as training data, the IPANN model may be used for extrapolation.