Precise predictive property models of tensile properties of highly-demanded TMR (thermo-mechanically-rolled) micro-alloyed steels are indispensable for cost-effective steel grade design and industrial online-automated control of AHSS product properties during processing. To improve the prediction power of existing models, the current study has embraced robust artificial neural networks “ANN” modeling and data-mined an extensive industrial database of cleaned and validated 9,732 record from an industrial HSM (hot strip mill). MATLAB® has been used for modelling with statistical analysis using SPSS®.
Comparably to existing models in literature, two additional elaborations have been, uniquely introduced to current model in order to get “black-box” ANN models more “transparent” to the fundamentals of materials science which ultimately results in better predictive power of the developed model.
The first procedure is incorporating, though unavailable in most industrial databases, the microstructural parameters (final ferrite grain size, volume fraction and simulated prior austenite grain size “PAGS”) as calculated from other empirical models in literature. The second procedure, unlike just feeding raw processing parameters (TMR temperatures and cooling rate) into ANN modelling tool, data have been metallurgically-conditioned by normalizing against critical transformation temperatures to reflect the contribution of driving force.
In terms of goodness of fit, residual error analysis and sensitivity analysis, the developed models show comparably better results regarding ultimate tensile and yield strength than other models in literature or those that regard metallurgical-conditioning manifesting the significance of proposed metallurgical normalization of raw processing parameters in modelling.