The prediction of the mechanical properties of hot-rolled strip products by means of hybrid methods

The aim of this multipartner, multinational project was to acquire the essential metallurgical know-how needed to develop and establish the physical-neural models that would predict the variation of mechanical properties across the length and width of the product. The mechanical properties of steels in the hot rolled condition should be predicted for a wide range of material compositions - based on the commercial steel grades of the partners. Work performed earlier has shown that it is possible to describe many relationships related to mechanical properties relatively well using semi-empirical physical models. This approach has certain limitations, however, because generally all the required information is not available. On the other hand, neural models are based completely on mathematics and the underlying metallurgical phenomena cannot always be analysed reliably. Hence, by combining physical models with neural networks or by developing neural modelling together with physical models, it is possible to generate a new tool, the hybrid method. It should enable the better development of processes and products and significantly reduce testing costs. In order to establish a firmer basis for physically based statistical regression models, an extensive literature survey was performed. Empirical equations available from the literature to describe the microstructural evolution in the course of hot rolling and subsequent cooling, from reheating to cooling on the run-out table, were collected and complementary data were generated. In the microstructural model, the input is the time-temperature-strain sequence during hot rolling. Then, the predicted microstructure is used as an input to the flow stress model that calculates the stress-strain relationship, thereby giving R p 0 . 2 , R m and Ag. A neural network model able to account for various types of steel was the second main objective in the development of the hybrid model. Before modelling, the steels were classified into five classes based on the chemical composition: C-Mn, HSLA, ferritic-rolled, Ti-alloyed and V-alloyed. As a result of mathematical analyses (e.g. S.O.M), it was decided to use the chemical composition, and the rolling and coiling temperatures as well as the final thickness as input vectors in the neural network modelling. The test results showed clearly that the neural network models were operating successfully for various steel types. It was seen that the hybrid model functioned well when output from Thermo-Calc software was used as the input data for the neural networks. These calculations can create fast and reliable neural network models, developed using the MEFNET program. The main objective was to verify and refine the data before bringing it into the neural network or other models, to detect complex relationships between the process parameters and to derive metallurgical process models using neural network-based modelling. The main idea of this project was to develop hybrid models that combine physical models including Thermo-Calc software with neural networks. Furthermore, a hybrid model was also developed that specifies the results of neural network calculations using physical models. However, it was realised that using this preliminary hybrid model for this kind of application, the accuracy of the predictions of mechanical properties of hot-rolled products is not improved essentially, so that this development stage can only be taken as an intermediate one.