Microstructure prediction of semi-solid AZ91D magnesium alloy prepared by mechanical stirring based on regression analysis and neural network

In the present investigation two smart prediction tools, namely the multiple regression analysis and general regression neural network models were developed to predict average grain size and shape factor of the semi-solid AZ91D magnesium alloy microstructure prepared by mechanical stirring. The process parameters (stirring temperature, stirring rate, stirring time) were considered as input variables to establish predictive models. The models were developed using the multiple regression analysis was employed to determine the significance of process parameters on microstructure. In the general regression neural network models, the k-fold cross validation method is used to optimize the smoothing factor. The neural network models were trained, validated and tested. The results show the general regression neural network models achieve higher prediction accuracy for predicted error within 5 % compared with regression models within 10 %, which suggests that the model is more reliable. Finally, the accuracy of models was demonstrated based on experimental verification, asserting that they can provide a foundation for developing a comprehensive prediction system to optimize the structural and processing of semi-solid magnesium alloys.