Deep Neural Network-Based Approach for Modeling, Predicting, and Validating Weld Quality and Mechanical Properties of Friction Stir Welded Dissimilar Materials

Abstract

The present investigation highlights the development of a suitable and novel deep neural network-based learning model for accurately predicting the weld quality and mechanical properties of difficult-to-join dissimilar materials. Optimized experimental welding parameters in the available literature were taken as input in the deep neural network (DNN). A feed-forward and back-propagated DNN was developed to apprehend the high non-complexity present in friction stir welding of dissimilar materials. Unlike most neural networks, activation functions were altered between layers, effectively capturing non-linearity. The developed model was used to design an experimental condition for dissimilar friction stir welding of aluminum and titanium. Microstructural characterization of the weld was performed to comprehend the influence of parameters on the quality of the joint produced. A close correlation between the machine-learning model and the experimental results was established. The coefficient of determination \(R^2\) between the predicted strength and the actual strength was 0.95 on the training dataset and 0.9 on the testing dataset. Similarly, \(R^2\) between the predicted strength and the actual strength for the experimental dataset was 0.91, thus making the model suitable for predicting experimental conditions and corresponding mechanical properties with the highest accuracy for any unknown dissimilar friction stir welds.