The application of machine learning algorithms in understanding the effect of core/shell technique on improving powder compactability

This study systemically investigated the application of core/shell technique to improve powder compactability. A 28-run Design-of-Experiment (DoE) was conducted to evaluate the effects of the type of core and shell materials and their concentrations on tensile strength and brittleness index. Six machine learning algorithms were used to model the relationships of product profile outputs and raw material attribute inputs: response surface methodology (RSM), Support Vector Machine (SVM), and four different types of artificial neural networks (ANN), namely, Backpropagation Neural Network (BPNN), Genetic Algorithm Based BPNN (GA-BPNN), Mind Evolutionary Algorithm Based BPNN (MEA-BPNN), and Extreme Learning Machine (ELM). Their predictive and generalization performance were compared with the training dataset as well as an external dataset. The results indicated that the core/shell technique significantly improved powder compactability over the physical mixture. All machine learning algorithms being evaluated provided acceptable predictability and capability of generalization; furthermore, the ANN algorithms were shown to be more capable of handling convoluted and non-linear patterns of dataset (i.e. the DoE dataset in this study). Using these models, the relationship of product profile outputs and raw material attribute inputs were disclosed and visualized.