Classification of mechanism of reinforcement in the fiber-matrix interface: Application of Machine Learning on nanoindentation data

Carbon fiber reinforced polymer manufacturing is emerging, with multiple studies to focus on the design of interfacial reinforcement to ensure the maximum of composite properties, but also respectively to be able to align with zero defect manufacturing. The controversy on the engineering approach is a data-driven task that can be efficiently tackled by involving Artificial Intelligence in order to establish unbiased structure-property relations. In the present study, nanoindentation mapping data were processed with Machine Learning classification models to identify the interfacial reinforcement. The data preparation included normalization and sorting out of highly similar data with k-means clustering, since nanoindentation on epoxy matrix does not enhance insight on the mechanism of reinforcement. The trained models included neural networks, classification trees, and support vector machines. Realization of models' performance was evaluated on the test dataset as screening to obtain best fitted models for each algorithm. Transfer learning potential was demonstrated by extrapolating the prediction of best trained models to a validation dataset at different indentation depth with support vector machines outperforming the other models. Overall accuracy was 67% on the test dataset, F1 Score was 65% in the prediction of reinforcement mechanism classes and 72% in case of pristine specimen, while accuracy on validation dataset was 72.7%. Prediction metrics were comparable to other case studies of real-world classification problems. Computational time-cost for tuning and training was sustainable and equal to 2.3 min.