Identification and classification of disordered carbon materials in a composite matrix through machine learning approach integrated with Raman mapping

Abstract

Identification and classification of different types of highly disordered carbon materials present in a polymer matrix with similar Raman spectra have been carried out using a machine learning approach. Convolutional neural network (CNN) has been used for the classification of disordered carbon materials such as graphene oxide (GO), functionalized carbon nanotube (f-CNT), carbon fiber (Cf), carbon black (CB), pyrolytic carbon (PyC), coke, and mesocarbon microbeads (MCMB). The hyperparameters of the machine learning model have been optimized by the Bayesian optimization algorithm. CNN gave an accuracy of 91.9 %, F1 score of 92 %, precision of 92.1 %, recall of 92 % and area-under-curve (AUC) of 0.99. Gradient-weighted Class Activation Mapping (Grad-CAM) has been utilized for getting the explainability in the classification process of the CNN model. The CNN along with Raman area scanning could successfully map different disordered carbon materials in the polymer matrix composite.