Classification of opacity for polymer nanocomposite films via deep neural network (DNN) classifiers

Abstract

Deep learning methods are of utmost importance in the field of nanotechnology due to their practical applications provide insights into an optimal design of nanomaterials with multi-characteristics. In the context of present research, we propose fully connected deep neural network (DNN) classifiers which have the capability to classify light transmission from polymer nanocomposite films made up of polystyrene (PS) latex particles and multi-walled carbon nanotubes (MWCNTs). For this purpose, collected spectroscopic data were first divided into three classes based on transmitted light intensity, mass fraction of MWCNT nanofillers, annealing temperature, and particle diameter of PS latexes. Bayesian optimizer has then been implemented for each proposed DNN classifier and the most optimal hyperparameters such as activation functions and hidden layer sizes, which provide the best classification accuracy, were acquired through trial-and-error method. Accuracy, cross-entropy loss, precision, recall and F1-score metrics together with confusion matrix and area under curve computed for receiver operating characteristic (ROC) curves have been extensively employed to assess the performance of proposed DNN classifiers. The highest accuracy, precision, recall and F1-score metrics can be achieved for both training and testing data sets when two hidden layer sizes are set to 30 and 20, respectively and sigmoid functions are used in those of hidden layer units. Computational results have indicated that DNNs can be exploited to classify optical transparency of film samples even with a limited amount of experimental data.