Optimizing polymer classification through FTIR spectra and feature-scaled machine learning models

The research is based on the characterization of polymer materials by Fourier Transform Infrared (FTIR) spectroscopy which is a method of characterizing polymers in terms of their characteristic molecular vibration patterns. Three widely used thermoplastics ABS (Acrylonitrile Butadiene Styrene), PP (Polypropylene) and Nylon-66 were selected and analyzed on the basis of FTIR transmittance spectra. Wavenumbers and the values of percent transmittance were obtained and converted into feature vectors. Gaussian noise was used to boost and clean the data. Machine learning methods that were used to classify the materials included Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), random forest, and logistic regression, where 80% of the data were used to train the machine learning model and 20% to test it. Accuracy measures and confusion matrices were used to determine the performance of the models. Random Forest was the best in terms of performance in Nylon-66 classification; however, the overall accuracy was influenced by the error in classifying ABS and PP. Consequently, k-NN (70) and the Logistic Regression (69) have shown a high performance compared to the Random Forest (57) in terms of overall accuracy. This study reveals that even though the Random Forest could be used to classify a particular set of polymers, polymers that have very similar spectral characteristics are still difficult to differentiate, indicating that the work should be improved by increasing the size of the data set and the number of extracted features.