Abstract
Fourier Transform Infrared Spectroscopy (FTIR) is a powerful technique for polymer analysis. Conventional FTIR analysis struggles with complex patterns, and expertise is required, limiting scalability in high-throughput environments. Machine Learning (ML) offers a promising route to accelerate data processing by objectively identifying and classifying spectral patterns. This work aims to identify polymers from their FTIR spectra with ML. A novel methodology inspired by chemical intuition is proposed, combining unsupervised dimensionality reduction of the FTIR spectra, bond fraction prediction via deep learning, and polymer identification through matching bond fraction predictions with a reference file. Several architectures are explored, with direct polymer classification used as a benchmark. Additionally, a Neural Network (NN) is designed to predict the oxidative degradation state of poly(ethylene) and poly(propylene) samples. For the polymer bond fraction prediction, the best results are obtained using a NN with latent values of the autoencoded FTIR spectra as input, achieving a square root of the mean squared error of 0.023, and correspond to the bond fractions from the polymer's repeating unit. The highest classification accuracy (75%) is obtained after augmenting both the reference file and spectral training data, using Euclidean distance as a matching method to the reference file. The oxidation detection algorithm reached 100% accuracy by restricting the spectral range input to the carbonyl region (1, 800-1, 550 cm-1), where oxidation features appear. The proposed methodology enables the identification of both seen and unseen polymers during training, outperforming direct classification methods and establishing a scalable framework for automated polymer identification.