Explainable Deep Learning and Targeted Spectral Augmentation for Mid-Infrared Classification of Additive-Containing Polymers

Deep learning architectures are increasingly used for spectral classification because of their ability to achieve high prediction accuracies and adapt to complex, heterogeneous data. However, their limited interpretability and sensitivity to data set heterogeneity remain as barriers to chemical insight and deployment in real-world technologies. Here, we evaluate a one-dimensional convolutional neural network (1D-CNN) trained on mid-infrared (MIR) spectra of virgin polymers and postconsumer plastic waste, with emphasis on additive-induced spectral variability that is often underrepresented in pristine polymer spectral libraries and simulated databases. Using one-dimensional gradient-weighted class activation mapping (1D Grad-CAM++), we identify the MIR spectral features most influential for classification of poly(ethylene terephthalate), high-density polyethylene, low-density polyethylene, polypropylene, and polystyrene. Fourier-transform infrared (FTIR) spectra were projected into a two-dimensional space using t-distributed stochastic neighbor embedding (t-SNE) to identify class overlap and spectral feature contamination. To improve model performance, polyolefin formulations containing common polymer additives, including erucamide (CH₃(CH₂)₇CH═CH(CH₂)₁₁CONH₂), a slip agent, and calcium carbonate (CaCO₃), a mineral filler, were extruded using a melt-mixer and incorporated as “chemistry-informed” training data. This targeted experimental augmentation improved the prediction accuracy of standard machine learning classifiers, including random forest, k-nearest neighbors, support vector machine, and logistic regression, from 79.4% to 86.6% without altering the training, validation, or testing parameters. Overall, this work establishes an integrated workflow that combines explainable deep learning, low-dimensional embedding, and targeted experimental data acquisition to improve MIR-based classification of additive-containing polymers. More broadly, the approach provides a practical foundation for autonomous material classification technologies designed for small, chemically heterogeneous spectral data sets.