Shapley Additive exPlanations-Integrated Convolutional Neural Networks for Chemically Interpretable Fourier-Transform Infrared-Based Microplastic Characterization

Fourier-transform infrared (FTIR) spectroscopy is a widely adopted technique for microplastic (MP) identification due to its molecular sensitivity and nondestructive nature. However, many FTIR-machine learning (ML) approaches operate as black-box classifiers, providing limited transparency regarding the spectral regions that drive model decisions and thereby constraining confidence in automated analysis. We present CNN-SHAP fusion, an explainable deep-learning framework that integrates one-dimensional convolutional neural networks (1D-CNNs) with SHapley Additive exPlanations (SHAP) to support postacquisition, attribution-based interpretation of FTIR spectra. The framework combines CNN-derived spectral embeddings with SHAP-weighted wavenumber representations as meta-features within an ensemble learning architecture, enabling the systematic evaluation of model sensitivity across the infrared spectrum. Using standardized preprocessing and a balanced dataset comprising six common polymers (HDPE, LDPE, PET, PP, PS, and PVC) measured under controlled laboratory conditions, CNN-SHAP fusion achieves a mean cross-validated classification accuracy of 99.6%. Attribution analysis indicates that model predictions are primarily influenced by spectral regions that are diagnostically relevant for polymer identification, including aliphatic C-H stretching, carbonyl-associated bands, aromatic features, and characteristic fingerprint-region patterns. These attribution profiles are consistent with established FTIR assignments for the polymers examined and remain stable across the cross-validation folds. CNN-SHAP fusion provides a transparent and reproducible framework for FTIR-based microplastic classification in laboratory and batch-processing workflows. By linking predictive performance with spectrally interpretable attribution, the approach supports the informed evaluation of model behavior and establishes a foundation for future validation under more complex and environmentally representative conditions.