Artificial Intelligence-Assisted FT-IR Spectral Sensing: From Molecular Fingerprints to Quantitative and Interpretable Chemical Information

Fourier-transform infrared (FT-IR) spectroscopy provides rapid, accessible, and chemically informative molecular fingerprints, but quantitative analysis of complex samples is often hindered by band overlap, baseline variation, matrix effects, and nonlinear relationships between spectra and target properties. This review examines how artificial intelligence (AI), including machine learning, deep learning, and explainable artificial intelligence (XAI), can convert FT-IR fingerprints into quantitative and interpretable chemical information. Recent studies were organized according to analytical function and examined with respect to spectral acquisition and data organization, preprocessing and input preparation, model development, validation, uncertainty, transferability, and chemical interpretation. AI-assisted FT-IR has been applied to composition and concentration prediction, adulteration and contamination-level estimation, physicochemical and material-attribute prediction, and process-, time-, and state-dependent prediction. Across these applications, robust performance depends on representative independent samples, leakage-free validation, transparent preprocessing, uncertainty assessment, calibration transfer, and spectroscopically plausible interpretation of model-relevant regions. Data augmentation, transfer learning, and XAI may improve model robustness and interpretability, but they require validation with independent real samples. Overall, AI can extend FT-IR from qualitative fingerprinting to quantitative and predictive spectral sensing, provided that future studies emphasize method-centered validation, reproducibility, transferability, and chemically meaningful explanation rather than relying solely on classification accuracy and regression error metrics.