Expert-Free Deep Learning of Fluorescence Spectra for Direct Quantification of Organic Micropollutants in Complex Waters

We present an expert-free, end-to-end deep learning framework for the simultaneous quantification of three fluorescent organic micropollutants (OMPs)─ciprofloxacin (CIP), naproxen(NAP), and zolpidem(ZOL)─directly from standardized excitation–emission matrix (EEM) spectra. Using a publicly available data set of natural water and municipal wastewater samples spiked with target compounds (0–50 μg L–1), a lightweight convolutional neural network achieves robust multitarget prediction with a mean overall out-of-fold R2 of 0.984 (CIP: 0.978, NAP: 0.980, ZOL: 0.993) using repeated stratified 5-fold cross-validation, without requiring PARAFAC decomposition or expert-guided spectral interpretation. Prediction accuracy is concentration-dependent, demonstrating practical utility predominantly at moderate-to-high levels (>5 μg L–1), particularly for highly fluorescent compounds. While detection limits (1.0–3.8 μg L–1) preclude routine monitoring of trace environmental concentrations, this framework offers a rapid, automated complementary approach to chromatography for process monitoring in high-strength wastewater streams (e.g., industrial effluents or WWTP influents) and data-intensive laboratory investigations into OMPs fate and removal.