Physics-guided deep learning for improved stability and interpretability in EEM fluorescence decomposition

Three-dimensional excitation-emission matrix (EEM) fluorescence spectroscopy is widely used to characterize dissolved organic matter (DOM). In practice, parallel factor analysis (PARAFAC) often requires stringent quality control and extensive preprocessing, and data-driven machine-learning approaches are constrained by the limited size and sparsity of available EEM datasets. A unified, physically consistent framework is described for EEM decomposition, centered on dual normalization that decouples spectral shape from intensity and supports a stable inverse mapping across the evaluated data settings. Supervised training is conducted on synthetic mixtures generated from OpenFluor reference spectra, and the resulting inverse solver is evaluated on experimentally measured datasets collected under a limited set of measurement settings. Accurate reconstructions are obtained on synthetic data, with mean squared error on the order of 10^-6. On real measurements with limited preprocessing, it maintains stable performance, with mean absolute reconstruction error of 10^-3-10^-2 over the effective region. The recovered component spectra closely match reference fluorescence features, with a mean maximum similarity, a component-level measure of spectral similarity, of about 0.94. The results indicate that synthetic calibration with physically constrained outputs can support stable and interpretable EEM decomposition under the evaluated conditions when real training data are limited. The proposed framework may be conceptually extendable to other linearly mixed spectroscopic modalities.