Robust S3Former deep learning model for the direct diagnosis and prediction of natural organic matter (NOM) from three-dimensional excitation-emission-matrix (3D-EEM) data.

The non-destructive, three-dimensional excitation-emission matrix (3D-EEM) based on fluorescence spectroscopy has been widely used in natural organic matter (NOM) monitoring in aquatic environments. However, the direct recognition of the species and concentration of NOM from 3D-EEM data remains challenging, especially under more complex scenarios such as in NOM mixtures and during ultrafiltration. In this work, first, we constructed a 3D-EEM spectral dataset for multi-task learning under various conditions. Then, we integrated the Transformer framework comprising a sparse spatial spectrum (S3) -aware attention mechanism and developed a novel deep learning model (S3Former) for identifying and predicting NOM species and concentrations. Our S3Former model excellently diagnosed sampled unary and binary NOM species (100% and 97.2%, respectively) and their concentrations (75% and 65.6%, respectively). The prediction accuracy, stability, and parameter efficiency of S3Former outperformed ten convolutional neural networks (CNN) models and the bare Transformer model. Finally, we applied the S3Former model to predict the NOM concentration ranges of ultrafiltration permeates, a more challenging prediction scenario due to the change in NOM molecular weights. Notably, our model demonstrated its robustness with good accuracy (up to 77% for humic acid). These results highlighted the potential of using S3Former to enhance automated water quality monitoring and control in water treatment.