Intelligent analysis and interpretation of aquatic organic matter fluorescence excitation-emission matrix.

Although fluorescence excitation-emission matrix (EEM) has been widely applied in the field of water environment, fluorescence decomposition, identification, and interpretation still rely heavily on manual experience. This study constructed an automatic and intelligent processing flow for analyzing EEMs. Firstly, the Rayleigh scattering was identified and removed using visual learning, and the missing values were interpolated with machine learning. The parallel factor analysis (PARAFAC) was automated to enable componential decomposition, and the obtained components were identified and interpreted with fluorescence database and generative artificial intelligence. The results indicated that the visual learning based on YOLO, superior to traditional methods (RobustPARAFAC and CMDR_PCA), could automatically, accurately, and effectively remove the Rayleigh scattering. The full-automatic PARAFAC analysis developed on Python platform performed comparably to the semi-artificial approach (using drEEM in Matlab) in decomposing the overlapping components. A fluorescence component database and the corresponding knowledge base were constructed with opensource fluorescence data and the related literature knowledge for component matching and evidence-based retrieval. Generative artificial intelligence (Gemini, Qwen, ChatGPT, Claude, GLM, DeepSeek) was employed to allow result attribution and semantic interpretation. With all the functional modules integrated, a three-dimensional fluorescence AI-assisted EEM analysis workflow was constructed and validated in scenarios including rivers, lakes, laboratory-prepared samples, and activated sludge. The AI-assisted EEM analysis workflow can significantly improve the automation, consistency, and interpretability of fluorescence EEM analysis, providing a feasible path for the implementation of intelligent water fluorescence analysis.