Prediction of groundwater nitrogen metabolic potential based on fluorescence information: An integrated semi-supervised learning-automated machine learning approach

Nitrogen pollution is a major environmental challenge in groundwater systems, and identifying nitrogen metabolic potential is essential for understanding its development and transformation. However, functional characterization relies on metagenomic data, while conventional water-quality monitoring does not directly quantify pathway-specific N-cycle gene abundance. Fluorescence spectroscopy offers a rapid and convenient alternative, but its predictive potential for groundwater nitrogen metabolism remains insufficiently explored. In this study, 3D-EEM fluorescence spectroscopy was integrated with the H2O automated machine learning (AutoML) framework and five reference machine-learning algorithms to predict seven N-cycle gene-abundance targets. Two datasets were used: an original dataset comprising 91 measured samples and an SSL-expanded dataset constructed by screening a candidate pool of 152 records from previous studies through semi-supervised learning (SSL). By generating pseudo-labels based on spectral-functional relationships learned from the measured samples, SSL expanded the training set and enhanced the model’s ability to capture nonlinear relationships. The SSL-H2O AutoML model provided pathway-dependent improvements, with R 2 values of 0.940 and 0.885 for denitrification and nitrogen fixation, respectively. Following SSL-based dataset expansion, the normalized importance of C2 doubled from 16.2% to 32.4%, indicating that SSL strengthened the model’s recognition of this key predictor. These results demonstrate the advantage of SSL under limited labeled data and highlight the potential of machine learning for predicting groundwater nitrogen metabolic potential. This framework provides a rapid and cost-effective supplement to metagenomic characterization and supports applications of data-driven approaches in groundwater research.