Disclosing microbial nitrogen metabolism in groundwater via automated machine learning-based analysis: A fluorescence...

Microbial nitrogen metabolism represents a complex and pivotal process underpinning nitrogen pollution in groundwater. Nonetheless, its accurate prediction is often impeded by the limited effectiveness of conventional water quality parameters as reliable proxies. This study develops Automated machine learning (AutoML) frameworks that utilize three-dimensional excitation-emission matrix fluorescence spectroscopy to directly predict microbial nitrogen metabolic pathways in groundwater. Here, we quantified the nitrogen metabolic functions of 176 samples, and compared the predictive performance of various machine learning models, including four classical algorithms and two AutoML frameworks, using conventional water quality parameters against fluorescence spectral features as inputs. While the H2O AutoML framework incorporating unfold principal component analysis derived fluorescence peaks exhibited superior predictive accuracy (R² = 0.72 - 0.86) for most metabolic functions, the Tree-based Pipeline Optimization Tool AutoML framework exhibited enhanced prediction efficacy specifically for ureolytic metabolism. Furthermore, interpretable and causal analyses pinpointed key fluorescent features and quantified actionable thresholds e.g., normalized P1 > 1.5 for denitrification stabilization; normalized FI > 0.06 for ureolysis optimization. This study establishes a novel paradigm for predicting nitrogen metabolism in groundwater, providing actionable thresholds for optimizing nitrogen management in anthropogenically impacted aquifers.