From genesis to dissolution: A diagnostic-predictive machine learning framework for the complete lifecycle of...

Mucilage events represent disruptive ecological and socio-economic phenomena in semi-enclosed basins. The Northwestern Adriatic Sea serves as a sentinel system for understanding these outbreaks, providing a proxy for other anthropogenically stressed and warming basins. Despite decades of research, the intricate ecological dynamics dictating the mucilage lifecycle remain poorly understood. This study applies a Machine Learning approach (Balanced Random Forest, Boruta feature selection) to a 25-year weekly dataset (2001-2025) to disentangle the interactions governing mucilage onset, persistence, and disappearance. Our models, validated through repeated cross-validation and Out-of-Bag error estimation, achieved a high predictive accuracy (AUC > 0.90), demonstrating that mucilage is a predictable outcome of "ecological memory" (cumulative influence of multi-month antecedent conditions) rather than a stochastic process. Mucilage onset follows a temporal cascade of drivers: (i) a loading phase driven by total soluble nitrogen (12-week lag); (ii) a priming phase dependent on prolonged water column stability (N²) (4-8 week lag); and (iii) a triggering phase activated by phosphorus accumulation (4-5 week lag) and a sharp thermal threshold exceeding 25 °C (1-2 week lag) abruptly triggering a high-probability outbreak. Mucilage persistence is maintained by water column stratification and low salinity (4 week lag), alongside nitrogen depletion 1 week before. The event's termination is mechanically determined by the breakdown of the pycnocline and the intrusion of mixed waters, occurring when N² drops below the threshold of 2.9 × 10^-3 s^-2. Beyond offering a tool for proactive management, this study highlights the basin's vulnerability to climate change, framing mucilage as the deterministic physiological response to a sequential cascade of nutrient loading, physical confinement, and thermal stress.