When oxygen vanishes from the Baltic Sea's deepest waters, entire ecosystems collapse. Fish die en masse. The water itself becomes toxic. For decades, researchers have watched this cycle repeat in Eckernförde Bay and beyond, unable to predict with confidence when the next crisis would strike. Now, a team at Kiel University has found an unexpected solution: the worst ocean models are actually worth keeping.
The oxygen depletion plaguing the Baltic Sea stems from a simple, tragic chain of events. Agricultural runoff and wastewater have overfertilized the water for decades, triggering massive algal blooms. When those blooms die and decompose, the process consumes nearly all dissolved oxygen in deeper layers. As ocean temperatures climb with climate change, these dead zones threaten to worsen. Scientists use complex mathematical models to predict oxygen levels and prepare for ecological disasters, but no single model can capture the full reality of how water circulates and biology unfolds beneath the surface.
Lead author Dr. Ulrike Löptien and her colleagues at Kiel University's Institute of Geosciences took an unusual approach. Rather than discarding their worst-performing models—the ones that failed to match real observations—they kept them all and fed the entire collection into a machine learning algorithm called a random forest. The algorithm learned from actual measurement data where and when each model, even the weakest ones, contained useful information about oxygen dynamics. What they discovered was remarkable: combining models using artificial intelligence dramatically improved predictions in ways that simple averaging never could.
The findings, published on June 8 in Scientific Reports, challenge a fundamental assumption in environmental science. "These models are far from useless," Dr. Löptien explains. "When combined intelligently using machine learning methods, the so-called low performers can provide crucial information about rare or extreme environmental conditions and thereby improve overall predictions." Some of the weakest models proved especially valuable for forecasting extreme oxygen crashes—the very events researchers most urgently need to anticipate.
The researchers deliberately constructed models with extreme parameter settings and unrealistic assumptions, deliberately excluding biological processes entirely in some versions. Traditionally, such models would be thrown away as garbage. Instead, the random forest algorithm recognized that their skewed perspectives on reality contained hidden gems: information about edge cases and outliers that high-performing models simply missed. By weighting each model differently depending on conditions, the machine learning method outperformed standard ensemble averaging by a significant margin.
The implications ripple far beyond Eckernförde Bay. "The diversity within a model ensemble can be of greater benefit than simply relying on high performers," says co-author Dr. Heiner Dietze. Environmental forecasters worldwide use similar ensembles to predict storms, crop yields, and ecosystem changes. This work suggests they may be throwing away valuable information by discarding their worst models. For the Baltic Sea specifically, better oxygen forecasts could help coastal communities prepare for fish kills, guide restoration efforts, and protect the marine life that depends on a functioning ecosystem. As climate change intensifies, that kind of early warning becomes increasingly precious.