The ocean is undergoing significant physical and chemical changes, posing massive challenges to marine biodiversity. But are our AI and machine learning models truly capturing the full scope of this threat?

Knowledge gap

While ocean warming is widely recognized, climate change is also profoundly altering ocean chemistry through acidification and deoxygenation. However, our literature review found that nearly 80% of studies predicting marine species redistributions completely overlook pH and dissolved oxygen levels, relying almost exclusively on temperature. This widespread omission potentially fails to adequately capture the true drivers of marine species’ redistributions.

Main approach

To address this, we conducted a global comparative analysis focusing on 268 cold-water coral species. We utilized machine learning maximum entropy models (MaxEnt) to project potential future distributions across different shared socioeconomic pathways and dispersal scenarios. We then directly compared the outcomes of models developed with a full suite of marine predictors against those that excluded pH and oxygen.

Technological challenge - how we tackle the study

Comparing models with different numbers of predictors is technically challenging, as adding variables can artificially inflate apparent model performance. We tackled this by employing an information criterion approach (AICc), which measures goodness-of-fit while strictly penalizing unnecessary model complexity. Additionally, we applied spatial block cross-validation to assess how well the models perform on withheld data and transfer to novel, non-analog future environmental conditions.

Main finding

Our results demonstrate that, irrespective of the specific scenario, timeframe, or dispersal assumption, models incorporating pH and dissolved oxygen consistently projected an 11.5% to 21.4% higher impact from climate change. For example, under a high-emission scenario by the end of the century, these comprehensive models projected an average geographic range contraction of 48.2% for cold-water corals, compared to only 26.8% in models lacking those two variables.

Main implications

Relying solely on temperature predictors risks severely underestimating extinction risks and misrepresenting future habitat shifts. Future biodiversity forecasts must systematically integrate a broader suite of ecologically meaningful variables to accurately identify climate refugia. Improving these predictive tools is vital for informing conservation planning, prioritizing management interventions, and designing resilient marine protected areas to support global sustainable development goals.