Using Explainable AI and Transfer Learning to understand and predict the maintenance of Atlantic blocking with limited observational data

Explainable AI can identify physical causes of extreme weather: The researchers used SHAP (Shapley Additive ExPlanation) analysis to reveal that specific atmospheric patterns—high-pressure anomalies in the American Southeast and North Atlantic separated by a trough over Atlantic Canada—are key predictors of prolonged Atlantic blocking events, validating findings from traditional meteorological analysis.

AI models reveal gaps in climate model performance: The study demonstrates that current climate models underestimate how long blocking events persist, and explainable AI methods can help diagnose why these deficiencies exist by identifying which physical features the models are missing or misrepresenting.

Transfer learning addresses real-world data scarcity: When real observational data (ERA5 reanalysis) proved insufficient for accurate predictions, the researchers improved results by pre-training their AI model on simulated data from an idealized model first—a practical solution for situations where limited real-world observations constrain AI development.

Explainability enables validation against established science: SHAP analysis allowed researchers to confirm their AI's predictions aligned with decades of prior meteorological research on blocking precursors, demonstrating that explainable AI isn't just a "black box" but can produce scientifically credible insights.

Responsible AI development requires multiple data sources: The work illustrates how responsible AI for climate science may require combining simulated, observational, and reanalysis data strategically rather than relying on any single source alone.