AI Peers Into the Black Box of Heatwaves and Reveals What Really Drives Them

When a heatwave shatters records, the question that follows is almost always the same: what caused it? Was it a stubborn high-pressure system parked over the region, a parched landscape baking under the sun, or simply the relentless background warming from rising greenhouse gases? Scientists have long known that all three ingredients matter, but weighing their individual contributions has been notoriously difficult, because the drivers interact in tangled, nonlinear ways. Now a team at the Barcelona Supercomputing Center and the University of Barcelona has turned to a surprising ally: explainable artificial intelligence, a branch of machine learning designed not just to make predictions but to justify them. Their study, published in the journal Weather and Climate Dynamics, offers one of the most detailed quantitative decompositions yet of what actually powers summer heat extremes across Europe and North Africa.

The researchers, led by Arnau Garcia Mesa, focused on six locations chosen to span a wide range of climatic behavior: Córdoba in southern Spain, Marrakech in Morocco, Lyon in France, Hannover in Germany, Belgrade in Serbia, and Stockholm in Sweden. These sites run the gamut from water-limited regimes, where soils are almost always dry in summer, to energy-limited regimes, where abundant moisture means the land surface and atmosphere interact very differently. For each location, the team assembled more than seven decades of data from the ERA5 and ERA5-Land reanalyses, covering daily maximum temperatures, large-scale atmospheric circulation fields, soil moisture at three depth levels, and seasonal carbon dioxide concentrations measured at NOAA’s Mauna Loa Observatory.

The machine learning framework at the heart of the study is a hybrid architecture the authors call the CombinedModel. Atmospheric data, which arrives as spatial maps of geopotential height at 500 and 200 hectopascals and sea-level pressure, is processed by a ConvNeXt convolutional neural network, a modern image-recognition architecture that had not previously been applied to driver quantification in this context. Local land-surface data and carbon dioxide values, which arrive as simple time series, are handled by a multilayer perceptron. The outputs of both branches are merged in a final classification network that decides, day by day, whether conditions resemble a hot extreme, defined as a daily maximum temperature exceeding the 90th percentile of the 1950–2000 climatology. Because extreme days are rare, the team used a weighted loss function to prevent the model from simply defaulting to predicting ordinary weather, and they trained an ensemble of twenty members to tame the randomness of network initialization.

But prediction alone was not the goal. The real innovation lies in the interpretability layer. Using SHAP values, short for SHapley Additive exPlanations, the researchers borrowed a technique from game theory that assigns each input feature a precise, additive share of the model’s output. Positive SHAP values push the model toward predicting an extreme; negative values push it away. Because the values are additive, they can be summed across grid points and variables to produce a global accounting of driver importance. The team computed these explanations only for the fifty percent most confident extreme predictions in the 2014–2024 test period, ensuring the attribution rested on predictions the network itself trusted.

The headline result is strikingly consistent: atmospheric circulation dominates everywhere. Across all six locations, circulation features accounted for 66 to 86 percent of the total explanatory power, with the geopotential at 500 hectopascals, the mid-tropospheric field that meteorologists use to spot blocking highs and ridges, contributing the most by a wide margin. This aligns with a growing body of work linking European heatwaves to persistent double jets over Eurasia, deep North Atlantic depressions, and Saharan warm-air intrusions, and it echoes a 2025 global analysis finding that 500-hectopascal geopotential is the dominant heat-extreme driver over most of the world’s land area. In other words, the large-scale shape of the atmosphere is the primary engine of a hot day, almost regardless of where you are.

The land surface, by contrast, tells a sharply regional story. Soil moisture’s share of the explanation follows a clear northward gradient: a negligible 0.2 percent in Marrakech, a moderate 9.4 percent in Córdoba, and a substantial 15.4 percent in Lyon, with similarly elevated values in the temperate and northern sites. The pattern makes physical sense. In already-arid Marrakech, summer soils are so consistently dry that year-to-year variability is small, leaving little for the land surface to modulate. In wetter transitional climates, however, a dry soil anomaly can dramatically amplify a heatwave by suppressing evaporation and shifting the surface energy balance toward sensible heating. The SHAP analysis captured this mechanism directly: for the shallowest soil layer, negative moisture anomalies were consistently associated with positive contributions to extreme-heat predictions, while deeper layers showed weaker and less robust signals, likely because neighboring soil layers are so strongly correlated that the model cannot cleanly separate their effects.

Carbon dioxide emerged as the third actor, and its role was revealing. The CO2 feature carried more than ten percent of the explanatory weight at most sites, and its importance was highest precisely where the yearly count of extreme days showed the strongest warming trend. By including CO2 explicitly, the framework controlled for the long-term anthropogenic signal without having to detrend the data, allowing a cleaner attribution of the physical drivers operating on daily timescales. The researchers stress that the explainability outputs represent the model’s interpretation rather than a direct measurement of reality, but the fact that the identified contributions reproduce well-established physical relationships lends the approach considerable credibility.

Robustness checks formed a major part of the study. When the team redefined extremes using the 80th percentile threshold, the same architecture recovered the same picture: circulation still dominated at 64 to 84 percent, and the northward gradient in land importance persisted, though soil moisture’s relative share shrank, suggesting that land-atmosphere feedbacks are not fully activated in milder heat events. A 95th-percentile model, hampered by too few extreme samples, proved unreliable and was set aside. The team also swapped reanalysis data for observations, using the E-OBS gridded temperature dataset and drought indices computed from it as soil-moisture proxies. The Standardized Precipitation Evapotranspiration Index, which incorporates evapotranspiration and thus better represents the surface water balance, reproduced the original results closely, while the precipitation-only Standardized Precipitation Index captured less of the land signal, a finding with practical implications for applying the framework to climate models, where soil layer definitions differ.

Two case studies brought the method down to the scale of individual disasters. For the August 2021 heatwave that pushed temperatures in Córdoba province to a scorching 47.6 degrees Celsius, the model attributed the event overwhelmingly to circulation, with positive SHAP values appearing precisely over the Iberian Peninsula where positive 500-hectopascal geopotential anomalies were located, a remarkable demonstration of spatial locality in the network’s reasoning. Intriguingly, CO2 ranked as the second most important feature even for this single event, hinting at a detectable anthropogenic fingerprint in an individual heatwave. The July and August 2018 heatwave in Hannover told a different story: the shallow soil-moisture layer played a much larger role, with SHAP values at times rivaling the circulation signal, consistent with the wet transitional regime of northern Germany, where the exceptional 2018 drought turned abundant soils into amplifiers of extreme heat.

The authors see their framework as more than an attribution exercise. Because it is data-driven yet physically interpretable, it could be applied to climate model output, to subseasonal-to-seasonal prediction systems, or to new regions, helping to diagnose where models capture or miss key feedbacks and to target climate risk assessments where they matter most. Future work, they note, should test whether the results hold across different network architectures and whether adding humidity fluxes or aerosol information sharpens the picture further. For now, the study offers a compelling demonstration that artificial intelligence, when forced to explain itself, can do more than forecast the weather; it can tell us, in numbers, why the weather turned deadly.

Subject of Research: Drivers of summer heat extremes in Europe and North Africa quantified with explainable machine learning

Article Title: Quantifying atmospheric and land drivers of hot temperature extremes through explainable Artificial Intelligence

Article References: Garcia Mesa, A., Palma, L., Donat, M., Materia, S., Gràvalos Talló, B., & Marcos Matamoros, R. (2026). Quantifying atmospheric and land drivers of hot temperature extremes through explainable Artificial Intelligence. Weather and Climate Dynamics, 7(3), 1709-1731. https://doi.org/10.5194/wcd-7-1709-2026

Image Credits: AI Generated

DOI: 10.5194/wcd-7-1709-2026

Keywords: heatwaves, explainable AI, SHAP values, atmospheric circulation, soil moisture, land-atmosphere coupling, climate change, CO2, machine learning, temperature extremes, ERA5, Europe