Auditing how an AI weather model uses lower-stratospheric and moisture information: structure-destroying interventions in GraphCast

Weather forecasting has long been the domain of massive supercomputers solving complex equations that describe how air moves, heats up, and cools down. In recent years, a new kind of weather forecaster has emerged, one that does not solve those equations but instead learns to predict the future by studying vast archives of past weather data. These artificial intelligence models, such as one called GraphCast, have become remarkably accurate, rivaling the traditional physics-based systems that have guided meteorology for decades. However, because these AI models learn patterns rather than following explicit physical laws, a critical question remains: do they actually understand the physics of the atmosphere, or are they simply finding shortcuts in the data? To answer this, scientists must look inside the "black box" of the model to see exactly which pieces of information it relies on to make its predictions. Two specific parts of the atmosphere are of particular interest: the lower stratosphere, the layer of air just above the weather systems we experience, and the moisture content within the air, which fuels storms and shapes climate patterns.

A recent study set out to audit how GraphCast uses information from these two regions. The researchers treated the AI model like a virtual laboratory, running thousands of simulated forecasts where they deliberately altered the input data to see how the model reacted. They focused on a specific version of the model that looks at the atmosphere up to a height of 50 hPa, which includes the lower stratosphere. To test the importance of the lower stratosphere, the researchers performed a series of experiments where they replaced the detailed, swirling weather patterns in that upper layer with a smooth, averaged-out version that contained no specific structures. They also tried replacing the real data with random noise or with weather patterns from a different year. The goal was to see if the model's predictions for the weather below would collapse without the specific, organized structure of the air above, or if it would simply ignore that layer entirely.

The results revealed that the model pays close attention to the organized structure of the lower stratosphere. When the researchers removed the swirling eddies and replaced them with a smooth average, the model's forecast for the height of the pressure surface at 500 hPa—a standard measure of weather patterns in the mid-latitudes—deviated significantly from its normal performance. After five days, this deviation grew to 55 meters, a substantial error in meteorological terms. Crucially, the model reacted much more strongly to this loss of structure than it did to random, incoherent noise that had the same overall energy. This suggests the model is not just sensing the general warmth or coldness of the upper air, but is specifically tracking the coherent, balanced swirls of the atmosphere. Interestingly, the model treated a smooth, climatological average of the upper atmosphere the same as a real but wrong weather pattern from a different year; both caused similar errors. This implies that for the model, having a realistic but incorrect structure is just as damaging as having no structure at all.

The study also compared the influence of the upper atmosphere to the influence of the middle atmosphere, where most weather happens. When the researchers introduced disturbances of equal strength in the middle layers versus the upper layers, the middle layers had a much stronger immediate effect on the forecast. However, the influence of the upper layers was not negligible, and it was found to be most significant during the winter months in the Northern Hemisphere. This aligns with the understanding that the connection between the stratosphere and the weather below is strongest when the polar jet stream is active. The researchers found that the model uses this upper-level information to shape the large-scale weather systems that move across continents, rather than just reacting to local conditions.

In a second set of experiments, the team investigated how the model handles moisture. They systematically dried out or moistened the entire atmosphere in their simulations, scaling the amount of water vapor up or down across all levels. They found a clear, graded response: drying the air weakened the intensity of weather systems, while adding moisture strengthened them. This behavior matched physical expectations, as the release of heat when water vapor turns into rain is a primary engine for storms. To ensure this was a genuine physical response and not an artifact of the model, they compared it to a control experiment where they added random temperature noise of the same mathematical strength. The temperature noise caused the model to fail catastrophically, driving precipitation rates into impossible negative numbers and destroying the forecast skill. In contrast, the moisture changes produced a sensible, proportional reaction, confirming that the model has learned a robust relationship between humidity and storm intensity.

The study also uncovered a fascinating asymmetry in how the model recovers from these changes. When the atmosphere was dried out, the model took a long time to restore the missing moisture, with the deficit fading slowly over several days. However, when the atmosphere was made too wet, the model corrected the excess much faster, removing the surplus moisture in a matter of hours through a burst of simulated rain. Despite this rapid cleanup, about one-third of the extra moisture persisted even after ten days, suggesting that while the model is good at reacting to wet conditions, it retains a lingering memory of the excess. These findings provide a clear map of how a leading AI weather model processes information, showing that it relies on the specific structure of the upper atmosphere and maintains a realistic, albeit imperfect, relationship with global moisture. By proving that the model uses these physical cues in a structured way, the research offers a new method for auditing artificial intelligence, ensuring that these powerful tools are not just memorizing the past, but are learning the rules that govern our weather.