Clouds: The Ghost in Climate Science’s Machine
▼ Summary
– Machine learning and AI are being used by climate scientists to improve climate models, with some researchers integrating AI into physics-based models and others developing AI tools that predict directly from data.
– There is a critical urgency among scientists to improve climate predictions, as a perfect model in the distant future will not help solve the current, rapidly unfolding climate crisis.
– The wide variation in climate warming predictions (from 2°C to 6°C) is largely due to how different models account for clouds, with even small errors in cloud cover leading to vastly different temperature forecasts.
– Even the most advanced supercomputer models cannot directly simulate clouds at the necessary scale, forcing scientists to use estimated parameters in their equations, a subjective and labor-intensive process.
– To automate and improve cloud parameter selection, researchers are creating extensive libraries of simulated cloud data using computationally intensive models, collaborating with organizations like Google to generate the needed volume.
The challenge of accurately predicting future climate hinges on a single, elusive factor: clouds. More than half of the variation between major climate projections stems from how simulations handle clouds, a discrepancy that can mean the difference between a manageable 2 degrees Celsius of warming and a catastrophic 6 degrees. This uncertainty has spurred climate scientists to turn to artificial intelligence, hoping these new tools can finally capture the complex, small-scale behavior of clouds that traditional physics-based models have struggled to represent.
Leading these efforts are physicists like Tapio Schneider of Caltech and Christopher Bretherton of the University of Washington. While both utilize AI, their approaches differ. Schneider is integrating machine learning into next-generation physics models to better estimate cloud effects. Bretherton, skeptical that equations will ever fully capture cloud dynamics, is pioneering AI tools that predict the future directly from observational data, largely bypassing traditional physics. Despite their different methods, they share a pressing concern. “Climate is changing fast,” Bretherton noted. “Having a perfect model in 100 years will not be useful for solving the climate crisis.”
The stakes are immense. Current model predictions for warming over the next half-century vary wildly, and this range is primarily a cloud problem. A discrepancy of just two or three percent in simulated cloud cover can lead to a difference of several degrees in projected warming, according to George Matheou, a physicist at the University of Connecticut. This tiny margin separates a difficult but adaptable future from one of profound instability for human societies.
Modern climate models account for the atmosphere, oceans, land, and ice. Yet, directly simulating clouds remains computationally impossible. Clouds form and evolve on scales of meters, influenced by minute air currents. As Schneider explains, capturing them directly would require computing power a hundred billion times greater than what exists today. Instead, scientists use a workaround: they add estimated parameters to their core physics equations. These parameters are meant to indirectly represent the influence of clouds, steering the digital atmosphere to behave as if clouds were present. Researchers painstakingly adjust these parameters until a model’s past predictions align with historical records. However, with sparse data, the process relies heavily on educated guesswork. “You have to guess a little bit,” Matheou admitted.
To transform this subjective art into a more rigorous science, Schneider founded the Climate Modeling Alliance (CLIMA). The goal is to use AI to automate the selection of optimal parameters. This requires vast amounts of training data on diverse cloud types from around the globe. Since flying research planes through actual clouds is expensive and limited, scientists rely on the next best option: highly detailed simulations known as large-eddy simulations (LES). These LES models are the gold standard for simulating cloud turbulence, but they are extraordinarily computationally expensive, producing only small snapshots of cloud behavior over limited areas and brief times.
For years, the scientific community had generated just a handful of these high-quality cloud simulations, far too few to comprehensively understand cloud behavior or train a sophisticated AI system. Facing this data shortage, Schneider sought a powerful collaborator, turning to researchers at Google for their computational expertise and resources to help build the extensive library of simulated clouds needed to teach the machines.
(Source: Quanta Magazine)