In 2008, atmospheric scientist Chris Bretherton flew through clouds above the Atacama Desert aboard an instrument-packed C-130 aircraft, gathering data on ice, water vapor, and air pressure. His goal was not sightseeing but confronting one of climate science’s hardest problems: understanding clouds. Nearly two decades later, as global temperatures have risen by about half a degree Celsius, clouds remain the single largest source of uncertainty in climate projections — a challenge often described as cloud feedback uncertainty.
Clouds both cool the planet by reflecting sunlight and warm it by trapping heat, making their net effect extraordinarily difficult to predict. Even the world’s most powerful supercomputers cannot directly simulate clouds at the tiny scales where crucial processes unfold. As a result, climate models rely on approximations called “parameters” to estimate cloud behavior. Small errors in these estimates can translate into dramatically different warming outcomes — sometimes shifting projections by several degrees Celsius — reinforcing why cloud feedback uncertainty dominates disagreements among climate models.
Two leading approaches are now competing to overcome this barrier. Tapio Schneider of Caltech leads efforts within the Climate Modeling Alliance (CLIMA) to improve traditional physics-based models using artificial intelligence. His team combines fluid dynamics equations with machine learning trained on thousands of high-resolution “large-eddy simulations” (LES), which capture cloud turbulence over small regions. With help from Google’s tensor processing units, researchers generated a library of more than 8,000 simulated clouds from across the Pacific, allowing AI systems to automatically tune cloud parameters. Early results suggest CLIMA’s new global model could be twice as accurate as previous generations, potentially narrowing cloud feedback uncertainty.
Bretherton, now at the Allen Institute for Artificial Intelligence, is pursuing a more radical path. Skeptical that equations like Navier–Stokes can ever fully capture cloud complexity, he is developing AI systems that learn directly from real atmospheric data. Inspired by breakthroughs in AI weather forecasting, Bretherton’s team created the Ai2 Climate Emulator (ACE2), a neural network trained on decades of observations. ACE2 can generate realistic atmospheric evolution — including storms and large-scale circulation — in minutes rather than hours, and performs comparably to conventional models for seasonal forecasts.
This data-driven approach offers remarkable speed and flexibility, but it also raises concerns. Neural networks approximate physics rather than obey it exactly, so small errors could compound over long timescales. Moreover, climate change involves conditions outside historical experience, and AI models trained on past data may struggle with unprecedented futures. Some scientists argue that physics-based models, despite their flaws, provide more reliable grounding. The debate reflects a deeper tension: whether future climate prediction should be rooted primarily in physical laws, empirical data, or hybrids of both — all in service of reducing cloud feedback uncertainty.
Increasingly, researchers see these methods as complementary. AI can accelerate physics-based simulations by factors of 100 to 1,000, enabling massive ensembles that explore many possible climate futures. Rather than predicting a single outcome, scientists aim to map probabilities: which scenarios are likely, which are extreme, and how human emissions reshape those odds.
The stakes are enormous. Depending on how clouds respond to warming, Earth could be headed toward roughly 2°C of warming — challenging but potentially manageable — or toward 6°C, a level that could destabilize civilization. While next-generation models promise sharper insight into cloud feedback uncertainty, they cannot decide how humanity responds. As Schneider notes, better forecasts may tell us how dangerous the future could become, but whether that knowledge drives action remains an open question.
www.quantamagazine.org/climate-physicists-face-the-ghosts-in-their-machines-clouds-20260220/