The persistent uncertainty surrounding climate change isn’t about denying the warming trend, but about precisely how much the planet will warm. For decades, climate models have consistently pointed to a significant rise in global temperatures, yet the range of predicted outcomes remains stubbornly wide. This isn’t a failing of the models themselves, but a consequence of a single, deceptively simple component of the Earth’s system: clouds. These ephemeral formations, simultaneously reflecting sunlight and trapping heat, introduce a level of complexity that pushes even the world’s most powerful supercomputers to their limits. The challenge isn’t simply to include clouds in simulations, but to accurately represent their intricate behavior – a task that’s now driving a fascinating convergence of physics, computer science, and artificial intelligence.
The difficulty stems from the scale at which clouds operate. Modern climate simulations meticulously account for the atmosphere, oceans, land, and ice, but resolving the subtle interplay of air currents and water vapor within clouds requires computational power far beyond our current capabilities. As George Matheou, a physicist at the University of Connecticut, explains, “If you are off by a few percent – 2 or 3% – of cloud cover, you will get warming that is several degrees Celsius different.” This sensitivity underscores the critical need for more accurate cloud representation, yet directly simulating clouds at their natural scale – meters across, shaped by tiny air currents – would demand computing resources exceeding anything available today, or likely to be available within the foreseeable future.
For years, climate scientists have relied on “workarounds,” adding non-physical parameters to existing fluid dynamics equations (specifically, the Navier-Stokes equations) to indirectly capture the effects of clouds. These parameters are essentially educated guesses, tweaked until the model produces results that align with historical data. However, this process is inherently subjective, relying heavily on intuition and limited data. The establishment of the Climate Modeling Alliance (CLIMA) in 2019, led by Tapio Schneider at the California Institute of Technology, aimed to automate this parameter selection using machine learning. But even with advanced algorithms, the need for comprehensive cloud data remained a significant bottleneck. Obtaining sufficient real-world cloud observations through aircraft and ground-based instruments is expensive and logistically challenging.
This piece references the quantamagazine.org report.
To overcome this data scarcity, researchers turned to high-resolution simulations known as large-eddy simulations (LES). While computationally intensive, LES provides a detailed, albeit limited, view of cloud turbulence. However, even generating a handful of these simulations proved costly. A breakthrough came through a collaboration with Sheide Chammas and his team at Google, who developed a highly efficient LES algorithm running on custom tensor processing units. This allowed them to generate a library of over 8,000 digital clouds, representing diverse conditions across the Pacific Ocean and all four seasons. Schneider hailed this library as “game-changing,” providing a wealth of data for training machine learning algorithms. CLIMA has now used this data to configure new cloud parameters, resulting in a model that preliminary testing suggests is twice as accurate as its predecessors.
However, a parallel and increasingly prominent approach is challenging the very foundation of traditional climate modeling. Chris Bretherton, formerly studying clouds through LES techniques and now at the Allen Institute for Artificial Intelligence (Ai2), has become a leading proponent of bypassing the Navier-Stokes equations altogether. Inspired by recent advances in AI-driven weather forecasting, Bretherton’s team is developing neural networks trained directly on historical atmospheric data. This approach, embodied in the Ai2 Climate Emulator version 2 (ACE2), aims to predict future climate states without explicitly modeling the underlying physics. Recent tests show ACE2 can predict global temperatures and precipitation three months ahead with comparable accuracy to physics-based simulations, and at a fraction of the computational cost.
The key difference lies in how these models handle uncertainty. Traditional models strive for a complete, physically accurate representation of the climate system, while AI-driven models focus on identifying patterns and correlations within the data. While physics-based models offer a degree of interpretability – we understand why they predict certain outcomes – AI models are often “black boxes,” making predictions without revealing the underlying reasoning. This raises concerns about long-term reliability. Neural networks excel at reproducing patterns in their training data, but climate change is pushing the Earth system into uncharted territory, where past patterns may no longer hold. As Sarat Sreepathi of Oak Ridge National Laboratory points out, “How much confidence do you have [in predictions] based on physical principles?”
The current trajectory isn’t about choosing one approach over another, but about leveraging the strengths of both. Researchers are exploring hybrid models that combine the predictive power of AI with the physical constraints of traditional simulations. For example, neural networks can be trained on the output of physics-based models, accelerating calculations and improving efficiency. The ultimate goal isn’t simply to predict the climate with greater accuracy, but to better understand the range of possible futures and the factors that influence them. The speed and efficiency of AI-driven models are particularly valuable for assessing climate risks and informing policy decisions.
The implications of these advancements extend beyond the scientific community. While the average person won’t directly interact with these complex models, improvements in climate forecasting will ultimately impact everything from agricultural planning to infrastructure development. However, the most crucial question remains unanswered: will this enhanced understanding translate into meaningful action to mitigate climate change? As Schneider succinctly puts it, “It’s a question that goes beyond just the science of it. It’s hard to predict.” The next few years will be critical in determining whether we can not only predict the future climate, but also shape it for the better. Watch for the results of the CLIMA model unveiling in Japan this March, and more importantly, observe whether the increased precision of these forecasts leads to more ambitious climate policies and a faster transition to a sustainable future.
