The central challenge of climate science is a mismatch of scales. While we want to predict the state of the entire planet over decades, the most important drivers of climate—clouds, thunderstorms, and ocean eddies—happen at scales of a few kilometers or less.
Traditional Earth System Models (ESMs) divide the world into a grid, but computational limits usually restrict these grid cells to roughly 100 kilometers wide. Because a single 100km box might contain thousands of individual clouds, the model cannot "see" them. Instead, it must guess their collective effect using simplified mathematical shortcuts called parameterizations. As David A. Randall and colleagues noted in their work on the GEWEX Cloud System Study, these approximations are the single largest source of uncertainty in our climate predictions.
The Failure of Mathematical Shortcuts
Parameterizations are essentially "if-then" rules: if a grid cell reaches a certain humidity, the model assumes a certain amount of rain falls. However, these rules are often too simple to capture the chaotic reality of the atmosphere.
For example, coarse models often suffer from "drizzling" too much—they predict frequent light rain instead of the intense, localized bursts of real-world storms. Afolayan Akinsanola Akinsanola and Rao Kotamarthi have shown that when models are allowed to "permit" convection rather than approximate it, the accuracy of precipitation timing and extreme weather patterns improves dramatically.
Without this resolution, models struggle with the "cold trap" in the tropical troposphere that regulates water vapor entering the stratosphere. As Paulo Ceppi and others found, failing to resolve these small-scale convective processes leads to unreliable projections of how the atmosphere will respond to warming.
The Kilometer-Scale Revolution
The most direct solution is to shrink the grid. By moving to "kilometer-scale" or "storm-resolving" models, researchers can explicitly simulate the internal "engines" of storms.
Recent breakthroughs, such as the work by Oliver Fuhrer and Torsten Hoefler, have demonstrated that near-global simulations at 1 km resolution are now computationally feasible on massive GPU clusters. This shift allows models to resolve the texture and timing of thunderstorms naturally, rather than relying on hand-tuned formulas.
This vision is being formalized through initiatives like Earth Virtualization Engines (EVEs). As Björn Stevens and Andreas F. Prein argue, EVEs represent a technical shift toward high-fidelity, interactive climate data that can resolve the physical processes previously hidden within the subgrid.
Hybridizing Physics with Machine Learning
While 1km models are the gold standard, they are too expensive for the multi-century simulations needed to study long-term climate change. This has led to a new paradigm: Neural Earth System Modelling (NESYM).
The idea, proposed by Christopher Irrgang and Niklas Boers, is to use machine learning to "learn" the behavior of high-resolution simulations and then plug those learned patterns into fast, low-resolution models. This works through several mechanisms:
- Nudging and Correction: Researchers like Anna Kwa and Noah Brenowitz use neural networks to "nudge" low-resolution models toward the accuracy of high-resolution ones, reducing errors in temperature and rain by up to 25%.
- Generative Downscaling: Using tools like Generative Adversarial Networks (GANs) or diffusion models, researchers can restore realistic spatial patterns to patchy rainfall data that coarse models typically "smear" out.
- Stochastic Emulation: Because weather is chaotic, a single "correct" answer for a subgrid process doesn't exist. Balu Nadiga has shown that treating subgrid physics as a probabilistic distribution using GANs better captures the natural variability of the atmosphere.
The Stability Challenge
The transition to AI-driven climate modeling isn't seamless. A persistent problem is that an ML model might be very accurate at predicting a single timestep (offline accuracy) but cause the entire climate simulation to "crash" or drift into non-physical states when run for years (online stability).
To solve this, researchers are developing "physics-informed" architectures. For instance, CondensNet, developed by Xin Wang and Kalli Furtado, uses adaptive physical constraints to ensure that water vapor is always conserved, enabling stable 50-year simulations that are hundreds of times faster than traditional high-fidelity models.
As Tom Beucler and others noted in a recent large-scale competition, while the field has mastered offline accuracy, the next frontier is ensuring these hybrid models can reliably simulate the complex, coupled interactions of the Earth system over the long term.