Global climate models (GCMs) are our primary tools for understanding the future of our planet, yet they face a fundamental physical limitation: the resolution gap. While a GCM might divide the Earth into a grid of cells spanning 50 to 100 kilometers, the processes that actually drive the climate—such as the formation of a single cloud, the turbulence of a gust of wind, or the collision of microscopic water droplets—occur at scales of meters or even millimeters.
As Randall et al. et al. explain in Confronting Models with Data: The GEWEX Cloud Systems Study, this scale mismatch forces researchers to rely on parameterizations. These are simplified mathematical "shortcuts" that represent the collective effect of small-scale (sub-grid) processes on the large-scale grid. Without them, models cannot simulate rain, heat exchange, or cloud cover.
The Problem with Heuristics
Traditionally, parameterizations have been built using heuristic formulas—simplified physical rules of thumb. For example, autoconversion parameterization is used to estimate how quickly small cloud droplets merge into raindrops. As Bhowmik et al. et al. note in their work on Indian Summer Monsoon rainfall simulation, these formulas are essential because models cannot resolve individual droplets, yet they are a primary source of uncertainty in predicting major weather patterns like monsoons.
The danger of these simplifications is that they often fail to capture the non-linear, complex nature of the atmosphere. Hohenegger et al. et al. demonstrated in Towards climate simulations at cloud-resolving scales that traditional models can even get the sign of physical feedbacks wrong; while coarse models predicted a positive feedback between soil moisture and rain, high-resolution simulations that resolved convection showed a negative one.
Embracing Randomness and Complexity
One way to improve these approximations is to stop treating them as purely deterministic. In Stochastic climate theory and modeling, Fogt et al. et al. argue that adding "stochasticity" (mathematical randomness) is not just a patch for missing data, but a theoretical necessity. By representing sub-grid processes as random fluctuations, models can better capture the natural variability of the weather and fix systematic biases that lead to "blurred" or overly stable climate states.
This complexity is particularly evident in how clouds organize themselves. Caroline Muller describes in Spontaneous Aggregation of Convective Storms how tropical clouds spontaneously "clump" together. Because this clumping is highly sensitive to microphysical parameterizations, small errors in how we represent a single cloud can lead to massive discrepancies in how we predict global climate sensitivity.
The Machine Learning Revolution
The most recent shift in the field is the move toward hybrid modeling, where machine learning (ML) replaces or enhances traditional physics-based parameterizations. By training neural networks on high-fidelity, "cloud-resolving" data, researchers can create emulators that are both fast and physically accurate.
- Data-Driven Closures: Giles et al. et al. showed in Embedding machine-learnt sub-grid variability improves climate model precipitation patterns that using Gaussian processes to inform coarse models about sub-grid fluctuations can reduce tropical precipitation errors by 20%.
- Neural GCMs: In Neural general circulation models for weather and climate, Kochkov et al. et al. introduced a model that integrates ML directly into the atmospheric equations, achieving the accuracy of state-of-the-art systems with massive computational savings.
- Stability Challenges: However, this transition is not seamless. Jakhar et al. et al. point out in Learning Closed‐Form Equations for Subgrid‐Scale Closures that while ML can discover highly accurate equations for sub-grid fluxes, these equations often lead to numerical instability when actually run in a simulation.
Why Resolution Isn't a Silver Bullet
It is tempting to think that as supercomputers get faster, we can simply shrink the grid until parameterization is no longer needed. However, Schneider et al. et al. argue in Harnessing AI and computing to advance climate modelling and prediction that this is a "computational dead end."
Even at a 1-kilometer resolution—which is currently the frontier of global modeling—fundamental processes like cloud microphysics and turbulence still occur at scales thousands of times smaller (millimeters). Consequently, the need to parameterize and calibrate unresolved physics remains a permanent feature of climate science. The future of the field lies not just in smaller grids, but in smarter, data-informed ways to bridge the gap between the microscopic and the global.
Key Foundational Work
Confronting Models with Data: The GEWEX Cloud Systems Study (2003) — contributed 2 concept elements to this direction.