For decades, global climate models have operated like low-resolution cameras. With grid cells typically spanning 100 kilometers, they can capture the broad strokes of planetary circulation but are "blind" to the fundamental engines of the atmosphere: individual clouds and convective storms. To compensate, scientists have relied on parameterizations—mathematical shortcuts that use statistical averages to guess what is happening inside a grid box.
A new generation of research is moving past these approximations. By pushing horizontal resolutions to the kilometer scale (typically 1–5 km), researchers are developing convection-permitting models (CPMs). These models don't just see the weather more clearly; they simulate the physics of the atmosphere from first principles.
From Statistical Guesses to Explicit Physics
The primary motivation for this shift is the failure of traditional parameterizations to capture the "texture" of the atmosphere. As David A. Randall and colleagues noted in their work on the GEWEX Cloud Systems Study, clouds are the single largest source of uncertainty in climate prediction.
In a coarse model, a thunderstorm is represented by a single number representing "average upward motion." In a kilometer-scale model, the storm is resolved as a dynamic fluid system. This transition to explicit convection modeling allows the simulation to solve the governing equations of cloud dynamics directly. As Cathy Hohenegger and Christoph Schär demonstrated in Towards climate simulations at cloud-resolving scales, this isn't just a quantitative improvement—it can fundamentally change the model's conclusions. They found that while coarse models predicted a positive feedback between soil moisture and rain, high-resolution simulations revealed a negative one, correcting a long-standing bias in how we predict local climate impacts.
Resolving the "Gears" of Extreme Weather
The intuition behind kilometer-scale modeling is that by resolving the "gears" of the atmosphere, the larger "machine" behaves more realistically. This is particularly evident in three areas:
- Tropical Cyclones: Standard datasets like ERA5 often "smooth over" the intense inner cores of hurricanes. Lara Tobias-Tarsh and Rao Kotamarthi showed that convection-permitting simulations can capture Category 3+ hurricanes that are entirely missing in coarser models, providing a more accurate "upper tail" for risk assessment.
- Complex Topography: Coarse models treat mountain ranges like smooth bumps. Afolayan Akinsanola Akinsanola and Jiali Wang have shown that 4-km resolution allows for the direct representation of orographic lift, where air is physically forced upward by mountains to create localized precipitation. This drastically improves precipitation timing and intensity across regions like the Himalayas and the Rockies.
- Storm Lifecycle: Coarse models often struggle with the "timing" of rain, making it rain too early or too often. By resolving storm-scale physics, CPMs can maintain the lifecycle of convective systems as they propagate across plains, capturing the "rain bombs" and organized convection that drive flash floods.
The Computational Leap: Exascale and AI
Running a global model at 1 km resolution is a massive computational challenge, requiring roughly 1,000 times more power than a 100-km model. This has led to a "journey" of engineering, as described by Xiaohui Duan and Yuxuan Li, who successfully ported the Community Earth System Model to 40 million cores. Similarly, Oliver Fuhrer and Torsten Hoefler established a performance baseline using thousands of GPUs to make these simulations feasible.
Parallel to this, AI is emerging as a way to bypass the brute-force cost of numerical simulation. Foundation models like Aurora, described by Ana Lučić, and FengWu-GHR by Tao Han, are being trained on high-resolution data to learn the "physical signatures" of weather. These models can produce kilometer-scale forecasts at a fraction of the energy cost of traditional supercomputers.
Earth Virtualization Engines
The ultimate goal of this research direction is the creation of Earth Virtualization Engines (EVE). As proposed by Björn Stevens and Andreas F. Prein, EVE represents a shift toward interactive, high-fidelity climate data.
By resolving atmospheric convection and ocean eddies from first principles, these "engines" aim to provide actionable information for local-scale adaptation. Instead of a vague global average, a policymaker could see how a specific city's high-rise canopy interacts with a heatwave, using models like the multi-layer coupling developed by Robert Schoetter and Valéry Masson. This transition from "discovery-oriented" to "service-oriented" science (Guy Brasseur et al., Climate science for 2050) ensures that climate models are no longer just academic tools, but precise instruments for navigating a warming world.