Weather forecasting has always been unpredictable — and climate modeling even more so. However, advances in modeling techniques and computing power have greatly improved our ability to anticipate nature’s behavior.
Researchers at the Max Planck Institute unveil a near–kilometer-scale model combining weather and climate simulation.
A Near–Kilometer-Scale Model of Earth’s Systems
The model’s resolution is technically 1.25 kilometers per grid cell — close enough to call it kilometer-scale. The model splits Earth into 336 million surface cells and 336 million atmospheric cells, totaling 672 million points.
For each of these cells, the researchers ran interconnected simulations to represent Earth’s key dynamic systems, dividing them into two groups: “fast” and “slow.”
The “fast” systems encompass the energy and water cycles — essentially, the weather. Capturing these processes accurately requires extremely fine resolution, such as the 1.25-kilometer scale achieved by the new model.
Using the ICON Model for High-Resolution Simulations
To build their simulation, the team employed the ICOsahedral Nonhydrostatic (ICON) model, developed jointly by the German Weather Service and the Max Planck Institute for Meteorology.
Diving into the details of climate modeling helps clarify the ideas behind their work:
The “slow” processes, by contrast, involve the carbon cycle and long-term shifts in the biosphere and ocean chemistry. These operate over years or decades — a stark difference from the few minutes it might take a thunderstorm to pass from one 1.25 km cell to the next.
The true innovation of the study lies in combining these fast and slow systems within a single framework. Traditionally, such comprehensive models could only run feasibly at resolutions above 40 kilometers due to computational limits.
So how did the researchers manage it? Through a blend of advanced software engineering and cutting-edge computing hardware — some of the most powerful chips available.
Modernizing Legacy Fortran Code for Climate Modeling
The foundation of their model was code originally written in Fortran — a programming language that often challenges anyone attempting to modernize legacy systems from before the 1990s.
Over time, the original code had become cluttered with add-ons that made it incompatible with modern computing architectures. To overcome this, the researchers turned to a framework called Data-Centric Parallel Programming (DaCe), which manages data in a way optimized for contemporary hardware.
Meanwhile, science communicator Simon Clark explored whether a climate model could run on much simpler hardware — like a Raspberry Pi:
The researchers ran their simulations on two cutting-edge supercomputers — JUPITER in Germany and Alps in Switzerland — both powered by Nvidia’s new GH200 Grace Hopper chips.
Each GH200 combines a GPU (the “Hopper,” designed for high-speed parallel processing like that used in AI training) with a CPU (the “Grace,” developed by ARM) to divide computing tasks efficiently.
This split allowed the team to assign the “fast” models — such as weather-related processes that update quickly — to the GPUs, while the slower, long-term carbon cycle simulations ran on the CPUs in parallel.
Harnessing Supercomputers to Simulate Earth at Unprecedented Scale
Using 20,480 GH200 superchips, the researchers simulated 145.7 days of Earth in a single day, calculating nearly one trillion variables.
Of course, that also means models of this sophistication won’t be running on local weather stations anytime soon. Such computational resources are rare and often prioritized by tech companies for AI development rather than climate research.
Still, the sheer scale and success of this project are remarkable — a major milestone that hints at a future where ultra-high-resolution climate simulations might become the norm.
Read the original article on: Science Alert