The First Century of Storm-Resolving Climate Science

In the basement of Météo-France, in Toulouse, a supercomputer spent months simulating a century of Earth's atmosphere. Not at the blurry 100-kilometer resolution that has defined climate science for decades, but at 2.6 kilometers — fine enough to draw individual thunderstorms, to trace the lift of air over the Pyrenees and the Alps, to see the climate of a small island in the Pacific rather than averaging it into oblivion. When the calculation finished, researchers had produced something unprecedented: 124 years of global climate data at kilometer-scale resolution, spanning from 1976 to 2099.

The result, published by David Saint-Martin, Olivier Geoffroy, and Gildas Dayon, is not merely a technical stunt. It represents something more profound — a shift in what climate scientists can actually see. "This simulation was specifically developed to complement the existing simulations integrated into the latest version of the French climate services," the authors write. But its implications stretch far beyond France's borders.

For most of the history of climate modeling, scientists have faced a fundamental tradeoff. To model the entire globe, you need to simplify. The standard approach, exemplified by the CMIP6 models that inform the IPCC reports, divides Earth into grid cells roughly the size of entire U.S. states — 50 to 100 kilometers across. At that resolution, processes smaller than the grid cell, like the rising plumes of hot air that make thunderstorms, must be approximated using mathematical shortcuts called parameterizations. These parameterizations are one of the biggest sources of uncertainty in climate projections. Nobody knows exactly how to express a thunderstorm as an equation.

Go finer, and this problem changes. At around 5 kilometers and below, a substantial fraction of convective motions — the upward drafts and downward cold pools that organize thunderstorms — can be "explicitly resolved." That means the model calculates them from first principles rather than guessing. Deep convection, the engine of tropical rainfall and hurricanes, stops being an estimate and starts being a result. The same is true for how air flows over mountains, how sea breezes organize along coastlines, how small islands perturb the atmosphere around them. All of these phenomena have always existed in climate models, but at coarse resolution, they were shadows of their real selves. At kilometer-scale, they start to look like the actual Earth.

This has been understood theoretically for years. But running a global model at kilometer resolution is extraordinarily computationally demanding, like trying to track every molecule in an ocean rather than modeling it as a continuous fluid. Until recently, the best anyone had managed was a few years of simulation — enough to test the models, not enough to study climate change over decades.

The Science

The ARP-ECM2 model used in this study is a global atmospheric model developed by French researchers. It draws on a legacy stretching back to the ARPEGE model, which has been the backbone of European weather forecasting for decades. But ARP-GEM2 is something different: a version designed specifically for efficient climate simulation across a range of scales, from the familiar 100-kilometer grids of classical climate models down to a few kilometers. The key innovation is a "semi-implicit, semi-Lagrangian spectral dynamical core" — a mathematical architecture that allows large time steps without sacrificing accuracy, making it practical to run multi-decade simulations at high resolution.

The simulation covered the period 1976 to 2099, though it did so in four segments of roughly 30 years each. The reason is a constraint that highlights both the power and the limitation of this approach: the model does not include an interactive ocean. Instead, the researchers prescribed sea surface temperatures from a lower-resolution coupled model called NorESM-MM, which itself participated in CMIP6. For the historical period (1976-2014), they used observed SSTs passed through that model's simulation; for the future (2015-2099), they used SSTs from a high-emissions scenario (SSP5-8.5). This is computationally convenient — you can split the run into independent pieces — but it means the simulation's large-scale climate evolution is tied to the driving model rather than emerging freely.

The simulation ran on ECMWF's Bull Sequana XH2000 supercomputer. At 2.6-kilometer resolution, one simulated year required about 650,000 CPU-hours and ran at roughly 200 simulated days per day of wall-clock time, using 110 compute nodes. Each 30-year segment took about two months to complete. This might sound expensive, but the authors note that the entire project consumed only about 2 percent of Météo-France's annual computing budget. Such simulations are no longer restricted to resource-intensive frontier experiments.

The choice of NorESM-MM as the driving model was not arbitrary. The researchers tested SSTs from multiple CMIP6 models and found that NorESM-MM's global mean surface temperature matched observations most closely. This matters because ARP-GEM had already been calibrated against observed SSTs, and using a driving model with similar temperature biases minimizes the mismatch between calibration and application.

To evaluate the simulation, the researchers compared it against observational datasets and against the CMIP6 ensemble. They examined six key variables: surface temperature, precipitation, net shortwave and longwave radiation at the top of the atmosphere, total cloud cover, and upper-level wind speed. Root-mean-square errors (RMSEs) were calculated against reference datasets for the period 1985-2014. The results appear in Figure 2, which shows ARP-GEM2's performance alongside the spread of 40 CMIP6 models.

The researchers also examined precipitation patterns in detail across six regions: metropolitan France, the Caribbean (Guadeloupe and Martinique), French Guiana, the Indian Ocean islands (Mayotte and Réunion), New Caledonia, and French Polynesia. These domains span the globe, which is precisely the point of a global simulation: France has territories scattered across every ocean basin, and a single model can represent them all consistently, rather than patching together regional simulations that may use different physics and boundary conditions.

For the future climate analysis, the researchers focused on the period 2064-2099 relative to 1985-2020 — a contrast between a world that has warmed substantially and the present. They compared ARP-GEM2's projected changes against NorESM-MM's to see what the added resolution actually changes when projecting regional precipitation shifts.

What They Found

The headline result is that ARP-GEM2 at 2.6 kilometers ranks among the best-performing models in the CMIP6 ensemble. This is a striking finding. The CMIP6 ensemble includes roughly 40 models, most running at roughly 100-kilometer resolution — 40 times coarser than ARP-GEM2 — and developed over decades by large international teams. That ARP-GEM2 is not merely competitive but among the leaders in error scores suggests that the move to kilometer-scale is not merely a technical exercise; it produces a better product.

Look at the individual variables. For surface temperature, the model's error is well below the CMIP6 median. For top-of-atmosphere longwave radiation — essentially, how much heat the Earth sends back to space — the error is similarly low. For the 200-hPa zonal wind, a measure of the jet stream's strength and position, ARP-GEM2 sits near the 25th percentile of the CMIP6 spread. Precipitation shows a slightly higher error than the best models, a common challenge at kilometer-scale where convective rainfall can be overly vigorous if not carefully calibrated.

But the real story emerges in the regional precipitation maps. At coarse resolution, French Guiana is a blur. The Caribbean islands are merged into a brown smear. The volcanic peaks of Réunion and New Caledonia are invisible. At 2.6 kilometers, these features snap into focus. The model captures the orographic lift of the Andes over French Guiana, producing realistic rainfall gradients from coast to interior. It represents the interaction between the trade winds and the mountainous islands of the Caribbean and Pacific. It resolves the narrow coastal rainbands that form when stable air is lifted over terrain — a phenomenon that shapes ecosystems and water resources but is completely invisible in coarse models.

"The kilometer-scale resolution of ARP-GEM allows it to capture orographic effects as well as island and coastal effects," the authors note. "This capability is particularly evident on small islands in the Caribbean, New Caledonia, French Polynesia, or in the Indian Ocean."

This matters because many of France's overseas territories are small islands where the climate is strongly shaped by local features — a volcanic peak here, a coral reef fringed lagoon there — that bulk up the atmospheric flow. A coarse model sees an island as a single point; a fine model sees it as a landscape with windward and leeward sides, with valleys and ridges, with thermal circulations that drive local rainfall. The difference is not cosmetic. It determines whether local water resource managers, agricultural extension agents, or emergency planners get useful information or useless noise.

The future climate projections reinforce this pattern. Both ARP-GEM2 and the driving NorESM-MM model show the same large-scale fingerprint of warming: the wet regions get wetter, the dry regions get drier — a pattern consistent with basic thermodynamics and visible in virtually all climate models. But the high-resolution simulation reveals structure that the coarse model misses. Over the Caribbean islands, for instance, ARP-GEM2 simulates changes in precipitation that differ from NorESM-MM's broad regional average. The fine model captures how islands will interact with changing large-scale flow, producing local signals that the coarse model cannot anticipate.

"ARP-GEM captures changes over these small islands that are difficult or impossible to anticipate from NorESM-MM alone or even from the CMIP6 multi-model ensemble," the authors write. This is the core claim: that kilometer-scale resolution is not just prettier pictures but genuinely new information about how regional climates will change.

The temperature trajectory (shown in Figure 1) follows NorESM-MM closely, as expected given the prescribed SSTs. The global mean warming reaches 1.5°C above pre-industrial levels during the period 2042-2061, and 3°C during 2074-2093 — two of the major global warming thresholds that policy discussions often reference. Because the simulation spans these periods, it provides sample climates for risk assessment at multiple levels of warming, not just a single future scenario.

Why This Changes Things

Climate services are the unglamorous backbone of climate adaptation. When a city planner designs a drainage system, they need to know how intense future rainfall will become. When an insurance company prices risk for coastal property, they need to know how sea level rise will interact with storm surge. When a Pacific island nation negotiates with international lenders for resilience infrastructure, they need credible projections of how much drier or wetter their climate will become. Climate services produce this information: downscaled projections, tailored datasets, technical guidance that translates scientific output into decision-ready format.

France's DRIAS platform — one of the more mature national climate services — currently relies on a mix of coarse global models and regional downscaling. Regional downscaling, typically run over a limited area at high resolution (10-20 km), adds local detail by nesting a fine mesh inside a coarse global model. This works well for Europe, where France has invested in the infrastructure. But France also has territories in the Caribbean, the Amazon basin, the Indian Ocean, and the South Pacific. Running regional climate simulations for each of these is expensive, operationally complex, and scientifically awkward: each regional simulation has its own physics and boundary conditions, making it hard to compare across domains or to attribute differences to local processes versus model artifacts.

A global kilometer-scale simulation solves the consistency problem. One model, one set of physics, global coverage. Every island is in the same framework as every continent. The deep convection over the Amazon is modeled the same way as the shallow trade-wind cumulus over Tahiti. This internal consistency matters for interpreting results. If a future projection looks different in Tahiti versus metropolitan France, you know it reflects the actual climate physics, not a difference in model configuration.

The computational cost — 2 percent of Météo-France's annual budget — makes this approach feasible for operational climate services, not just research experiments. A decade ago, a simulation like this would have required a top-10 supercomputer and been the subject of a flagship paper. Now it fits within routine service allocation. This is not a one-off; it is a proof of concept for a production capability.

The gains are not only for tropical islands. The simulation captures coastal processes along metropolitan France's shores, including the narrow rainbands that form when winds from the Atlantic encounter the coasts of Brittany or Normandy. It resolves the Alps, the Pyrenees, and the Massif Central in ways that inform high-altitude hydrology and avalanche risk. For a country with both alpine skiing and low-lying coastal cities, understanding mountain snowfall and coastal storm surge at high resolution is not a luxury.

More broadly, the demonstration that kilometer-scale global models can perform comparably to — or better than — the CMIP6 ensemble reframes a long-running debate. For years, the conventional wisdom held that global kilometer-scale models were research tools, too expensive for production climate services, and that regional downscaling was the pragmatic path forward. This study challenges that view. The authors argue that global kilometer-scale simulations offer "added value relative to coarser models" and that their "unified framework promotes simplicity, ease of use, storage efficiency, and reduced development effort."

This has implications well beyond France. Climate services exist in every major economy; many are grappling with the same tension between coarse global coverage and fine local resolution. If global kilometer-scale simulation is feasible at the national level — not just at the handful of institutions that run supercomputers for research — then the entire architecture of climate services could shift. A single global simulation at 2-3 km could be shared among nations, with each country extracting the regional information it needs, rather than each country running its own regional downscaling with all the associated overhead.

"It can also support collaboration among national climate services through the sharing of climate information from global simulations," the authors note. That sentence is quietly revolutionary. It implies a future where the world's climate services are not a patchwork of national efforts with incompatible methods, but a coherent system built on shared, high-resolution, physically consistent global output.

What's Next

No study is complete, and this one honestly lists its limitations. The most consequential is the absence of ocean coupling. The simulation uses prescribed sea surface temperatures from NorESM-MM, which means the model's large-scale climate evolution is locked to that of the driving model. It cannot produce its own El Niño events or Atlantic Multidecadal Variability; it inherits them from the prescribed SSTs. This is a significant constraint for studying climate variability and for projections over the next few decades, where internal variability — the natural swings in ocean temperature that cause decade-to-decade wiggles in global temperature — can be as large as the forced signal from greenhouse gases.

Coupling ARP-GEM2 to a dynamic ocean model is the obvious next step and is flagged as future work. Moon et al. (2025) and others have already demonstrated multi-year coupled simulations at comparable resolutions, so this is a tractable problem. The transition from prescribed SSTs to interactive oceans will likely improve the simulation's fidelity, particularly in regions where ocean-atmosphere coupling is strong: the tropics, the mid-latitude storm tracks, and the high latitudes.

A second limitation is that this is a single simulation, not an ensemble. Climate models are inherently probabilistic; small perturbations in initial conditions or model physics can produce different trajectories, and the spread of an ensemble captures the range of plausible futures. A single simulation cannot quantify this uncertainty. Running multiple ensemble members at kilometer-scale resolution would be dramatically more expensive and is currently impractical. The authors acknowledge this gap and suggest that pattern scaling techniques — methods that estimate regional climate change from global mean temperature — can partially compensate. But the fundamental limitation remains.

A third area for improvement is precipitation, where the model's error is slightly higher than for other variables. This is a recognized challenge in kilometer-scale models, which tend to produce convective rainfall that is too intense or too frequent. Calibration efforts have narrowed the gap, but precipitation remains the variable most sensitive to resolution and convective parameterization.

Looking further ahead, the authors see this work as a proof of concept for a broader transition. "The present study demonstrates the feasibility of this approach and highlights the benefits of this new generation of climate modeling for climate services," they conclude. That is a careful, restrained statement for what is actually a fairly significant claim: that global kilometer-scale simulation is not just possible but ready to be incorporated into the operational toolkit of climate services.

If they are right, the implications cascade. Better regional projections for small islands means better adaptation plans for some of the world's most climate-vulnerable populations. Consistent global coverage means that a farmer in Senegal and a city planner in Vancouver can access projections built on the same underlying physics. Lower computational cost means that more simulations can be run — different scenarios, different emissions pathways, different climate sensitivities — giving decision-makers a fuller picture of the range of futures they must prepare for.

The supercomputer in Toulouse has run its century. Now the question is what the world does with what it learned.