Ranking the World’s Climate Models for South America: Four Models Rise Above the Rest

Climate models are the instruments through which humanity peers into its own future, and a new study has now graded those instruments against one of the most climatically complex continents on Earth. In research published in Theoretical and Applied Climatology, a team of Brazilian and international scientists led by Maria Leidinice da Silva of the International Centre for Theoretical Physics evaluated seventeen global climate models from the sixth phase of the Coupled Model Intercomparison Project, known as CMIP6, to determine how faithfully each one reproduces the present-day climate of South America. The verdict is nuanced but consequential: model skill varies dramatically from region to region and from variable to variable, and a handful of models—CNRM-CM6-1, GFDL-ESM4, MPI-ESM1-2-HR and NorESM2-MM—consistently outperform their peers across the continent.

The stakes of this kind of evaluation are far higher than an academic exercise in scorekeeping. Global climate models are the engines that generate the projections used by the Intergovernmental Panel on Climate Change, and they also supply the boundary conditions that drive regional climate models in so-called dynamical downscaling. If a global model misrepresents the circulation patterns, moisture transport or precipitation regimes of a particular region, every downstream projection inherits those errors. For South America, a continent that contains the Amazon rainforest, the vast wetlands of the La Plata Basin, the semiarid interior of northeastern Brazil and the dramatic seasonal reversal of the South American Monsoon, the choice of which model to trust is not a technical footnote. It shapes estimates of future water availability, agricultural viability, forest resilience and flood risk for hundreds of millions of people.

The research team framed their assessment around the historical period from 1979 to 2014, comparing each model’s simulations against ERA5, the state-of-the-art atmospheric reanalysis produced by the European Centre for Medium-Range Weather Forecasts. Reanalyses blend observations with numerical weather prediction to create a physically consistent record of the past climate, and ERA5 has become a standard benchmark for model evaluation worldwide. By measuring how closely each CMIP6 model tracks ERA5 across multiple atmospheric variables and vertical levels, the researchers could quantify not just whether a model gets surface temperature roughly right, but whether it captures the three-dimensional structure of the atmosphere—the winds aloft, the geopotential heights, the vertical distribution of moisture—that governs how weather systems actually form and evolve over the continent.

What makes the study genuinely distinctive is its multilevel design. Rather than confining the analysis to a single variable at the surface, the authors examined the atmosphere at different heights, recognizing that a model can simulate plausible surface conditions while simultaneously botching the upper-level circulation that steers storms and organizes convection. This vertical dimension matters enormously for South America, where the Bolivian High, an anticyclone that forms aloft during the austral summer, is intimately linked to the monsoon system and to deep convection over the Amazon. A model that fails to reproduce upper-level circulation may still appear competent in surface statistics while getting the fundamental dynamics wrong—a trap that single-variable evaluations routinely miss.

To make the comparison tractable, the researchers divided South America into five subdomains that capture the continent’s principal climatic regimes: the northern Amazon Basin, the southern Amazon Basin, the Northeast region of Brazil, the South American Monsoon domain, and the La Plata Basin. Each of these regions poses its own modeling challenge. The Amazon tests how well a model handles the coupling between forest, soil moisture and deep convection in a regime where precipitation is largely recycled from evapotranspiration. Northeast Brazil, with its semiarid climate punctuated by extreme droughts, probes the representation of the intertropical convergence zone and its seasonal migrations. The monsoon domain demands that a model capture a full seasonal cycle of onset, maturity and demise, while the La Plata Basin, one of the world’s great agricultural heartlands, is notorious for organized mesoscale convective systems that produce some of the most intense rainfall events on the planet.

Across these regions and variables, the team deployed a battery of statistical metrics—measures of bias, spatial pattern correlation, and the ability to reproduce observed variability—and then synthesized them through a comprehensive rating index, or CRI. This framework condenses many separate performance measures into a single score for each model in each region, allowing a consistent ranking that would otherwise be impossible when different metrics point in different directions. The approach addresses a long-standing problem in climate science: model evaluation often produces contradictory conclusions depending on which metric an analyst happens to emphasize, and a transparent, integrated index reduces the subjectivity inherent in model selection. It also echoes a broader movement in the field, reflected in recent benchmarking literature, toward systematic, multi-metric assessment rather than ad hoc comparisons.

The results reveal a striking geographic pattern. Model performance diverges most strongly in the tropical subdomains, where the interplay of convection, land surface feedbacks and large-scale circulation creates a punishing test for coarse-resolution global models. In the subtropical regions, by contrast, the models show considerably higher consistency with one another and with the reanalysis, suggesting that the large-scale dynamics there are more robustly captured. This dispersion in the tropics is a familiar refrain in climate modeling—convection remains one of the hardest processes to represent because it occurs at scales far smaller than a typical model grid cell and must be approximated through parameterizations—but the study quantifies just how much this limitation separates individual models from one another over South America.

At the top of the rankings, four models distinguish themselves. CNRM-CM6-1, developed by the French research community; GFDL-ESM4, from the United States Geophysical Fluid Dynamics Laboratory; MPI-ESM1-2-HR, the high-resolution configuration of the Max Planck Institute Earth System Model; and NorESM2-MM, the medium-resolution variant of the Norwegian Earth System Model, all deliver superior performance across multiple variables and multiple regions. Their common thread is instructive: several of these models incorporate relatively sophisticated atmospheric components or finer horizontal resolution, which helps them resolve the terrain, coastlines and circulation features that shape South American climate. At the other end of the spectrum, INM-CM4-8, KIOST and NESM3 show persistent limitations, particularly in simulating upper-level circulation and the precipitation processes that depend on it. The authors emphasize that these weaker models are not without value in a global context, but their specific failures over South America make them poor candidates for regional applications on this continent.

Perhaps the most practically important finding is that the best-ranked models also reproduce the observed monthly precipitation climatology and its variability more faithfully than their lower-ranked counterparts. Precipitation is the variable that matters most for water resources, hydropower and agriculture in South America, and it is notoriously difficult to simulate because it depends on the delicate interaction of moisture convergence, atmospheric stability and convection. The fact that the comprehensive rating index successfully identifies models that excel specifically in precipitation suggests that the framework captures something real about model quality rather than merely rewarding statistical conformity in easier variables. The authors explicitly position their ranking as a foundation for selecting boundary conditions for dynamical downscaling over South America, giving regional modeling groups a defensible, evidence-based starting point for their experiments.

The broader lesson of the study is one that climate scientists have been articulating with increasing urgency: there is no universally best climate model, and ensembles should not be treated as interchangeable collections of equals. A model that performs admirably over Europe or East Asia may stumble over the Amazon, and a model that excels at surface temperature may misrepresent the winds five kilometers above the surface. The authors’ central conclusion—that models should be chosen based on their demonstrated performance in the region of interest—represents a shift from the traditional practice of using all available models indiscriminately toward a more discriminating, regionally informed selection strategy. For a continent whose climate future hinges on the fate of the Amazon forest, the reliability of the monsoon and the behavior of the La Plata Basin’s rivers, knowing which models deserve a seat at the table is not a luxury. It is the difference between projections that inform sound adaptation policy and projections that quietly mislead it.

Subject of Research: Evaluation of CMIP6 global climate model performance over South America

Article Title: Assessment of CMIP6 models’ performance for South America: multilevel perspective

Article References: Assessment of CMIP6 models’ performance for South America: multilevel perspective. (n.d.). https://doi.org/10.1007/s00704-026-06585-1

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06585-1

Keywords: CMIP6, climate models, South America, ERA5, precipitation, Amazon Basin, South American Monsoon, La Plata Basin, model evaluation, dynamical downscaling, reanalysis, climate projections