How to Unlock Local Detail in Coarse Climate Projections with NVIDIA Earth-2

Global climate models are good at the big picture—but local climate extremes, like hurricanes and typhoons, often disappear in the details. Those patterns are still there—you just need the right tools to unlock them in high-resolution climate data.

Using NVIDIA Earth‑2, this blog post shows you how to downscale coarse climate projections into higher-resolution, bias‑corrected fields—revealing local detail not resolved in the raw data. 

Why downscaling climate projections is key to risk assessment

High-resolution projections play a key role in assessing physical climate risk, informing decisions from infrastructure planning to agricultural adaptation. However, running global models at fine resolution is computationally prohibitive—requiring significant compute, storage, and time. Coupled Model Intercomparison Project Phase 6 (CMIP6) provides the most widely used global climate projections—underpinning IPCC reports and sector-specific risk models—but its outputs are often too coarse to capture short-lived weather events, like tropical cyclones or extreme heat. 

Small ensemble sizes compound the problem. When only a few simulations are available, rare but extreme events can be missed, reducing the reliability of tail-risk estimations. Many weather workflows build on ERA5, a reanalysis dataset. CMIP6 outputs often differ in grid, cadence, pressure levels, and variable definitions, creating a gap between available projections and climate risk assessment needs. For researchers and practitioners, closing this gap calls for accelerated computing and AI-driven tools that can bridge coarse climate projections and analysis-ready data.

NVIDIA Earth-2 provides a full-stack platform to build and deploy AI-powered weather and climate applications. NVIDIA CorrDiff—a model within this ecosystem—replaces compute-intensive and time-consuming high-resolution simulations with a generative downscaling model that transforms coarse CMIP6 output into high-resolution fields. It consists of a regression model that predicts the conditional mean, and a diffusion model that predicts the residuals between the regression output and the ground truth. 

CorrDiff performs core climate data transformations—including spatial and temporal downscaling, bias correction, and variable synthesis simultaneously—by training to map from biased climate model outputs to observation-constrained reanalysis, like ERA5. As a model-to-model translation tool, it can generate large ensembles from a single input sample, providing uncertainty estimates that help assess tail risks. 

This blog post walks you through training CorrDiff for CMIP6 downscaling step by step, including dataset selection and configuration. 

It also shows how S&P Global Energy, a leader in climate hazard assessment, is developing this workflow with NVIDIA to produce data grounded in observation for use in risk analytics. 

Input and target datasets for CorrDiff training

Training CorrDiff requires paired low-resolution input and high-resolution target datasets covering the same time period and weather evolution. You’ll need an observation-assimilated climate model run for training. Free-running climate simulations don’t align with observed patterns. 

Using CanESM5 assimilated hindcasts as input

This guide uses data from the CanESM5 climate model. The output is on a 64×128 global grid at approximately 2.8° resolution (~300 km at the equator). This captures large-scale climate patterns but is too coarse to resolve the finer-scale features needed for climate risk assessment. 

Select data that balances availability across historical and future scenarios, temporal resolution close to ERA5’s hourly output, and vertical coverage. Be sure to include pressure-level variables to describe the 3D atmospheric state. 

You’ll use daily data from CanESM5’s assimilated hindcast runs (experiment_id: dcppA-assim). These runs use ERA-Interim to maintain CanESM5’s large-scale circulation aligned with observed weather. Because CanESM5 provides daily data and ERA5 provides hourly data, use three consecutive daily timesteps—the day before, the day of, and the day after the target hour as input.

Surface inputs include temperature, humidity, precipitation, radiation fluxes, wind, cloud cover, and cryosphere variables. Pressure-level inputs (ta, ua, va, zg, hus, wap) are provided at 1000, 850, 700, 500, 250, 100, 50, and 10 hPa.

ERA5 reanalysis as the high-resolution target

The target dataset is ERA5, which provides hourly data at 0.25° resolution (~31 km at the equator) on a 721×1440 grid. With comprehensive coverage of variables, it is a standard target for many AI weather models, including FourCastNet and HENS. Given an input resolution of ~2.8° and a target resolution of ~0.25°, this achieves approximately 11x super-resolution per spatial dimension. 

On the output side, use the same variables as HENS plus sea surface temperature (SST). SST missing values over land are filled using a smoothed nearest-neighbor interpolation to avoid missing values and artificial gradients at coastlines. 

Additional inputs and preprocessing 

Beyond CanESM5 variables, it’s important to include temporal and geographic context channels (Table 1). 

Trigonometric encodings provide continuous spherical coordinates. The hour-of-day channel guides temporal weighting. For example, midnight predictions rely more on “yesterday” while hour 23 draws more from “tomorrow.” 

Apply three preprocessing steps:  

1. Merge snow cover and sea ice concentration into a single variable to avoid missing values over land.  
2. Normalize each channel using z-scores for stable training.
3. Upsample bilinearly to the ERA5 grid, as CorrDiff requires matched input and output grids.

CanESM5 and the ERA5 core dataset have 38 years of overlap. Combined with 10 ensemble members from CanESM5, this yields approximately 138,700 training samples—well above the recommended 50,000 minimum. The final input comprises 231 channels: 222 from CanESM5 (74 variables × 3 timesteps), 2 time-dependent, and 7 static. 

How to train CorrDiff for climate downscaling

Training CorrDiff involves five steps: data loading, model configuration, regression training, regression evaluation, and diffusion training. The full pipeline takes tens of GPU-hours for small datasets or ~2,000 GPU-hours for larger ones, like in this guide.

1. Data loading
   1. Build your custom dataloader using the NVIDIA PhysicsNeMo DownscalingDataset as a template.
   2. Keep per-sample work lightweight.
   3. Precompute dataset-wide statistics (e.g., climatologies).
   4. If data loading becomes a bottleneck, consider optimizing or switching to a faster loader.
2. Configure the model and datasets
   1. Edit Hydra-based YAML configuration files as described in the documentation to define datasets, model architecture, training hyperparameters, and data splits.
   2. If required, override settings at runtime using Hydra ++ syntax.
3. Training the regression model
   1. Run python train.py with the regression config and train a UNet to predict the conditional mean of the mapping.
   2. Track validation loss for early stopping to avoid overfitting.
   3. For scaling tips and faster training, see the CorrDiff training performance optimization post.
4. Evaluating the regression model
   1. Run the trained regression model on the validation set.
   2. Compute the standard deviation of prediction errors per output channel.
   3. Use these to set sigma_data in the diffusion model’s loss function.
   4. This balances optimization across variables with different error scales.
5. Training the diffusion model
   1. Run python train.py with the diffusion config, using sigma_data from the regression evaluation step.
   2. This stage learns to add realistic fine-scale variation.

Running inference and evaluating downscaled output for CanESM5 SSP585

Use NVIDIA Earth2Studio—an open-source Python package for running AI weather and climate models—for inference. It includes a CorrDiff wrapper and a CMIP6 dataset that downloads from ESGF servers automatically. The package is easily extensible with custom sources and wrappers. 

Running downscaling for CanESM5 SSP585 is straightforward: 

import torch
import numpy as np
import xarray as xr
from earth2studio.models.dx import CorrDiffCMIP6
from earth2studio.data import CMIP6MultiRealm, CMIP6, fetch_data

# Load model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = CorrDiffCMIP6.load_model(
    CorrDiffCMIP6.load_default_package(), 
    output_lead_times=np.array([np.timedelta64(-12, "h"), np.timedelta64(-6, "h")])
)
model.seed = 1 # Set seed for reproducibility
model.number_of_samples = 1 # Modify number of samples if needed
model = model.to(device)
# Build CMIP6 multi-realm data source, about 60 Gbs of data will be fetched
cmip6_kwargs = dict(experiment_id="ssp585", source_id="CanESM5", variant_label="r1i1p2f1", exact_time_match=True)
data = CMIP6MultiRealm([CMIP6(table_id=t, **cmip6_kwargs) for t in ("day", "Eday", "SIday")])
x, coords = fetch_data(
    source=data,
    time=np.array([np.datetime64("2037-09-06T12:00")]), # CMIP daily data provided at 12:00 UTC
    lead_time=model.input_coords()["lead_time"],
    variable=model.input_coords()["variable"],
    device=device,
)
# Run model forward pass
out, out_coords = model(x, coords)
da = xr.DataArray(data=out.cpu().numpy(), coords=out_coords, dims=list(out_coords.keys()))

This code loads the CorrDiff model and the CMIP6 datasets (atmosphere and sea ice realms), configures the experiment and variant IDs, and runs the downscaling workflow for the specified timestamps. Results are returned as xarray DataArray.

To learn more about Earth2Studio, sign up for the online course Applying AI Weather Models with NVIDIA Earth-2 from the NVIDIA Deep Learning Institute. 

Figure 1 shows the model’s capabilities on a single timestamp. The left panel shows raw CMIP6 10 m wind speed at ~2.8° resolution. The right panel shows CorrDiff output at ~0.25° resolution, revealing a hurricane in the Caribbean and a typhoon in the Pacific—neither is resolved in the coarse input. 

CorrDiff recovers these systems by detecting subtle large-scale fingerprints that correlate with cyclone presence. While ERA5 has known limitations for tropical cyclones, this result shows how fine-scale signals can be extracted from coarse climate model output. 

You can also evaluate the model quantitatively. For the 2010 test set, outputs are downscaled four times daily (00, 06, 12, 18 UTC), averaged into daily means, and compared against bilinearly interpolated CMIP6 as a baseline. Table 2 summarizes results for near-surface air temperature (T2m) and 10m wind components (U10m, V10m). 

CorrDiff outperforms the baseline on Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), with strong temperature improvements. The T2m bias drops from nearly 1 K to -0.11 K. For wind components, the baseline shows a lower bias, but this reflects the minimal wind bias in CanESM5’s assimilated runs rather than a limitation of CorrDiff. 

CorrDiff can generate ensembles of plausible outputs from a single input sample. Table 3 reports hourly metrics using an 8-member ensemble, including the Continuous Ranked Probability Score (CRPS), which generalizes MAE to probabilistic forecasts. CRPS rewards accuracy and well-calibrated spread—lower values indicate better performance. Values below the deterministic MAE show that the ensemble spread reflects uncertainty, not noise. 

These quantitative evaluations use historical data where ERA5 provides ground truth. To test generalization, apply the model to SSP585 projections (r1i1p1f1), which extend beyond the training period to 2100. 

Follow the same protocol: downscale four times daily, aggregate to daily means, then compute global annual averages. Figure 2 shows the time series for T2m, U10m, and V10m.

During the historical period, the pattern matches the 2010 findings. CanESM5 shows a warm bias, while CorrDiff output aligns closely with ERA5. In the SSP585 projection, the CorrDiff correction remains stable, suggesting the learned bias adjustment holds in future conditions. However, variability in the correction increases as the projections move further from the training period, especially for wind components.

While encouraging, the results warrant caution. The model hasn’t been trained on future climate data, and some distribution shifts are inevitable. For applications requiring high confidence in future projections, additional validation strategies—such as rolling-window cross-validation—can help quantify extrapolation limits. 

Applications, takeaways, and next steps for climate downscaling

The next question is how these high-resolution fields— especially large, correlated ensembles—translate into downstream decision workflows. S&P Global Energy is using CorrDiff and FourCastNet models to generate large sets of climate ensembles for probabilistic portfolio-level impact and resilience analysis. These large ensembles enable the team to represent ranges and structures of plausible future climate states with a set of initial conditions, including rare but high-impact extremes. 

With access to hundreds or thousands of realizations, S&P Global Energy can better define and quantify variability, tail behavior, and joint risks across assets—critical for modeling nonlinear, highly correlated climate impacts. They are developing a scalable capability for probabilistic, portfolio-level risk analysis using very large ensembles. The ensembles serve as inputs to internal impact functions, translating climate variables into losses such as damage to buildings and infrastructure, disruption of energy systems or transportation networks, and stress on community supply chains supporting essential services. 

With this technology, it is possible to improve resilience and turn climate risk into actionable insight. While this work is still in active development, these technologies can help organizations and decision-makers better understand, prepare for, and respond to climate risks.

Get started

This guide covered how CorrDiff downscales coarse climate model output by combining 11x spatial super-resolution, daily-to-hourly temporal downscaling, variable synthesis, and bias correction in a single workflow. 

 You can now apply the same approach to your own climate downscaling projects. Get started: 

1. Run inference using Earth2Studio to downscale CMIP6 scenarios with the pretrained CorrDiff model.
2. Train your own model with PhysicsNeMo to customize CorrDiff for your specific datasets and needs.
3. Watch the tutorial video Get Started with NVIDIA Earth-2 in 5 Minutes.