NVIDIA Earth-2 enables researchers and climate scientists to unlock local climate details from coarse global climate model outputs using AI-powered downscaling. While global models like CMIP6 provide essential insights into long-term climate trends, they often lack the resolution to capture critical local phenomena such as hurricanes, typhoons, and extreme heat events. These fine-scale features are vital for accurate climate risk assessment in sectors like infrastructure, agriculture, and energy. To address this gap, NVIDIA has developed CorrDiff, a generative AI model within the Earth-2 platform that transforms low-resolution CMIP6 data into high-resolution, bias-corrected fields. CorrDiff performs multiple tasks simultaneously—spatial and temporal downscaling, bias correction, and variable synthesis—by learning to map from biased climate model outputs to observation-constrained reanalysis data like ERA5. The training process uses paired datasets: CanESM5 assimilated hindcasts as input and ERA5 reanalysis as the high-resolution target. CanESM5 provides daily data at approximately 2.8° resolution (~300 km), while ERA5 offers hourly data at 0.25° (~31 km), enabling about 11x super-resolution per dimension. The input includes surface and pressure-level variables across three consecutive days, along with contextual features such as solar zenith angle, hour of day, geopotential height, distance to ocean, land-sea mask, and trigonometric encodings of longitude and latitude. Preprocessing steps include merging snow and sea ice into a single variable, normalizing data using z-scores, and upsampling to match the ERA5 grid. With 38 years of overlap and 10 ensemble members, the dataset yields over 138,000 training samples—sufficient for robust model training. Training CorrDiff involves five key steps: data loading, model configuration, regression training, regression evaluation, and diffusion training. The regression component uses a UNet to predict the mean of the high-resolution output, while the diffusion model adds realistic fine-scale variability. The model is trained to minimize error across all variables, with sigma_data set based on regression evaluation. Once trained, CorrDiff can generate large ensembles from a single input, providing uncertainty estimates crucial for assessing tail risks. Using NVIDIA Earth2Studio, users can run inference with minimal code. For example, downscaled outputs for CanESM5 SSP585 scenarios reveal tropical cyclones in the Caribbean and Pacific—features absent in the original coarse data. Quantitative evaluations show CorrDiff significantly improves accuracy over traditional interpolation methods. For near-surface temperature, bias drops from nearly 1 K to -0.11 K, and RMSE decreases from 3.19 K to 1.55 K. Wind components also show improved performance, with lower RMSE and better-calibrated ensemble spread as measured by CRPS. When applied to future projections, CorrDiff maintains consistent bias correction through 2100, though variability in corrections increases over time, highlighting the need for caution in long-term extrapolation. Techniques like rolling-window validation can help assess reliability. Organizations like S&P Global Energy are already using CorrDiff to generate large ensembles for portfolio-level climate risk analysis. These ensembles enable probabilistic modeling of extreme events and their cascading impacts on infrastructure, supply chains, and energy systems—turning climate projections into actionable resilience strategies. To get started, users can leverage Earth2Studio to run pre-trained CorrDiff models on CMIP6 data, or train custom versions using PhysicsNeMo. The platform offers a scalable, open-source workflow for integrating AI-driven climate downscaling into real-world decision-making.