This study introduces a generative diffusion model called CorrDiff, designed to downscale coarse-resolution (25-km) global weather data to higher resolution (2km) over a specific subset of the globe. CorrDiff operates in two main steps: regression and generation. In the regression step, it approximates the mean, while in the generation step, it further refines the mean and generates the distribution, adding fine-scale details stochastically. This methodology resembles the common practice in fluid dynamics of decomposing physical variables into their mean and perturbations.
Context
The adoption of CorrDiff presents significant potential benefits, including accelerated regional forecasts, improved climate downscaling, and the provision of high-resolution regional forecasts in data-scarce regions by leveraging training data from adjacent areas. Notably, CorrDiff demonstrates sample efficiency, effectively learning from just four years of data. Furthermore, running on a single GPU, CorrDiff is at least 22 times faster and 1,300 times more energy-efficient than the numerical model used to produce its high-resolution training data, which typically runs on 928 CPU cores.
Arxiv/Blog/Paper Link
Detailed Description
This study introduces a generative diffusion model called CorrDiff, designed to downscale coarse-resolution (25-km) global weather data to higher resolution (2km) over a specific subset of the globe. CorrDiff operates in two main steps: regression and generation. In the regression step, it approximates the mean, while in the generation step, it further refines the mean and generates the distribution, adding fine-scale details stochastically. This methodology resembles the common practice in fluid dynamics of decomposing physical variables into their mean and perturbations.
Context
The adoption of CorrDiff presents significant potential benefits, including accelerated regional forecasts, improved climate downscaling, and the provision of high-resolution regional forecasts in data-scarce regions by leveraging training data from adjacent areas. Notably, CorrDiff demonstrates sample efficiency, effectively learning from just four years of data. Furthermore, running on a single GPU, CorrDiff is at least 22 times faster and 1,300 times more energy-efficient than the numerical model used to produce its high-resolution training data, which typically runs on 928 CPU cores.
The code for CorrDiff is available on GitHub at NVIDIA/modulus/corrdiff.