Abstract:Deterministic regression-based downscaling models for climate variables often suffer from spectral bias, which can be mitigated by generative models like diffusion models. To enable efficient and reliable simulation of extreme weather events, it is crucial to achieve rapid turnaround, dynamical consistency, and accurate spatio-temporal spectral recovery. We propose an efficient correction diffusion model, TAUDiff, that combines a deterministic spatio-temporal model for mean field downscaling with a smaller generative diffusion model for recovering the fine-scale stochastic features. We demonstrate the efficacy of this approach on downscaling atmospheric wind velocity fields obtained from coarse GCM simulations. Our approach can not only ensure quicker simulation of extreme events but also reduce overall carbon footprint due to low inference times.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the statistical downscaling of extreme weather events. Specifically, the existing downscaling methods for climate variables based on deterministic regression models are often affected by spectral bias. Especially when dealing with extreme weather events, these models have difficulty in capturing fine - scale stochastic atmospheric processes. To solve these problems, the author proposes an improved model named TAUDiff. ### Main problems and challenges: 1. **Spectral bias**: Deterministic regression models have low - frequency spectral bias when dealing with fine - scale random features, resulting in inaccurate simulations of extreme weather events. 2. **Dynamic consistency**: Existing downscaling models lack the ability to model time, which may lead to the damage of the dynamic consistency of downscaled data (for example, the distortion of storm front propagation). 3. **Computational efficiency**: Although large - scale generative models can improve the output quality, their high training and inference computational costs limit their large - scale use in practical applications. ### Goals of the TAUDiff model: - **Rapid simulation of extreme events**: Achieve rapid extreme weather event simulation through an efficient downscaling method. - **Dynamic consistency**: Ensure that the time - series characteristics of downscaled data are consistent with historical observational data. - **Accurate spatio - temporal spectral recovery**: Improve the accuracy of spectral feature recovery of downscaled data in space and time. - **Reduce carbon footprint**: Reduce the carbon emissions of the model by reducing the inference time and computational resource requirements. ### Solutions: The TAUDiff model combines two main components: 1. **Mean - field model**: Adopt a video prediction model (such as the SimVP architecture) to ensure mean - field downscaling with dynamic consistency. 2. **Calibrated diffusion model**: Use a smaller guided denoising diffusion model to generate fine - scale random features and make up for the deficiencies of the mean - field model. Through this method, TAUDiff can not only simulate extreme weather events quickly and accurately, but also significantly reduce computational costs and carbon emissions.

TAUDiff: Improving statistical downscaling for extreme weather events using generative diffusion models