Downscaling: Diffusion model

Left: target precipitation (from a physical model). Right: Estimate from an ML diffusion model

This video shows two versions of the same precipitation field, one from a physical model (left) and one from a machine-learning diffusion model trained on the physical model outputs (right). The aim is to judge the quality of the diffusion model - how well does it reproduce the physical model results.

The precipitation field is a sequence of daily averages. For the physical model that’s a daily average of high-frequency model output - the diffusion model calculates daily averages directly. To make a video, the daily averages are interpolated to hourly values, and then the hourly values are plotted as a sequence of frames.