Year: 2025
Type: dissertation
Abstract: Diffusion models are a type of deep learning method applicable for generating high-quality image data. They have also been proven suitable for super-resolution tasks, where low-resolution inputs are transformed to high-resolution outputs. Diffusion models are trained to generate samples from Gaussian noise by learning to reverse a noising, or diffusion, process. Flow matching is another technique related to diffusion, where the data is linearly transformed from the input distribution to achieve the target sample. Climate downscaling refers to a model aiming to improve the resolution of coarse climate data. These models can be used to gain more detailed information from coarse climate inputs, such as predictions of future climate scenarios. In this thesis, a climate downscaling method was implemented using both diffusion models and flow matching. For temperature data, both methods proved efficient, with flow matching achieving the lowest error. For more complex atmospheric water data, the initial experiments were also comparable to baseline solutions. Additionally, suitable physical constraints were found for both models, resulting in moderate improvements in the error. These generative models show potential for being applied to larger-scale climate downscaling.
Source: LUTPub (LUT University)
Author: Toropainen, Helmi
Language: English
Subfield: Atmospheric Science
Field: Earth and Planetary Sciences
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7029062557