Seminario CSTCP: "Advancements in Machine Learning Techniques for Regional Climate Downscaling", Dr. Erika Coppola (ICTP)

Tipologia evento: 
Data evento
Data inizio evento: 
29/03/2023 - 16:00
Data fine evento: 
29/03/2023 - 17:00
Data pubblicazione evento
Pubblicato il: 

Advancements in Machine Learning Techniques for Regional Climate Downscaling: Improving Accuracy and Efficiency for High-Resolution Climate Projections 

Dr. Erika Coppola  



The latest developments in machine learning methods for regional climate downscaling will be illustrated, which are used to produce high-resolution climate projections. Climate downscaling is an essential tool for understanding the local impacts of global warming at the local and regional scale however, the traditional downscaling approaches have limitations in terms of computational cost and accuracy. Machine learning has emerged as a promising technique to improve the accuracy and efficiency of downscaling and easing the burden of computational costs. The proposed methods leverage deep learning frameworks, such as convolutional, recurrent, and graph neural networks, to emulate the dynamical models used in climate simulations. By using input data from low-resolution simulations and target data from high-resolution observations, these frameworks can generate high-resolution climate projections at a significantly lower computational cost. The input data used in the methods typically include temperature, humidity, wind components, and atmospheric parameters, while the target data include precipitation observations. Recent studies have shown that these machine learning approaches can produce highly realistic and consistent high-resolution temperature and precipitation fields, making them valuable tools for climate projection. Moreover, the use of machine learning can improve the understanding of the relationship between high- and low-resolution simulations, allowing for exploration of future climates and poorly known regions. The development of these methods opens the door to further research, including exploring different machine learning architectures and designing optimal simulation sets to ensure the robustness of the emulators. Ultimately, these methods offer a promising solution for accurate and efficient regional climate downscaling and high-resolution climate projection.


Stanza 204, Miramare Campus (Leonardo Building, strada Costiera 11, Trieste)




Valerio Vitale


YouTube Channel of the series: here

Ultimo aggiornamento: 27-03-2023 - 14:14