Ilenia Manco(She/Her)Centro Euro-Mediterraneo sui Cambiamenti Climatici
Ilenia Manco, with a background in meteorology and oceanography, is a researcher at the CMCC (Euro-Mediterranean Center on Climate Change) and is currently completing her Ph.D. in Future Earth Climate Change and Societal Challenges at the University of Bologna, Italy. Specializing in meteorological and climatological modeling at various spatial scales, from the sensitivity of parameterizations to validation techniques of results, she has expanded her approach to this discipline with the use of sophisticated AI techniques. She has developed an innovative Deep Learning algorithm based on a Generative Adversarial Network (GAN), enabling statistical downscaling at high and very high resolutions for various regional domains and datasets. Her work in the field of AI is expanding with the development of algorithms ranging from lightning event prediction to precipitation nuclei clustering. |
 |
TalksAI-assisted Climate Downscaling of the ERA5 Reanalysis for Rapid Assessment23-Apr-24The state-of-the-art Global Circulation Models (GCMs) are still operated at such coarse spatial resolutions that necessitate refinement to assess regional climate changes and their impacts. This limitation is primarily attributed to the representation of regional-scale topography and meteorological processes, particularly those associated with extreme events. Conventional dynamical downscaling methods are computationally intensive. In contrast, though computationally efficient, statistical approaches often compromise spatial coherence. To address these limitations, this study introduces an innovative application of Generative Adversarial Networks (GANs). GANs consist of two interconnected components: a generative model and a discriminative model. The generative model, here in represented by the ERA5 climate reanalysis (~31 km), learns to produce high-resolution data. The discriminator, utilizing the VHR-REA_IT dataset (~2.2 km), distinguishes between real high-resolution data and data generated by the GAN (ERA5-DownGAN). This pioneering study utilizes the developed GAN architecture to downscale ERA5 to a resolution of ~2.2 km, particularly for the fields of 2m temperature, 10m wind, and total cumulative precipitation. The training phase (01/1990-12/2000) enables the generative model to learn high-resolution data production, while the testing phase (01/2001-12/2005) evaluates the GAN's performance against VHR-REA_IT. The computational domain focuses on the Italian Peninsula, encompassing parts of northern and central Europe and northern Africa. A set of conventional error metrics, and graphical representations enabling assessment of spatial and temporal correlation between datasets and percentile distribution, is analyzed to evaluate the performance of GANs. The results obtained have shown promise, both in terms of pattern reconstruction and value range (mean, median, extremes), defining this architecture as a potentially viable alternative to dynamical downscaling. |