Can AI be enabled to perform dynamical downscaling? A latent diffusion model to mimic kilometer-scale COSMO5.0_CLM9 simulations

Geoscientific model development, Journal Year: 2025, Volume and Issue: 18(6), P. 2051 - 2078

Published: April 1, 2025

Abstract. Downscaling based on deep learning (DL) is a key application in Earth system modeling, enabling the generation of high-resolution fields from coarse numerical simulations at reduced computational costs compared to traditional regional models. Additionally, generative DL models can potentially provide uncertainty quantification through ensemble-like scenario generation, task prohibitive for conventional approaches. In this study, we apply latent diffusion model (LDM) demonstrate that recent advancements modeling enable deliver results comparable those dynamical models, given same input data, preserving realism fine-scale features and flow characteristics costs. We our LDM downscale ERA5 data over Italy up resolution 2 km. The target consist m temperature 10 horizontal wind components downscaling performed with COSMO-CLM. A selection predictors used as input, residual approach against reference U-Net leveraged applying LDM. performance baselines increasing complexity: quadratic interpolation ERA5, U-Net, adversarial network (GAN) built U-Net. Results highlight improvements introduced by architecture combined approach, outperforming all terms spatial error, frequency distributions, power spectra. These findings point out potential LDMs cost-effective, robust alternatives applications (e.g., climate projections), where resources are limited but critical.

Language: Английский

Modeling and observations of North Atlantic cyclones: Implications for U.S. Offshore wind energy

Journal of Renewable and Sustainable Energy, Journal Year: 2024, Volume and Issue: 16(5)

Published: Sept. 1, 2024

To meet the Biden-Harris administration's goal of deploying 30 GW offshore wind power by 2030 and 110 2050, expansion energy into U.S. territorial waters prone to tropical cyclones (TCs) extratropical (ETCs) is essential. This requires a deeper understanding cyclone-related risks development robust, resilient systems. paper provides comprehensive review state-of-the-science measurement modeling capabilities for studying TCs ETCs, their impacts across various spatial temporal scales. We explore environments influenced including near-surface vertical profiles critical variables that characterize these cyclones. The limitations Earth system mesoscale models are assessed effectiveness in capturing atmosphere–ocean–wave interactions influence TC/ETC-induced under changing climate. Additionally, we discuss microscale designed bridge scale gaps from weather (a few kilometers) turbine (dozens meters). also machine learning (ML)-based, data-driven simulating TC/ETC events at both Special attention given extreme metocean conditions like gusts, rapid direction changes, high waves, which pose threats infrastructure. Finally, outlines research challenges future directions needed enhance resilience design next-generation turbines against conditions.

Language: Английский

A Generative Super‐Resolution Model for Enhancing Tropical Cyclone Wind Field Intensity and Resolution

Journal of Geophysical Research Machine Learning and Computation, Journal Year: 2024, Volume and Issue: 1(4)

Published: Nov. 20, 2024

Abstract Extreme winds associated with tropical cyclones (TCs) can cause significant loss of life and economic damage globally, highlighting the need for accurate, high‐resolution modeling forecasting wind. However, due to their coarse horizontal resolution, most global climate weather models suffer from chronic underprediction TC wind speeds, limiting use impact analysis energy modeling. In this study, we introduce a cascading deep learning framework designed downscale fields given low‐resolution data. Our approach maps 85 events ERA5 data (0.25° resolution) (0.05° observations at 6‐hr intervals. The initial component is debiasing neural network model accurate speed using second employs generative super‐resolution strategy based on conditional denoising diffusion probabilistic (DDPM) enhance spatial resolution produce ensemble estimates. able accurately intensity realistic radial profiles fine‐scale structures fields, percentage mean bias −3.74% compared observations. downscaling enables prediction widely available data, allowing past assessment future risks.

Language: Английский