AI-Based Regional Emulation for Kilometer-Scale Dynamical Downscaling

TL;DR

An AI-based Limited-Area Model developed for dynamical downscaling over the Southern Great Plains and the southeastern United States provides a practical and transferable example of adapting AI weather prediction models for regional climate applications.

Abstract

An AI-based Limited-Area Model (LAM) is developed for dynamical downscaling over the Southern Great Plains and the southeastern United States, with strong generalization abilities under diverse boundary conditions. The model is trained using 0.25-degree, 3-hourly ERA5 as forcings and CONUS404 as targets in 1980--2019, producing 4-km, hourly dynamical downscaling outputs; it is also connected to a post-processing model to derive additional diagnostic variables. The model is evaluated across multiple forcing datasets, time periods, and climate regimes. For present-day downscaling in the 2021--2024 water years, the model produces stable multi-year simulations with no unrealistic drift; its deterministic verification scores are comparable to other weather-forecasting-oriented AI models. The model also generalizes robustly to a 1.0-degree, 6-hourly non-ERA5 forcing dataset, yielding only minor performance changes. Frontal cyclone and hurricane case studies further demonstrate that the model reconstructs realistic, interpretable weather-scale dynamical and thermodynamic structure from coarse boundary information. The AI-based LAM is further tested by downscaling 30-year global climate model runs in 1980--2010 and 2070--2100, and climate model ensembles in 2025-2027. In this application, the model remains stable at hourly downscaling frequencies for all 30 years and effectively captures future climate-change signals, indicating meaningful generalization across different climate regimes. When downscaling ensembles, the model produces well-posed ensemble distributions without collapsing the ensemble spread. Overall, the AI-based LAM of this study offers good downscaling performance and generalization abilities. It provides a practical and transferable example of adapting AI weather prediction models for regional climate applications.

Figures (12)

• Figure 1: (a) The 4-km grid spacing domain with shaded elevation. (b) Vegetation categories of the domain. (c) Architecture of the AI-based LAM. (d) Architecture of the U-Net-based post-processing model. (e) The dynamical downscaling workflow. (f) An example of time window matching between hourly dynamical downscaling and 6-hourly boundary forcings.
• Figure 2: (a) Domain-averaged and 30-day moving averaged hourly 2-m air temperature ("2-m Temp") RMSE in WY 2021--2024 for LAM-3H-ERA5 (blue), LAM-6H-ERA5 (red), and LAM-6H-GDAS (orange). Light blue shading indicates boreal winter. (b) The energy spectrum of 2-m air temperature in the 10--100 km wavelength range, where the black dashed line is computed from CONUS404. (c--d) As in (a--b), but for 700 hPa air temperature ("700 hPa Temp"). (e--f) As in (a--b), but for 10-m wind speed. (g--h) As in (a--b), but for 700 hPa wind speed. (i--j) As in (a--b), but for total precipitable water ("PWAT"). (k--l) As in (a--b), but for total water path ("Q").
• Figure 3: (a) 2-m air temperature ("2-m Temp") produced by LAM-3H-ERA5 averaged over WY 2021--2024. (b) The difference between (a) and the corresponding CONUS404 target.(c--d) As in (a--b), but for air temperature ("T") along the zonal cross-section from 33.3$^\circ$N, 102.8$^\circ$W to 33.2$^\circ$N, 91.1.0$^\circ$W (dashed gray line). (e--f) As in (a--b), but for 10-m wind speed. (g--h) As in (c--d), but for wind speed aloft ("SPD"). (i--j) As in (a--b), but for total precipitable water ("PWAT"). (k--l) As in (c--d), but for total water path ("Q").
• Figure 4: As in Figure \ref{['fig3']}, but for the downscaling experiment of LAM-6H-ERA5.
• Figure 5: As in Figure \ref{['fig3']}, but for the downscaling experiment of LAM-6H-GDAS.