Kilometer-scale convection-allowing model emulation using generative diffusion modeling.

Area of Science:

• Atmospheric Science
• Meteorology
• Machine Learning

Background:

• Storm-scale convection-allowing models (CAMs) are crucial for predicting extreme weather but are computationally intensive.
• Deep learning models have struggled to achieve skill in kilometer-scale atmospheric simulations, unlike coarser resolutions.
• Accurate prediction of severe weather phenomena like thunderstorms and mesoscale convective systems remains a challenge.

Purpose of the Study:

• To develop a deep learning model capable of skillful kilometer-scale atmospheric simulation for weather forecasting.
• To emulate the performance of operational convection-allowing models using a generative diffusion approach.
• To assess the model's ability to predict key atmospheric variables and convective phenomena.

Main Methods:

• Developed StormCast, a generative diffusion model, to emulate the High-Resolution Rapid Refresh (HRRR) operational CAM.
• Utilized autoregressive prediction of 99 state variables at kilometer scale with a 1-hour time step.
• Conditioned the model on 26 synoptic variables and incorporated dense vertical resolution in the atmospheric boundary layer.

Main Results:

• StormCast demonstrates successful learning of kilometer-scale atmospheric dynamics.
• Achieved competitive 1- to 6-hour forecast skill for composite radar reflectivity.
• Generated physically realistic convective cluster evolution, moist updrafts, and cold pool morphology.

Conclusions:

• Generative diffusion models show promise for improving kilometer-scale regional machine learning weather prediction.
• StormCast's performance suggests potential for enhancing dynamical downscaling in climate hazard research.
• This work opens new avenues for leveraging AI in high-resolution weather forecasting and climate impact studies.