[Paper Review] Kilometer-Scale Convection Allowing Model Emulation using Generative Diffusion Modeling

Jaideep Pathak, Yair Cohen+9 more|arXiv (Cornell University)|Aug 20, 2024

Simulation Techniques and Applications6 citations

TL;DR

StormCast uses a generative diffusion model to emulate km-scale convection from HRRR, enabling autoregressive 1-hour steps with 3 km resolution and 26 synoptic inputs, achieving competitive skill for 1–6 hour forecasts and realistic convective dynamics.

ABSTRACT

Storm-scale convection-allowing models (CAMs) are an important tool for predicting the evolution of thunderstorms and mesoscale convective systems that result in damaging extreme weather. By explicitly resolving convective dynamics within the atmosphere they afford meteorologists the nuance needed to provide outlook on hazard. Deep learning models have thus far not proven skilful at km-scale atmospheric simulation, despite being competitive at coarser resolution with state-of-the-art global, medium-range weather forecasting. We present a generative diffusion model called StormCast, which emulates the high-resolution rapid refresh (HRRR) model-NOAA's state-of-the-art 3km operational CAM. StormCast autoregressively predicts 99 state variables at km scale using a 1-hour time step, with dense vertical resolution in the atmospheric boundary layer, conditioned on 26 synoptic variables. We present evidence of successfully learnt km-scale dynamics including competitive 1-6 hour forecast skill for composite radar reflectivity alongside physically realistic convective cluster evolution, moist updrafts, and cold pool morphology. StormCast predictions maintain realistic power spectra for multiple predicted variables across multi-hour forecasts. Together, these results establish the potential for autoregressive ML to emulate CAMs -- opening up new km-scale frontiers for regional ML weather prediction and future climate hazard dynamical downscaling.

Motivation & Objective

• Motivate km-scale convection prediction and the need for fast, high-resolution emulators of CAMs.
• Propose a generative diffusion framework to emulate km-scale atmospheric states conditioned on synoptic-scale inputs.
• Demonstrate that StormCast can produce realistic convection and competitive short-range forecast skill.
• Show that the approach yields realistic power spectra and multivariate convective structures.

Proposed method

• Formulate the time-stepped conditional distribution p_theta(M_{t+1} | S_t, M_t) for high-resolution mesoscale states based on a 1-hour time step.
• Use a two-phase learning: (i) deterministic regression F_theta to estimate the conditional mean, and (ii) stochastic diffusion via elucidated diffusion models (EDM) to model residuals.
• Train on HRRR as the mesoscale target and ERA5 as synoptic conditioning data, with 3 km resolution and 125 m–500 m vertical spacing.
• Autoregressively sample M_{t+1} from p_theta using the combination of learned mean mu_{t+1} and diffusion-based residual r_{t+1}.

Experimental results

Research questions

• RQ1 Can a generative diffusion model learn km-scale atmospheric dynamics conditioned on synoptic-scale inputs to emulate a convection-allowing model?
• RQ2 Does StormCast produce realistic km-scale convection, with competitive short-range skill and physically consistent vertical structure?
• RQ3 Can the model generate probabilistic ensembles offering meaningful forecast uncertainty at km scales?

Key findings

• StormCast attains competitive 1–6 hour skill for composite radar reflectivity compared to HRRR at thresholds 20dBZ, 30dBZ, and 40dBZ.
• StormCast ensembles via PMM show realistic multivariate convective dynamics including moist updrafts and cold pool morphology.
• Power spectra and probability distributions for several variables remain realistic at lead times up to several hours, with diffusion improving small-scale variance.
• Case studies demonstrate physically plausible multivariate convection features, such as updrafts co-located with enthalpy anomalies and gust fronts.
• The model enables efficient generation of km-scale forecasts and ensembles, suggesting potential for regional ML weather prediction and dynamical downscaling.