Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HiRO-ACE: Fast and skillful AI emulation and downscaling trained on a 3 km global storm-resolving model

W. Andre Perkins, Anna Kwa, Jeremy McGibbon, Troy Arcomano, Spencer K. Clark, Oliver Watt-Meyer, Christopher S. Bretherton, Lucas M. Harris

TL;DR

HiRO-ACE delivers a fast, probabilistic pathway to kilometer-scale precipitation by coupling a stochastic 100 km global emulator (ACE2S) with a diffusion-based downscaling model (HiRO) trained on a decade of 3 km X-SHiELD output. ACE2S preserves grid-scale variability and climate fidelity, while HiRO learns to generate realistic 3 km textures from 100 km states, enabling plausible tropical cyclones, fronts, and convective features across most of the globe. The end-to-end framework achieves near-X-SHiELD extreme-precipitation statistics (up to the 99.99th percentile) and accurate time-mean spatial patterns with biases largely within internal variability, running hundreds of times faster than direct km-scale simulations. This enables large ensembles and long-term regional precipitation studies for local climate adaptation and extreme-event risk assessment, with potential extensions to multi-variable downscaling and climate-change contexts.

Abstract

Kilometer-scale simulations of the atmosphere are an important tool for assessing local weather extremes and climate impacts, but computational expense limits their use to small regions, short periods, and limited ensembles. Machine learning offers a pathway to efficiently emulate these high-resolution simulations. Here we introduce HiRO-ACE, a two-stage AI modeling framework combining a stochastic version of the Ai2 Climate Emulator (ACE2S) with diffusion-based downscaling (HiRO) to generate 3 km precipitation fields over arbitrary regions of the globe. Both components are trained on data derived from a decade of atmospheric simulation by X-SHiELD, a 3 km global storm-resolving model. HiRO performs a 32x downscaling--generating 3 km 6-hourly precipitation from coarse 100 km inputs by training on paired high-resolution and coarsened X-SHiELD outputs. ACE2S is a $1^\circ \times 1^\circ$ ($\sim$100 km) stochastic autoregressive global atmosphere emulator that maintains grid-scale precipitation variability consistent with coarsened X-SHiELD, enabling its outputs to be ingested by HiRO without additional tuning. HiRO-ACE reproduces the distribution of extreme precipitation rates through the 99.99th percentile, with time-mean precipitation biases below 10% almost everywhere. The framework generates plausible tropical cyclones, fronts, and convective events from poorly resolved coarse inputs. Its computational efficiency allows generation of 6-hourly high-resolution regional precipitation for decades of simulated climate within a single day using one H100 GPU, while the probabilistic design enables ensemble generation for quantifying uncertainty. This establishes an AI-enabled pathway for affordably leveraging the realism of expensive km-scale simulations to support local climate adaptation planning and extreme event risk assessment.

HiRO-ACE: Fast and skillful AI emulation and downscaling trained on a 3 km global storm-resolving model

TL;DR

HiRO-ACE delivers a fast, probabilistic pathway to kilometer-scale precipitation by coupling a stochastic 100 km global emulator (ACE2S) with a diffusion-based downscaling model (HiRO) trained on a decade of 3 km X-SHiELD output. ACE2S preserves grid-scale variability and climate fidelity, while HiRO learns to generate realistic 3 km textures from 100 km states, enabling plausible tropical cyclones, fronts, and convective features across most of the globe. The end-to-end framework achieves near-X-SHiELD extreme-precipitation statistics (up to the 99.99th percentile) and accurate time-mean spatial patterns with biases largely within internal variability, running hundreds of times faster than direct km-scale simulations. This enables large ensembles and long-term regional precipitation studies for local climate adaptation and extreme-event risk assessment, with potential extensions to multi-variable downscaling and climate-change contexts.

Abstract

Kilometer-scale simulations of the atmosphere are an important tool for assessing local weather extremes and climate impacts, but computational expense limits their use to small regions, short periods, and limited ensembles. Machine learning offers a pathway to efficiently emulate these high-resolution simulations. Here we introduce HiRO-ACE, a two-stage AI modeling framework combining a stochastic version of the Ai2 Climate Emulator (ACE2S) with diffusion-based downscaling (HiRO) to generate 3 km precipitation fields over arbitrary regions of the globe. Both components are trained on data derived from a decade of atmospheric simulation by X-SHiELD, a 3 km global storm-resolving model. HiRO performs a 32x downscaling--generating 3 km 6-hourly precipitation from coarse 100 km inputs by training on paired high-resolution and coarsened X-SHiELD outputs. ACE2S is a (100 km) stochastic autoregressive global atmosphere emulator that maintains grid-scale precipitation variability consistent with coarsened X-SHiELD, enabling its outputs to be ingested by HiRO without additional tuning. HiRO-ACE reproduces the distribution of extreme precipitation rates through the 99.99th percentile, with time-mean precipitation biases below 10% almost everywhere. The framework generates plausible tropical cyclones, fronts, and convective events from poorly resolved coarse inputs. Its computational efficiency allows generation of 6-hourly high-resolution regional precipitation for decades of simulated climate within a single day using one H100 GPU, while the probabilistic design enables ensemble generation for quantifying uncertainty. This establishes an AI-enabled pathway for affordably leveraging the realism of expensive km-scale simulations to support local climate adaptation planning and extreme event risk assessment.

Paper Structure

This paper contains 40 sections, 17 equations, 17 figures.

Table of Contents

  1. Introduction
  2. HiRO-ACE Overview
  3. HiRO-ACE Framework
  4. Model components
  5. Computational requirements
  6. Key Results
  7. Accurate frequency of km-scale extreme precipitation
  8. Accurate km-scale spatial pattern of time-mean precipitation
  9. Well-reconstructed km-scale precipitation features
  10. 11-year X-SHiELD reference simulation
  11. Stochastic ACE2S trained on X-SHiELD
  12. Model architecture and design
  13. Training Strategy
  14. ACE2S improves on ACE2
  15. HiRO downscaling trained on X-SHiELD
  16. ...and 25 more sections

Figures (17)

  • Figure 1: Schematic of the two-stage HiRO-ACE framework.
  • Figure 2: Global (65° S--65° N) fine-grid 6-hourly precipitation statistics during the 2023 hold-out year for X-SHiELD, a single HiRO 'perfect prediction' ensemble member, and a single HiRO-ACE ensemble member: (a) Histograms of binned 6-hr average surface precipitation rates over all 3 km grid columns and sampling times, with the 99th, 99.99th, 99.9999th percentiles of the data marked with vertical dashed lines; (b) Time and meridional mean of the zonal power spectrum.
  • Figure 3: Annual-mean precipitation over the 2023 hold-out year for (a) a single selected ensemble member of ACE2S, (b) coarsened X-SHiELD, (c) one HiRO-ACE ensemble member, (d) one HiRO perfect prediction ensemble member, and (e) fine-grid X-SHiELD.
  • Figure 4: Tropical cyclone case study near southern China showing (a) ACE2S 6-hour forecast of 6 h average surface precipitation rate at 100 km resolution, (b) HiRO downscaled output at 3 km resolution, and (c) target X-SHiELD precipitation at 3 km resolution.
  • Figure 5: Comparison of surface precipitation (a) histograms and (b) spherical power spectrum, averaged over all 6-hourly samples of 10-year runs of deterministic ACE2 (blue), ACE2S (orange), and the coarsened X-SHiELD target.
  • ...and 12 more figures