Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dynamical-generative downscaling of climate model ensembles

Ignacio Lopez-Gomez, Zhong Yi Wan, Leonardo Zepeda-Núñez, Tapio Schneider, John Anderson, Fei Sha

TL;DR

A paradigm that jointly exploits physics-based models and generative AI to drastically reduce the cost of downscaling climate projections, while retaining the skill of physics-based approaches is proposed, which enables translating large climate projection ensembles into impact-relevant climate risk assessments.

Abstract

Regional high-resolution climate projections are crucial for many applications, such as agriculture, hydrology, and natural hazard risk assessment. Dynamical downscaling, the state-of-the-art method to produce localized future climate information, involves running a regional climate model (RCM) driven by an Earth System Model (ESM), but it is too computationally expensive to apply to large climate projection ensembles. We propose a novel approach combining dynamical downscaling with generative artificial intelligence to reduce the cost and improve the uncertainty estimates of downscaled climate projections. In our framework, an RCM dynamically downscales ESM output to an intermediate resolution, followed by a generative diffusion model that further refines the resolution to the target scale. This approach leverages the generalizability of physics-based models and the sampling efficiency of diffusion models, enabling the downscaling of large multi-model ensembles. We evaluate our method against dynamically-downscaled climate projections from the CMIP6 ensemble. Our results demonstrate its ability to provide more accurate uncertainty bounds on future regional climate than alternatives such as dynamical downscaling of smaller ensembles, or traditional empirical statistical downscaling methods. We also show that dynamical-generative downscaling results in significantly lower errors than bias correction and spatial disaggregation (BCSD), and captures more accurately the spectra and multivariate correlations of meteorological fields. These characteristics make the dynamical-generative framework a flexible, accurate, and efficient way to downscale large ensembles of climate projections, currently out of reach for pure dynamical downscaling.

Dynamical-generative downscaling of climate model ensembles

TL;DR

A paradigm that jointly exploits physics-based models and generative AI to drastically reduce the cost of downscaling climate projections, while retaining the skill of physics-based approaches is proposed, which enables translating large climate projection ensembles into impact-relevant climate risk assessments.

Abstract

Regional high-resolution climate projections are crucial for many applications, such as agriculture, hydrology, and natural hazard risk assessment. Dynamical downscaling, the state-of-the-art method to produce localized future climate information, involves running a regional climate model (RCM) driven by an Earth System Model (ESM), but it is too computationally expensive to apply to large climate projection ensembles. We propose a novel approach combining dynamical downscaling with generative artificial intelligence to reduce the cost and improve the uncertainty estimates of downscaled climate projections. In our framework, an RCM dynamically downscales ESM output to an intermediate resolution, followed by a generative diffusion model that further refines the resolution to the target scale. This approach leverages the generalizability of physics-based models and the sampling efficiency of diffusion models, enabling the downscaling of large multi-model ensembles. We evaluate our method against dynamically-downscaled climate projections from the CMIP6 ensemble. Our results demonstrate its ability to provide more accurate uncertainty bounds on future regional climate than alternatives such as dynamical downscaling of smaller ensembles, or traditional empirical statistical downscaling methods. We also show that dynamical-generative downscaling results in significantly lower errors than bias correction and spatial disaggregation (BCSD), and captures more accurately the spectra and multivariate correlations of meteorological fields. These characteristics make the dynamical-generative framework a flexible, accurate, and efficient way to downscale large ensembles of climate projections, currently out of reach for pure dynamical downscaling.
Paper Structure (33 sections, 26 equations, 15 figures, 1 table)

This paper contains 33 sections, 26 equations, 15 figures, 1 table.

Table of Contents

  1. Introduction
  2. Results
  3. Generative downscaling skill
  4. Quantifying uncertainty in multi-model climate projections
  5. Regional analysis of compound extremes
  6. Discussion
  7. Methods
  8. Data for learning and evaluation
  9. Generative model design and training
  10. Baselines
  11. Method details
  12. Regional Residual Diffusion-based Downscaling
  13. Diffusion model formulation
  14. Model architecture
  15. Encoding
  16. ...and 18 more sections

Figures (15)

  • Figure 1: Schematic of the dynamical-generative downscaling framework. A regional climate model (RCM) is used to downscale global simulations from different ESMs to an intermediate grid. A generative artificial intelligence system (GenAI), such as a diffusion model, is then used to further downscale the RCM output to the desired resolution. The topographic height is shown at 100 km (left), 45 km (middle), and 9 km (right) resolution, to showcase landscape changes at the different scales of the process. Water bodies are highlighted in blue.
  • Figure 2: Land-averaged downscaling skill measured by CRPS (a-d), and radially averaged energy spectra (e-h) for select near-surface meteorological fields. Results are computed using 4-hourly data spanning years 2095-2096 of the multi-model SSP3-7.0 climate projection, and shown for cubic interpolation (Interp.), BCSD, and for 32-member R2-D2 ensembles.
  • Figure 3: Spatial distribution of downscaling bias (first and third columns) and CRPS (second and fourth columns), shown for near-surface temperature and precipitation. Results are computed using 4-hourly data spanning years 2095-2096 of the multi-model SSP3-7.0 climate projection, and shown for cubic interpolation, BCSD, and for 32-member R2-D2 ensembles. The insets show the land-averaged absolute value of each metric.
  • Figure 4: Assessment of downscaled multi-model ensemble distribution fidelity. Top: MAE of downscaled quantiles over land, with respect to the quantiles of the full dynamically downscaled ensemble. Bottom: Quantile-quantile plots at specific locations with respect to the full dynamically downscaled ensemble. Quantiles from 0.01 to 0.99 are computed using daily snapshots at 00 UTC for three-month periods of 2095. Results are shown for cubic interpolation, BCSD, the generative model R2-D2, and for the average over 4-member sub-ensembles. Uncertainty estimates represent the bootstrapped sample standard deviation.
  • Figure 5: Analysis of the strongest Santa Ana wind event in the multi-model projection for the period September-November 2095. Top: Topographic and coarse-resolution context of the event. The anomaly fields (b,c) show the 45-km resolution anomalies with respect to their September-November 2095 climatology. Bottom: Downscaled Santa Ana wildfire threat index (SAWTI) from WRF (d), R2-D2 (e), and BCSD (f). Quiver plots represent the magnitude and direction of 10 m winds, and dashed contours represent the 1200 m isohypse of the mountain ranges.
  • ...and 10 more figures