Generative artificial intelligence improves projections of climate extremes

TL;DR

FuXi-CMIPAlign introduces a generative, domain-adaptive downscaling framework that bridges ERA5 training data and EC-Earth inference to produce high-resolution, multivariate climate fields from coarse CMIP inputs. By combining Flow Matching with Maximum Mean Discrepancy loss, it achieves improved accuracy, stability, and generalization for extremes, including tropical cyclones, across historical and future emission scenarios. The approach yields substantial reductions in 99th-percentile errors for temperature, precipitation, and wind, and reveals scenario-dependent shifts in extreme-event distributions, with notable TC activity increases in key Pacific regions. This framework offers a robust tool for climate-risk assessment, adaptation planning, and more reliable projection of extreme events under changing emissions.

Abstract

Climate change is amplifying extreme events, posing escalating risks to biodiversity, human health, and food security. GCMs are essential for projecting future climate, yet their coarse resolution and high computational costs constrain their ability to represent extremes. Here, we introduce FuXi-CMIPAlign, a generative deep learning framework for downscaling CMIP outputs. The model integrates Flow Matching for generative modeling with domain adaptation via MMD loss to align feature distributions between training data and inference data, thereby mitigating input discrepancies and improving accuracy, stability, and generalization across emission scenarios. FuXi-CMIPAlign performs spatial, temporal, and multivariate downscaling, enabling more realistic simulation of compound extremes such as TCs.

Figures (4)

• Figure 1: Spatial distribution of 99th-percentile mean absolute error (MAE) and accumulative tropical cyclone (TC) tracks from 2005 to 2014. a-c, show MAE for 2-meter temperature (T2M), 6-hour accumulated total precipitation (TP), and 10-meter wind speed (WS10M), respectively. d shows TCs, with line colors indicating intensity categories and numbers (n/year) denoting the annual mean number of TC tracks detected over the 10-year period. Columns correspond to European Consortium-Earth (EC-Earth), noalign, and FuXi-CMIPAlign, from left to right.
• Figure 2: Spatiotemporal trends of high temperature, extreme precipitation, and strong winds over land under emission scenarios (2015–2100). These events are defined by the 99th percentile thresholds of 2-meter temperature (T2M), 6-hour accumulated total precipitation (TP), and 10-meter wind speed (WS10M), respectively. a, Temporal evolution of land area fraction affected by each event type. The gray line represents historical observations based on ERA5 data (2005–2014), while the colored lines indicate projections from FuXi-CMIPAlign across SSP126 to SSP585 scenarios (2015–2100). b, Correction relative to European Consortium-Earth (EC-Earth): circles denote EC-Earth, arrows point to FuXi-CMIPAlign, and arrow length indicates the magnitude of correction. Values above arrowheads quantify corrections, with positive values indicating increases and negative values decreases. c, Global land frequency of event distribution in 2060 under SSP370. d, Spatial frequency of extreme events in three climate-sensitive regions (Middle East, Northwestern Australia, and Siberia) under SSP370, along with annual frequencies of extreme events for 2030, 2060, and 2100 under different scenarios.
• Figure 3: Detectable tropical cyclone (TCs) tracks under emission Scenarios (2015-2100). a, Cumulative spatial distribution of TC tracks detected with FuXi-CMIPAlign under four emission scenarios (SSP126, SSP245, SSP370, SSP585) from 2015 to 2100. The color of TC tracks transitions from dark green to dark red, representing the intensity from tropical depression to severe typhoon (hurricane). b, The upper stacked graph shows the incremental annual mean number of TC tracks detected by FuXi-CMIPAlign relative to European Consortium-Earth (EC-Earth) across four scenarios, with colors indicating TC intensity categories. Bars in different colors represent different SSP scenarios, while different shades of the same color represent different intensities of TCs. The lower bar chart summarizes the annual mean number of Severe Tropical Storms and Typhoons (or Hurricanes) simulated in major ocean basins during three periods (2016–2030, 2031–2060, 2061–2100), with frequency statistics shown for regions of frequent TC activity. Basin abbreviations: ATL (Atlantic Ocean), SPO (South Pacific Ocean), IND (Indian Ocean), NEP (Northeast Pacific Ocean), NWP (Northwest Pacific Ocean).
• Figure 4: Overview of the FuXi-CMIPAlign's architecture. a, Data sources and timeline, including European Consortium-Earth (EC-Earth, historical and scenario data) and ERA5. The overlapping period between the Historical and Reanalysis datasets spans from 1940 to 2014. In this study, the data from 1940 to 2000 were utilized as the training set, while the years 2001 and 2002 were designated as the validation set to monitor the training process. The period from 2005 to 2014 was employed as the test set to evaluate the performance of the model. b, Training stage: The model learns to reconstruct ERA5 fields at 6-hourly, 0.25$^\circ$ resolution from degraded ERA5 inputs at daily, 70km resolution, while aligning feature distributions between degraded ERA5 and EC-Earth using Maximum Mean Discrepancy (MMD) loss. Digital elevation model (DEM) and temporal features are included as additional conditions. c, Testing stage: The trained model downscales EC-Earth (daily, 70km resolution), using DEM and temporal features to generate high-resolution outputs (ERA5 variables at 6-hourly, 0.25$^\circ$ resolution) through iterative refinement. Gaussian noise is utilized to initialize the sampling process.