Fast, Scale-Adaptive, and Uncertainty-Aware Downscaling of Earth System Model Fields with Generative Machine Learning

TL;DR

This work presents a consistency model (CM) for fast, scale-adaptive, and probabilistic downscaling of Earth system model precipitation fields. Trained only on ERA5 ground truth, the CM performs zero-shot downscaling of arbitrary ESM outputs in a single step, producing high-fidelity, uncertainty-enabled high-resolution fields without explicit physical constraints. Compared with diffusion-based baselines, CM achieves higher correlation with native high-resolution fields, enables ensemble-based uncertainty quantification, and offers enormous speedups (approximately three orders of magnitude faster) suitable for large ESM ensembles and long climate projections. The method generalizes to unseen climate states (e.g., SSP5-8.5) and provides a practical, scalable tool for impact assessment, weather prediction, and greener AI, with potential extensions to temporally conditioned and multi-variate downscaling.

Abstract

Accurate and high-resolution Earth system model (ESM) simulations are essential to assess the ecological and socio-economic impacts of anthropogenic climate change, but are computationally too expensive to be run at sufficiently high spatial resolution. Recent machine learning approaches have shown promising results in downscaling ESM simulations, outperforming state-of-the-art statistical approaches. However, existing methods require computationally costly retraining for each ESM and extrapolate poorly to climates unseen during training. We address these shortcomings by learning a consistency model (CM) that efficiently and accurately downscales arbitrary ESM simulations without retraining in a zero-shot manner. Our approach yields probabilistic downscaled fields at a resolution only limited by the observational reference data. We show that the CM outperforms state-of-the-art diffusion models at a fraction of computational cost while maintaining high controllability on the downscaling task. Further, our method generalizes to climate states unseen during training without explicitly formulated physical constraints.

Figures (5)

• Figure 1: Sketch of the consistency model for downscaling of Earth system model fields. (Upper panel) Unconditional training of the score-based diffusion and consistency models (CM) that learn to reverse a forward diffusion process. While the stochastic differential equation of the diffusion model requires an iterative integration over many steps, the CM only takes a single step to generate a global precipitation field from noise. (Lower panel) The unconditionally trained consistency model is used to downscale (upsample) a low-resolution ESM precipitation field to a four times higher resolution. By adding noise of a chosen variance to the ESM field, the spatial scale to be preserved in the ESM can be controlled: small noise variance implies a close pairing to the original ESM field with only small changes; a larger variance will result in changes at larger spatial scales and in weaker pairing to the ESM field.
• Figure 2: Qualitative comparison of single-day precipitation fields. (A) Daily precipitation from the ERA5 target dataset was used for training the generative models. (B) Same as (A) but at four times lower resolution for comparisons. (C) A precipitation field from a historical run of the POEM ESM interpolated to the target resolution and (D) on its native resolution of $3^\circ \times 3.75^\circ$. The POEM fields are unpaired with the ERA5 field from the same date or any other ERA5 field. Downscaled field from POEM (D) with the SDE bridge method (D)$\rightarrow$(E). (F) An upscaled (average pooled) representation of (E) for comparison with the original POEM field is shown in D and the Pearson correlation between the two. Downscaling POEM with the CM-based method (D)$\rightarrow$(G), and the respective pooled field (H). Note that the CM downscaling yields a higher correlation, and hence better consistency of the large-scale features, than the SDE method.
• Figure 3: Mean spatial power spectral densities (PSDs) of global precipitation fields. (A) Comparison of the PSDs for the target ERA5 reanalysis data (black), the POEM simulations interpolated to the same high-resolution grid (orange), the SDE bridge (cyan), and the CM downscaling (magenta). The vertical dashed lines mark the spatial scale at which the PSDs of POEM and ERA5 intersect and are thus a natural choice for the wavenumber $k^{\ast}$ up to which to correct, which in turn determines $t^{\ast}$, i.e. the noising strength in the diffusion models (see Eq. \ref{['main:eq:stop_time']} in the Methods). (B) CM downscaling (dashed lines) applied to be consistent with different spatial scales as a function of the noising strength $t$ over the entire range $[t_{\mathrm{min}}, t_{\mathrm{max}}]$. Noising small scales implies nearly reproducing the POEM simulations, while noising larger scales corresponds to a weaker pairing to the ESM (see Fig. \ref{['main:fig:method_overview']}).
• Figure 4: Comparison of global histograms and longitudinal mean precipitation. (A) Global histograms of relative precipitation frequency for the ERA5 reanalysis data (black), POEM simulations without applying the QDM-preprocessing (grey), POEM simulations with QDM (orange), the SDE bridge (cyan), and the CM (magenta). (B) Absolute errors of the histograms in (A) with respect to the ERA5 ground truth. (C) Precipitation averaged over time and longitudes for the same data as in (A). (D) Absolute errors of the latitude profile in (C). Both the SDE and the CM downscaling method are able to further improve upon the QDM-preprocessing in terms of bias correction, most notably for extreme precipitation.
• Figure 5: Sampling spread and generalization to unseen climates of the generative probabilistic downscaling process. (A) The ESM field interpolated to the target resolution. (B) and (C) show two different exemplary samples generated by the CM downscaling, preserving large-scale patterns and generating new patterns on smaller scales. (D) The ensemble mean of $10^3$ samples with the standard deviation is shown in (E). (F) Three-year rolling global mean normalized to the reference year 2020 of the very high emission scenario SSP5-8.5. The ESM (orange) shows an increase in global mean precipitation over the emission scenario, in line with the thermodynamic Clausius-Clapeyron relation traxl_role_2021. The CM downscaling (magenta) is able to preserve the trend with a high degree of accuracy, notably without the addition of any physical constraints in the CM network.