Generating realistic global precipitation fields from modelled atmospheric circulation

Technical Summary: Generating Realistic Global Precipitation Fields from Modelled Atmospheric Circulation

Problem Statement
Accurate high-resolution precipitation simulation is critical for assessing climate change impacts, particularly regarding extreme events like floods and droughts. However, Earth System Models (ESMs) struggle to resolve precipitation explicitly due to computational constraints. Current ESMs rely on column-based parameterizations that approximate subgrid-scale processes (e.g., convection) independently for each grid column. These schemes often fail to capture horizontal interactions, leading to significant biases, such as the double Intertropical Convergence Zone (ITCZ) bias, and an inability to represent fine-scale spatial variability and extreme event statistics. Furthermore, generating large ensembles of high-resolution projections is computationally prohibitive when running dynamical models at fine resolutions.

Methodology
The authors propose a two-stage, generative machine learning framework that infers high-resolution (0.25°) global daily precipitation fields from a small set of low-resolution (1°) prognostic atmospheric variables. The framework is trained exclusively on ERA5 reanalysis data and applied to ESM outputs via domain adaptation.

1. Stage 1: Deterministic Regression (UNet)

   • Input: Four 1° atmospheric variables: specific humidity at 850 hPa, 10m zonal and meridional wind components, and mean sea-level pressure.
   • Architecture: A memory-efficient UNet with ResNet blocks and a single bottleneck attention layer.
   • Objective: To learn a deterministic mapping from the coarse atmospheric state to a 1° precipitation field.
   • Bias Correction: The raw UNet output inherits systematic biases from the ESM inputs. To correct this, the authors apply Quantile Delta Mapping (QDM) using the historical period (1980–2014) to align the UNet output with 1° ERA5 precipitation.

2. Stage 2: Probabilistic Downscaling (Conditional Diffusion Model)

   • Input: The QDM-corrected 1° precipitation field (interpolated to 0.25°).
   • Architecture: A Denoising Diffusion Probabilistic Model (DDPM) based on an Efficient U-Net, conditioned on the coarse precipitation.
   • Training Strategy: The model is trained to reconstruct high-resolution (0.25°) ERA5 precipitation patterns. Crucially, during training, Gaussian noise is injected into the 1° conditioning input. This forces the model to learn to regenerate small-scale spatial patterns consistent with the large-scale structures provided by the condition, rather than simply memorizing the input.
   • Inference & Domain Adaptation: When applying the model to ESM data, the pipeline ensures the input distribution matches the training distribution. This involves:
     • Applying QDM to correct large-scale biases.
     • Interpolating to the target resolution.
     • Injecting the same magnitude of Gaussian noise used during training. This step effectively removes the smooth, biased small-scale features of the ESM output, allowing the diffusion model to "regenerate" realistic small-scale variability consistent with ERA5 statistics while preserving the large-scale dynamics of the ESM.

Key Results

• Bias Reduction: The framework significantly reduces spatial biases compared to the raw GFDL-ESM4 model. The mean absolute bias drops from 0.561 mm/day (GFDL) to 0.164 mm/day (QDM-corrected UNet) and further to 0.111 mm/day (Diffusion Model). The method effectively mitigates the double ITCZ bias prevalent in the GFDL model.
• Statistical Fidelity: The diffusion model successfully reproduces the spatial power spectral density (PSD) of ERA5, recovering small-scale variability (down to ~100 km) that is lost in the coarse ESM output. Histograms show improved representation of precipitation frequencies, particularly for extreme events (>150 mm/day), though a slight underestimation of the most extreme tails remains.
• Ensemble Generation: Unlike deterministic parameterizations, the diffusion model generates diverse, spatially consistent ensembles. Evaluation using the Continuous Ranked Probability Score (CRPS) shows the model outperforms deterministic baselines and provides a well-calibrated representation of uncertainty (spread-skill relationship).
• Climate Trend Preservation: When applied to future climate scenarios (SSP3-7.0), the framework preserves the large-scale climate change signals (trends) projected by the ESM while correcting small-scale spatial structures. The diffusion model dampens the unrealistically high inter-annual variance introduced by the QDM step in the regression model, resulting in global mean trends that align closely with the driving ESM projections.
• Generalizability: The approach was successfully applied to a different ESM (MPI-ESM-HR) without retraining, demonstrating robustness across models with similar spectral characteristics.

Significance and Claims
The paper claims to present a computationally efficient alternative to traditional column-based parameterizations. By leveraging generative machine learning, the framework can produce high-resolution (0.25°) probabilistic precipitation fields from coarse atmospheric variables in seconds, a task that would be prohibitively expensive with dynamical downscaling.

The authors emphasize that their method:

• Decouples scales: It allows impact models to receive high-resolution precipitation data without requiring the driving atmospheric model to run at that resolution.
• Corrects biases: It mitigates systematic ESM biases (like the double ITCZ) and restores fine-scale spatial variability that ESMs typically blur.
• Maintains consistency: It preserves the large-scale climate change signals projected by the ESMs while generating small-scale patterns consistent with observational reanalysis (ERA5).
• Offers a new paradigm: It demonstrates that complex precipitation patterns can be learned directly from large-scale atmospheric variables, offering a viable path for generating large ensembles of high-resolution climate projections for risk assessment.

The authors note limitations, including the assumption that the statistical relationship between large-scale circulation and small-scale precipitation remains valid under future warming, and the lack of explicit temporal constraints on fine-scale features (though large-scale temporal continuity is inherited from the atmospheric inputs). They acknowledge that ERA5 itself contains biases but remains the standard for training due to its global coverage and physical consistency.