Assessing the Geographic Generalization and Physical Consistency of Generative Models for Climate Downscaling

TL;DR

This work assesses the geographic generalization and physical consistency of state-of-the-art deep-learning downscalers for climate downscaling, revealing significant generalization gaps when transferring from Central Europe to other regions. It introduces a PSD-based loss as a frequency-domain regularizer to encourage recovery of small-scale structures, improving spectral fidelity and some physics-derived diagnostics. Despite gains, out-of-distribution regions still exhibit degraded accuracy and physical inconsistency, underscoring the challenge of transferability in high-resolution climate predictions. The findings offer a practical path toward more reliable downscaling by embedding frequency-domain regularization and physics-aware checks, with code available for reproducibility.

Abstract

Kilometer-scale weather data is crucial for real-world applications but remains computationally intensive to produce using traditional weather simulations. An emerging solution is to use deep learning models, which offer a faster alternative for climate downscaling. However, their reliability is still in question, as they are often evaluated using standard machine learning metrics rather than insights from atmospheric and weather physics. This paper benchmarks recent state-of-the-art deep learning models and introduces physics-inspired diagnostics to evaluate their performance and reliability, with a particular focus on geographic generalization and physical consistency. Our experiments show that, despite the seemingly strong performance of models such as CorrDiff, when trained on a limited set of European geographies (e.g., central Europe), they struggle to generalize to other regions such as Iberia, Morocco in the south, or Scandinavia in the north. They also fail to accurately capture second-order variables such as divergence and vorticity derived from predicted velocity fields. These deficiencies appear even in in-distribution geographies, indicating challenges in producing physically consistent predictions. We propose a simple initial solution: introducing a power spectral density loss function that empirically improves geographic generalization by encouraging the reconstruction of small-scale physical structures. The code for reproducing the experimental results can be found at https://github.com/CarloSaccardi/PSD-Downscaling

Figures (14)

• Figure 1: Input, target, and CRPS UNets-PSD prediction of $\mathbf{u}$ in Central Europe (in-distribution region) on 1st of January 2021
• Figure 2: Input, target, and CRPS UNets-PSD prediction of $\mathbf{u}$ in the Iberia, Morocco region (OOD) on 1st of January 2021. The prediction is much blurrier compared to in-distribution as shown Figure \ref{['fig:preds-CE-maintext']}.
• Figure 3: PSD comparison of down-scaled (top row) and physics-derived (bottom row) variables from CERRA (ground truth) and model predictions in Central Europe (in-distribution domain). Wavenumber $k$ is plotted on a logarithmic scale.
• Figure 4: Input, target, and predictions of $\mathbf{u}$ in Central Europe (in-distribution) using CorrDiff, Reg-CorrDiff-PSD, and CRPS UNets-PSD. CorrDiff yields a visually sharper prediction, while the others appear slightly blurrier.
• Figure 5: Input, target, and predictions of $\mathbf{t2m}$ in Central Europe (in-distribution) using CorrDiff, Reg-CorrDiff-PSD, and CRPS UNets-PSD. $\mathbf{t2m}$ is easier to predict, hence differences between outputs are hard to spot visually.