Settings

Now showing1 - 4 of 4
  • Some of the metrics are blocked by your 
    consent settings
    Item type:Publication, Access status: without files ,
    How do generative downscaling models use noise? Insights from conditional precipitation downscaling
    (American Meteorological Society (AMS), 2026-09-03)
    Singh, Shivam
    ;
    Papalexiou, Simon Michael
    ;
    Abdelmoaty, Hebatallah M.
    ;
    Hartvigsen, Tom
    ;
    Mamalakis, Antonios
    Generative deep learning models offer a promising framework for stochastic precipitation downscaling, but it remains unclear how they utilize large-scale conditioning inputs versus stochastic latent variables. This challenge limits physical interpretability and raises concerns about whether generated fine-scale variability reflects meaningful uncertainty or unconstrained randomness. Here, we investigate how different training objectives shape stochastic information usage in conditional precipitation downscaling, for a demanding task of 16× increase in spatial resolution. We analyze two models based on the same U-Net architecture: an MSE-trained U-Net and the same U-Net trained within a conditional Wasserstein generative adversarial network (WGAN). Using Integrated Gradients, noise ensemble sensitivity analysis, and layer-wise representation analysis, we quantify how each model uses low-resolution (LR) precipitation inputs and Gaussian noise to generate high-resolution precipitation fields. We find that the predicted mean precipitation is overwhelmingly attributed to the LR input in both models, whereas maximum intensity exhibits a relatively greater stochastic contribution. However, only the WGAN meaningfully exploits noise in practice to stochastically represent underdetermined aspects of the downscaling task. The MSE-trained U-Net exhibits comparatively limited and less organized utilization of stochastic input. Internal representation analysis further shows that these differences emerge primarily at the bottleneck and early decoder stages of the U-Net architecture. These results demonstrate that meaningful stochastic behavior in generative AI downscaling is not guaranteed by architectural design alone but is strongly shaped by the training objective. The study provides a framework for evaluating the physical credibility and trustworthiness of generative models in hydroclimatic applications.
    Publicationtype: Journal Article
      2
  • Some of the metrics are blocked by your 
    consent settings
    Item type:Publication, Access status: with files ,
    Comprehensive inter-comparison of generative AI models for super-resolution precipitation downscaling across hydroclimatic regimes
    (Copernicus Publications, 2026-08-17)
    Singh, Shivam
    ;
    Papalexiou, Simon Michael
    ;
    Abdelmoaty, Hebatallah M.
    ;
    Hartvigsen, Tom
    ;
    Mamalakis, Antonios
    High-resolution precipitation information is essential for hydrologic modeling, flood forecasting, and climate-risk assessment, yet global weather and climate models operate at spatial resolutions too coarse to resolve storm structure, intermittency, and extremes. Deep-learning-based statistical downscaling provides a computationally efficient alternative to dynamical downscaling, but deterministic convolutional neural networks often yield overly smooth predictions and underestimate fine-scale variability and extreme events. Generative deep-learning models, including generative adversarial networks and diffusion models, offer a promising alternative by enabling stochastic downscaling and explicit representation of uncertainty. This study presents a systematic intercomparison of three representative deep-learning architectures for precipitation super-resolution, namely a Convolutional U-Net as baseline, a conditional Wasserstein GAN (WGAN), and a conditional Denoising Diffusion Probabilistic Model (DDPM). Using a perfect-model experimental design based on ERA5-Land precipitation fields over climatologically distinct regions of the United States, models are trained over the Central Plains and Northwest domains and evaluated over an independent Northeast test domain under  8×and 16× downscaling factors, providing a stringent test of cross-regional generalization. Evaluation diagnostics span precipitation distributions, wet–dry occurrence, extremes, spatial autocorrelation, spectral structure, and ensemble-based uncertainty quantification. All three models preserve large-scale precipitation organization, with differences emerging primarily at fine spatial scales and in the representation of extremes and spatial dependence. U-Net provides stable and computationally efficient predictions but consistently smooths fine-scale variability and suppresses extreme precipitation. WGAN improves distributional fidelity and heavy-tail behavior with comparatively modest computational overhead. DDPM yields the most physically coherent spatial structure and natural ensemble diversity for uncertainty quantification, at a substantially higher computational cost. Analysis of seed-based variability further reveals that training uncertainty dominates over stochastic generation variability, underscoring the need for multi-seed evaluation in generative downscaling systems.
    Publicationtype: Journal Article
      12  2
  • Some of the metrics are blocked by your 
    consent settings
    Item type:Publication, Access status: with files ,
    Does nonstationarity affect GAN-based downscaling? Insights from high-resolution WRF simulations over the Canadian prairies
    (Society for Industrial & Applied Mathematics (SIAM), 2026-04-20)
    Abdelmoaty, Hebatallah M.
    ;
    Papalexiou, Simon Michael
    ;
    Mamalakis, Antonios
    ;
    Singh, Shivam
    ;
    Coia, Vincenzo
    Although downscaling models are typically trained on historical data and then applied to future projections, this practice raises concerns about their ability to reproduce nonstationary climate signals. While such limitations are expected for traditional statistical approaches, deep learning models have the potential to learn a more generalizable downscaling relationship, separating the transformation task from the nonstationary signal. In this study, we examine the sensitivity of deep learning–based downscaling to nonstationarity using Wasserstein generative adversarial networks (WGANs) to downscale hourly precipitation from 50 to 4 km over the Canadian Prairies. We trained two models on convection-permitting Weather Research and Forecasting (WRF) simulations: one on a historical period (2000–15) and another on a future period (2086–2100). Four setups were considered: hh (trained/tested on historical), ff (trained/tested on future), fh (trained on future/tested on historical), and hf (trained on historical/tested on future). Within-climate models (hh and ff) reproduced spatial heterogeneity more effectively, with better spatial correlations and realistic variability, while all four reproduced temporal autocorrelation. Extremes remain a challenge, yet with negative biases, in terms of hourly precipitation upper percentiles (75th, 95th, 99th, and 99.9th) and daily maxima. The hh and ff cases performed best; the fh case benefited from exposure to stronger extremes during training, while hf performed poorest due to weak generalization. Importantly, however, cross-climate model biases (hf and fh) were not substantially larger than within-climate ones. These findings highlight the challenge of training on stationary data for nonstationary climates but also demonstrate the potential of WGANs to mitigate such limitations.
    Publicationtype: Journal Article
      5  2
  • Some of the metrics are blocked by your 
    consent settings
    Item type:Publication, Access status: with files ,
    Machine unlearning: bias correction in neural network downscaled storms
    (Elsevier, 2026-02)
    Papalexiou, Simon Michael
    ;
    Mamalakis, Antonios
    Accurate precipitation at fine spatial resolutions is essential for hydrologic modeling and risk assessment, yet most precipitation data are available at coarse scales. Dynamical downscaling can improve spatial resolution but is computationally expensive while statistical downscaling struggles to reproduce high-resolution characteristics. Machine learning has been shown to offer an operational alternative for transforming coarse data into fine-scale fields, especially honoring spatiotemporal precipitation dependencies. Here, we show that combining machine learning with post-processing bias correction approaches—a form of “machine unlearning”— yields improved performance. We evaluate four machine-learning models—Linear Network (LNet), Fully Connected Network (FCNet), Convolutional Neural Network (UNet), and Wasserstein Generative Adversarial Network (WGAN)—for downscaling precipitation using synthetic benchmark storms with known marginal and spatiotemporal properties. Using synthetic fields provides full control over storm characteristics and enables rigorous evaluation. Raw outputs from all models struggle to reproduce wet/dry boundaries, statistics, and extremes, and only WGAN captures the complex spatiotemporal structure of fine-scale storms. We then apply linear and nonlinear bias corrections to enforce zeros, match the mean of positive values, and align full marginal distributions, including tails. This demonstrates that post-processing is crucial for reliable neural network outputs in operational settings. The results highlight WGAN’s potential for operational downscaling while emphasizing the need for systematic post-processing and careful validation before real-world application.
    Publicationtype: Journal Article
      30  18