Varda-single-1.0: deterministic data-driven weather forecasting at 1 km resolution over Switzerland's complex topography

physics.ao-ph cs.LG Alberto Pennino, Francesco Zanetta, Michele Cattaneo, Claire Merker, Radi Radev, Jonas Bhend, Louis Frey, Hugues de Laroussilhe, Oph\'elia Miralles, Carlos Osuna, Daniele Nerini, Andreas Pauling, Daniel Hupp, Ulrich Hamann, Mary McGlohon, Marti Bosch, Luca Lanzilao, Marco Arpagaus, Lukas Jansing, Daniel Leuenberger, Mark A. Liniger, Katrin Ehlert, Matthew Chantry, H\aa{}vard Homleid Haugen, Gert Mertes, Ana Prieto Nemesio, Mario Santa Cruz, Jasper Wijnands, Gabriel Moldovan, Harrison Cook, Oliver Fuhrer · Oct 1, 2026

What it does

Why it matters

It combines a 6-hourly autoregressive forecaster with a temporal downscaler to balance forecast horizon and temporal fidelity. The work demonstrates that machine learning models can match or exceed operational numerical weather prediction...

Main concern

Community signal

0 up · 0 down

AI Review AI reviewed

Plain-language introduction

Varda-single-1.0 is a deterministic, data-driven weather forecasting system built for the Alpine domain, delivering hourly forecasts at 1 km resolution via a stretched-grid Graph Transformer architecture. It combines a 6-hourly autoregressive forecaster with a temporal downscaler to balance forecast horizon and temporal fidelity. The work demonstrates that machine learning models can match or exceed operational numerical weather prediction baselines in complex topography, and it openly releases model weights and software to enable operational adoption and downstream research.

Critical review

Verdict

Bottom line

The paper presents a substantial, well-validated operational ML forecasting system with an honest appraisal of its limitations. Verified over a full year against operational analyses and station observations, the authors show their system is competitive with or improves on MeteoSwiss’ operational NWP baselines for most headline scores and variables, while explicitly acknowledging that it underestimates local wind maxima and produces overly smooth convective precipitation fields, consistent with the smoothing associated with squared-error training.

“Varda-single is competitive with or improves on MeteoSwiss’ operational numerical weather prediction baselines for most headline scores and variables.”

Pennino et al., Varda-single-1.0 · Abstract

“Despite its competitive aggregate scores, Varda-single underestimates some local wind maxima and produces overly smooth convective precipitation fields, consistent with the smoothing associated with squared-error training.”

Pennino et al., Varda-single-1.0 · Abstract

What holds up

The training curriculum is methodically staged from global ERA5 reanalysis to a 20-year kilometre-scale regional reanalysis and finally to operational analysis fine-tuning, yielding large skill gains. The two-stage forecaster-plus-downscaler design pragmatically limits error accumulation while providing hourly output, and the verification is unusually thorough, combining aggregated headline scores, spatial skill decompositions, and process-oriented case studies.

“Despite its brevity, this stage produces the single largest skill improvement of the whole curriculum: RMSE and bias both distinctly decrease for many surface variables (e.g., we see a 25% reduction in the RMSE of 2 m temperature at +60 h lead time).”

Pennino et al., Varda-single-1.0 · Section 3.5

“In order to gain insight into the model’s behaviour, we investigate three case studies in addition to the aggregated headline scores, and find particular weaknesses in Varda-single’s representation of local winds over complex terrain.”

Pennino et al., Varda-single-1.0 · Section 4

Main concerns

The primary concern is the mismatch between the model’s 6-hourly autoregressive step and its 1 km spatial resolution, which bypasses the short-timescale processes the high-resolution grid is meant to resolve. This temporal undersampling likely contributes to the systematic underprediction of wind extremes and overly smooth precipitation fields, issues the authors attribute to MSE loss and insufficient latent capacity in complex terrain. Additionally, the temporal downscaler is not fine-tuned on operational analyses, leaving a residual distribution gap for hourly variables, and the forecaster’s negative wind-speed bias is amplified when vector components are combined into speed.

“the 6-hourly temporal resolution of the forecaster does not allow us to fully exploit the high-resolution information of our 1 km grid, because of a mismatch in the characteristic spatio-temporal scales of the processes involved.”

Pennino et al., Varda-single-1.0 · Section 5

“Varda-single’s negative wind-speed bias may in part be explained by the fact that the model is trained to predict the two wind components. The MSE training objective introduces a conditional bias in both components, rewarding smaller absolute values in general.”

Pennino et al., Varda-single-1.0 · Section 4.2

Evidence and comparison

The evidence is extensive and generally persuasive: the authors compare against two operational NWP baselines and a global data-driven baseline (AIFS-single), using RMSE, ETS, station scorecards, spatial skill scores ($S(x) = 1 - \mathrm{MSE}_f / \mathrm{MSE}_b$), and object-based SAL metrics for precipitation. They fairly note that pointwise RMSE advantages partly reflect smoothing rather than genuine predictive skill, and they highlight discrepancies between analysis-based and station-based verification, showing intellectual honesty often absent in ML weather papers.

“Some of the skill as measured by RMSE is a direct consequence of training on MSE and the resulting smoothing of forecast fields.”

Pennino et al., Varda-single-1.0 · Section 4.1

“Varda-single thus approximately reproduces seasonal precipitation amounts but fails to capture the spatial structure of precipitation. This is more pronounced in summer (convection) and still considerable in winter.”

Pennino et al., Varda-single-1.0 · Section 4.3.3

Reproducibility

Reproducibility is strong. Model weights and inference configurations are released on Hugging Face with setup instructions for consumer-grade hardware, and the system is built entirely on open-source frameworks (Anemoi, PyTorch, PyTorch Geometric). Evaluation code is available in the evalml package. However, full independent reproduction of the training pipeline is hindered by reliance on a 20-year, 1 km regional reanalysis (REA-L-CH1) and operational KENDA-CH1 analyses that are not trivially accessible, and by the substantial GPU resources required for training (up to 64 NVIDIA A100 or GH200 GPUs).

“Model weights and configuration files are openly available on Hugging Face, together with setup instructions and examples for producing seamless hourly forecasts over Switzerland at 1 km resolution for lead times of several days, on consumer-grade hardware.”

Pennino et al., Varda-single-1.0 · Section 5

“Training of the operational emulator was performed on up to 16 nodes of the Balfrin cluster... Each node is equipped with 4 NVIDIA A100 GPUs... The temporal downscaler was trained on 16 nodes with 4 NVIDIA GH200 GPUs each.”

Pennino et al., Varda-single-1.0 · Appendix C

Abstract

We present Varda-single-1.0, a medium-range data-driven weather prediction system built for the Alpine domain. It provides hourly deterministic regional forecasts on a mesh of 1 km resolution and global forecasts on a 31 km mesh. The system comprises two independently trained stretched-grid Graph Transformer models with encoder-processor-decoder architecture, developed in the Anemoi framework: a 6-hourly autoregressive forecaster and a temporal downscaler reconstructing hourly forecasts between the forecaster's steps. Its training curriculum includes pre-training on ERA5 reanalysis data, followed by training on a 20-year kilometre-scale regional reanalysis, and finally fine-tuning on operational kilometre-scale analyses. Verified over one year against operational analyses and surface station observations, Varda-single is competitive with or improves on MeteoSwiss' operational numerical weather prediction baselines for most headline scores and variables. It broadly matches the skill of the high-resolution 1 km ICON-CH1-EPS control at lead times up to +33 h and generally outperforms the 2 km ICON-CH2-EPS control at lead times up to +120 h. Despite competitive aggregate scores, Varda-single underestimates some local wind maxima and produces overly smooth convective precipitation fields, consistent with the smoothing associated with squared-error training. To gain insight into the model's behaviour, we investigate three case studies beyond the aggregated headline scores, and find particular weaknesses in Varda-single's representation of local winds over complex terrain. Varda-single represents an important step in the development of high-resolution ML forecasting over complex terrain, in complementing the operational regional numerical weather prediction models of MeteoSwiss with data-driven models and in providing a pretrained model for researchers and user-specific applications.

Challenge the Review

Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.

Challenges are public to read, but only signed-in members can post them. Your challenge text is stored with your account for moderation, but usernames are not shown in the public thread.

No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.