Varda-single-1.0

Varda-single-1.0 is MeteoSwiss' medium-range, data-driven weather forecasting system for the Alpine domain. It produces hourly deterministic regional forecasts on a 1 km mesh over Switzerland and the surrounding region, together with global forecasts at 31 km resolution, complementing MeteoSwiss' operational ICON-CH1/CH2-EPS numerical weather prediction systems.

Varda-single-1.0 is built with Anemoi, the open-source framework for data-driven weather forecasting co-developed by ECMWF and a growing community of European national meteorological services.

Example Varda-single-1.0 forecast initialised on 2025-04-06 00 UTC, at +12 h lead time, showing 2 m temperature and 10 m wind speed from the global 31 km mesh down to the 1 km regional mesh over the Alps.

Table of contents

• Quickstart
• Model overview
• Evaluation
• Training details
• Citation

---

Quickstart

Run Varda-single-1.0 yourself

The notebook demonstrates end to end how to:

• retrieve the latest initial conditions from MeteoSwiss' and ECMWF's open data platforms,
• run the 6-hourly forecaster followed by the 1-hourly temporal downscaler with anemoi-inference,
• and plot the resulting forecast against ICON-CH1 for comparison.

A full run takes about 20 minutes end to end, including dependency installation and fetching initial conditions.

Hardware requirements

Running the model needs about 30 GB of GPU memory and 14 GB of system RAM per GPU device. Internet access is required to install dependencies and fetch initial conditions.

Before running on a hosted notebook platform, you need to configure your session with the appropriate hardware:

• Kaggle — Session options → Accelerator → GPU T4 ×2, and Internet → on.
• Colab — Runtime → Change runtime type → A100. Requires Colab Pro; the free T4 (16 GB) will run out of memory.

---

Model overview

Model description

Varda-single-1.0 consists of two independently trained Graph Transformer encoder-processor-decoder models, operating on a stretched grid (a global 31 km mesh refined to 1 km over Switzerland and the surrounding region):

• a 6-hourly autoregressive forecaster that predicts the atmospheric state at 6-hourly steps, and
• a 1-hourly temporal downscaler that reconstructs the five intermediate hourly states between any two consecutive 6-hourly forecasts (plus a sixth step, to correctly handle accumulated diagnostics such as precipitation).

This two-stage design trades off temporal resolution against error accumulation: an autoregressive model operating directly at 1-hourly steps would accumulate errors six times faster over a 120-hour forecast, whereas the temporal downscaler, conditioned on both endpoints of each 6-hour window, does not propagate errors forward.

• Developed by: MeteoSwiss
• Model type: Encoder-processor-decoder Graph Transformer, stretched-grid (global 31 km / regional 1 km)
• License: Model weights are published under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/.
• Paper: Varda-single-1.0: deterministic data-driven weather forecasting at 1 km resolution over Switzerland's complex topography — arXiv:2610.01835.

Model Sources

• Repository: Anemoi is an open-source framework for creating data-driven weather forecasting systems, co-developed by ECMWF and national meteorological services across Europe.
• Paper: Pennino, Zanetta, Cattaneo, Merker, Radev, Bhend, Frey, de Laroussilhe, Miralles, Osuna, Nerini, Pauling, Hupp, Hamann, McGlohon, Bosch, Lanzilao, Arpagaus, Jansing, Leuenberger, Liniger, Ehlert, Chantry, Haugen, Mertes, Prieto Nemesio, Santa Cruz, Wijnands, Moldovan, Cook, Fuhrer — Varda-single-1.0: deterministic data-driven weather forecasting at 1 km resolution over Switzerland's complex topography (arXiv:2610.01835).

---

Evaluation

Varda-single-1.0 was verified over one year (April 2025 – March 2026) against MeteoSwiss' operational analyses (KENDA-CH1) and SwissMetNet surface station observations, and compared against the operational ICON-CH1-EPS (1 km, up to +33 h) and ICON-CH2-EPS (2 km, up to +120 h) baselines.

Scorecards comparing Varda-single-1.0 against the ICON-CH1-CTRL (short range, left) and ICON-CH2-CTRL (medium range, right) operational baselines, stratified by region and lead time. Blue indicates Varda-single-1.0 performs better, red indicates the ICON baseline performs better; dot size encodes the magnitude of the relative difference.

---

Training details

Training data

Varda-single-1.0 is trained through a four-stage curriculum, transferring from global to regional, and from reanalysis to operational analyses:

1. Global pre-training on ERA5 (1979–2023) at 31 km resolution, on a uniform global graph.
2. Stretched-grid training on a cutout combining ERA5 with REA-L-CH1, a 20-year kilometre-scale regional reanalysis produced with the ICON model over the MeteoSwiss forecasting domain (2005–2025, 1 km, hourly).
3. Rollout training, extending the forecaster's autoregressive rollout window to reduce error accumulation over longer forecast horizons.
4. Operational fine-tuning on ECMWF IFS operational analyses (global) and MeteoSwiss' KENDA-CH1 operational analyses (regional, 1 km), which closes the distribution gap between the 20-year reanalysis and the operational analyses the model is initialised from at inference time.

The temporal downscaler goes through the first two stages only, and is therefore trained only on reanalysis data. It needs no rollout training, since it is not autoregressive, and does not undergo the operational fine-tuning stage, since global operational analyses are not available at hourly resolution.

Training procedure

Training stages of the 6-hour forecaster and the 1-hour temporal downscaler. The corresponding anemoi-training configs are in varda-forecaster-det-sgm/configs/training/ and varda-temporal-downscaler-det-sgm/configs/training/, numbered by stage.

---

Citation

If you use this model in your work, please cite:

Pennino, A., Zanetta, F., Cattaneo, M., Merker, C., Radev, R., Bhend, J., Frey, L., de Laroussilhe, H., Miralles, O., Osuna, C., Nerini, D., Pauling, A., Hupp, D., Hamann, U., McGlohon, M., Bosch, M., Lanzilao, L., Arpagaus, M., Jansing, L., Leuenberger, D., Liniger, M. A., Ehlert, K., Chantry, M., Haugen, H. H., Mertes, G., Prieto Nemesio, A., Santa Cruz, M., Wijnands, J., Moldovan, G., Cook, H., & Fuhrer, O. (2026). Varda-single-1.0: deterministic data-driven weather forecasting at 1 km resolution over Switzerland's complex topography. arXiv preprint arXiv:2610.01835.

BibTeX:

@misc{pennino2026vardasingle,
  title={{Varda-single-1.0}: deterministic data-driven weather forecasting at 1 km resolution over {Switzerland}'s complex topography},
  author={Pennino, Alberto and Zanetta, Francesco and Cattaneo, Michele and Merker, Claire and Radev, Radi and Bhend, Jonas and Frey, Louis and de Laroussilhe, Hugues and Miralles, Oph{\'e}lia and Osuna, Carlos and Nerini, Daniele and Pauling, Andreas and Hupp, Daniel and Hamann, Ulrich and McGlohon, Mary and Bosch, Marti and Lanzilao, Luca and Arpagaus, Marco and Jansing, Lukas and Leuenberger, Daniel and Liniger, Mark A. and Ehlert, Katrin and Chantry, Matthew and Haugen, H{\aa}vard Homleid and Mertes, Gert and Prieto Nemesio, Ana and Santa Cruz, Mario and Wijnands, Jasper and Moldovan, Gabriel and Cook, Harrison and Fuhrer, Oliver},
  year={2026},
  eprint={2610.01835},
  archivePrefix={arXiv},
  primaryClass={physics.ao-ph},
  url={https://arxiv.org/abs/2610.01835}
}