3 min read

WeatherGenerator Highlights 2025

WeatherGenerator Highlights 2025

2025 marked the first year for the WeatherGenerator project, with major advances across all areas - from model architecture to expanded datasets and growing community engagement. Let’s have a look at some of the project’s highlights of the year:

Data

  • Secured compute time on three major HPC systems: Jupiter, Leonardo, and Alps
  • Successfully onboarded diverse datasets including ERA5, CERRA, SEVIRI, and OPERA for model training
  • Integrated gridded analysis data from NWP and observations covering Norway (Met Norway)
  • Started data preparation for polar orbiting image data (AVHRR and VIIRS) and regional modeling with radar and satellite data (SMHI)
  • Developed custom data readers to integrate internal datasets for hydropower applications (Statkraft)
  • Implemented Kerchunk and Vzarr methods for efficient data reading (FZJ)
  • Optimized data reading efficiency across NetCDF, Zarr, and Kerchunk formats (FZJ)
  • Developed EBCC as novel error-bound compression method for weather data, with paper published on ArXiv and submitted to PNAS https://arxiv.org/abs/2510.22265 (ETHZ/SPCL)
  • Implemented Zarr plugin for EBCC compression (ETHZ/SPCL)
  • Converted energy datasets into .netcdf/anemoi formats for fine-tuning applications (Buluttan)

Model

  • Achieved highly flexible model training across wide range of datasets (ERA5, CERRA, SEVIRI, OPERA)
  • Implemented fine-tuning capabilities for pre-trained models on SYNOP stations
  • Significantly increased forecast performance through minor, computationally efficient architecture modifications (FZJ)
  • Extended model for pre-training with latent space masked token modeling
  • Developed latent space diffusion model for forecasting
  • Implemented 2D RoPE Positional Encoding for improved spatial understanding (FZJ)
  • Developed first version of LoRA (Low-Rank Adaptation) for efficient fine-tuning (FZJ)
  • Implemented PerceiverIO embedding layer for handling diverse input types (FZJ)
  • Added Mixture of Experts (MoE) layer implementation (FZJ)
  • Explored channel selection on ERA5 for optimized model efficiency (FZJ)
  • Cleaned up model code for better generalization and backward compatibility (KNMI)
  • Tested different masking and forecasting strategies (KNMI)
  • Modified adapter engine to separate masked and unmasked tokens (KNMI)
  • Implemented latent state as scratch space with register tokens (KNMI)
  • Developed functionality to convert WeatherGenerator output into standard formats (Met Norway)

Application

  • Applied WeatherGenerator to reanalysis application (AP11) with first successful outputs (Met Norway)
  • Published paper “High-Resolution Probabilistic Data-Driven Weather Modeling with a Stretched-Grid”: https://arxiv.org/abs/2511.23043 (Met Norway)
  • Designed and partially implemented scalable evaluation framework supporting 400+ wind/solar farms (Buluttan)
  • Achieved first successful end-to-end pipeline execution (Buluttan)
  • Integrated .zarr format into production workflows (Buluttan)
  • Successful application to Swedish National Space Agency for project promoting WeatherGenerator use among Swedish governmental organizations (SMHI)

Communication and Dissemination

Hackathon

In August, we hosted an internal hackathon that brought together partners from across the consortium. The event provided a valuable first encounter with the WeatherGenerator code for many participants, offering hands-on experience with in-house datasets and HPC workflows. The hackathon served as an excellent opportunity for partners to meet, collaborate, and gain practical insights into getting started with the project.

Looking ahead

As we enter 2026, the WeatherGenerator project continues to grow and evolve. Building on this year’s achievements, we are excited to expand our capabilities, broaden our applications, and strengthen our community. The foundation is set for an impactful year ahead.