Super-Resolving Coarse-Resolution Weather Forecasts with Flow Matching

ArchesWeatherSR is a flow matching–based generative super-resolution model for global weather forecasts. It takes a coarse-resolution (1.5°) forecast and generates an ensemble of plausible high-resolution (0.25°) fields, recovering fine-scale variability while preserving large-scale structure. As demonstrated on ArchesWeatherGen forecasts in the companion paper.

For more information, see the geoarches repository and documentation.

Installation

We recommend using uv to manage the environment. After cloning the repo, run:

git clone https://github.com/dataymeric/ArchesWeatherSR.git
cd archesweathersr
uv sync

uv sync installs all dependencies and the archesweathersr package itself in editable mode, which is required so that imports resolve correctly when running the training and inference scripts.

Training

Configuration is managed as in geoarches, using Hydra. The entry point is train.py.

Basic run

python train.py \
  module=archesweathersr \
  dataloader=era5downscaling-hdf5 \
  ++name=my_run

Data

ERA5 data was obtained from WeatherBench2. We recommend downloading the data in HDF5 format for use with the dataloaders.era5_hdf5 dataloader. We provide a small download script with scripts/dl_era.py.

Inference

Pretrained model

A pretrained model is available on Hugging Face. Download the weights with:

hf download dataymeric/ArchesWeatherSR --local-dir runs/archesweathersr

Then run inference pointing to that directory:

python train.py mode=test ++name=archesweathersr

Super-resolving ArchesWeatherGen forecasts

archesweathersr.inference.infer_forecasts super-resolves forecasts produced by ArchesWeatherGen. Each run processes one time slice (identified by --task-id) across all input files:

python -m archesweathersr.inference.infer_forecasts --task-id 0

To process multiple time slices in parallel (e.g. as a SLURM job array), pass $SLURM_ARRAY_TASK_ID as the task ID:

# run_sr.sbatch
#!/bin/bash
#SBATCH --array=0-599

python -m archesweathersr.inference.infer_forecasts --task-id $SLURM_ARRAY_TASK_ID

We provide a script to produce ArchesWeatherGen rollouts in the correct format for super-resolution with scripts/rollout_archesweathergen.py. This requires downloading the pretrained models.

Citation

@preprint{delefosse2026archesweathersr,
  title         = {Super-Resolving Coarse-Resolution Weather Forecasts With Flow Matching},
  author        = {Delefosse, Aymeric and Charantonis, Anastase and B{\'e}r{\'e}ziat, Dominique},
  year          = {2026},
  eprint        = {2604.00897},
  archiveprefix = {arXiv},
  primaryclass  = {cs.LG},
  doi           = {10.48550/arXiv.2604.00897}
}