Hourly downscaler for AI weather forecasts over France

Fills in the hours between the 6-hourly steps of a coarse 2 m temperature forecast, on the CERRA 5.5 km grid. Given the coarse field at both ends of a 6-hour window, how far into the window the hour is, and the same terrain maps as the spatial model, it predicts the correction to the straight line between the two steps. Each ensemble member keeps the same noise from one hour to the next, so a member's detail evolves smoothly instead of being redrawn every hour.

It starts from the spatial model EmmaScharfmann/france-temperature-downscaler (same U-Net, two extra input channels switched off at the start), which it also needs at inference time as the baseline. Code: weather-forecast-model-downscaler.

Task 6-hourly 0.25° 2 m temperature to hourly 5.5 km
Region France
Training data hourly CERRA (analyses plus their +1 h and +2 h forecasts) and ERA5, 2019
Validation 1 to 4 August 2020
Loss fair CRPS, then fine-tuned with CRPS on 2, 4, 8 and 16 pixel block averages
Parameters about 2.15 M

Results

Against CERRA over land, on the validation hours:

RMSE (K) CRPS (K) spread / error
coarse, linear in time 1.85 1.38
spatial model on that line 1.46 0.78 0.66
this model 1.19 0.62 0.74

On a 5-day AIFS-ENS forecast (run of 10 August 2025, 10 members, from ECMWF open data), scored every hour against 1,791 Météo-France stations:

RMSE (K) CRPS (K) spread / error
AIFS-ENS, linear in time 3.01 1.90 0.40
spatial model, linear in time 2.77 1.63 0.61
this model 2.43 1.39 0.65

This is a single forecast, so take the station numbers as indicative. reports/ holds the training history, the scores by hour of the window, the power spectra in space and time, and a movie of the 5-day forecast.

Use

from huggingface_hub import hf_hub_download
from downscale import hourly as hr

hourly = hr.HourlyDownscaler.load(hf_hub_download("kashif/france-temperature-hourly-downscaler", "model.pt"))
spatial = hr.load_spatial(hf_hub_download("EmmaScharfmann/france-temperature-downscaler", "model.pt"))
# forecast: (time, lead_time, ensemble, lat, lon) in kelvin, 6-hourly
coarse_hourly = hr.hourly_forecast(forecast)
fine = hr.downscale_hours(hourly, forecast, coarse_hourly.lead_time.values, samples=4, batch=8)

or from the command line:

python scripts/fetch_forecasts.py --config config_temporal.yaml --start 2025-08-10T00 --end 2025-08-10T00
python scripts/downscale_hourly.py --config config_temporal.yaml --hourly model.pt --spatial spatial.pt

config_temporal.yaml (with the config_spatial.yaml it extends) holds every setting used for training, and its header gives the two training commands.

Limitations

  • One year of hourly training data; the model overfits after about 8 epochs.
  • CERRA has a small jump every three hours, where an analysis replaces the +1 h and +2 h forecasts; the model learns from that target as it is.
  • Trained over land only; values over the sea are not meaningful.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for kashif/france-temperature-hourly-downscaler

Finetuned
(1)
this model