yyiming3/gencast_distillation

GenCast Distillation

GenCast Distillation is a single-step probabilistic weather forecasting model distilled from the multi-step Google DeepMind GenCast diffusion model.

Model file

Checkpoint format:

• Framework: JAX / Haiku
• Format: Google DeepMind GenCast-compatible CheckPoint NPZ
• Resolution: 1 degree
• Pressure levels: 13
• Parameters: distilled EMA student
• SHA256: fa5d5bb970f3650cec25857eecea58529b74b3ba6a69ce9765637d2be9173247

Base model

The student was distilled from:

• Model: GenCast 1p0deg <2019
• Developer: Google DeepMind
• Source: gs://dm_graphcast/gencast/params/GenCast 1p0deg <2019.npz
• Repository: https://github.com/google-deepmind/weathernext

The original GenCast model weights are distributed under CC BY-NC-SA 4.0.

Evaluation

The model is evaluated on:

Global Weather Forecasting

Global ensemble forecasting evaluated with WeatherBench2.

Typhoon Prediction

Typhoon-track prediction evaluated against IBTrACS, with optional TempestExtremes tracking and lead-time-dependent position-error analysis.

Usage

hf download yyiming3/GenCast-Distillation \
  student_gencast_dm.npz \
  --local-dir /path/to/weights

git clone https://github.com/yyimingucl/gencast_distillation
cd genCast_distillation
conda env create -f environment.yaml
conda activate dmd_gencast

DATA_DIR=/path/to/data
MODEL_CHECKPOINT_PATH=/path/to/weights/student_gencast_dm.npz

bash long_run_js/eval_gwf.sh