meteolibre-dev/meteolibre-rectified-flow

MeteoLibre Rectified Flow Model

In the folder models_shortcut/:

Model Description

• Model type: Rectified Flow Diffusion Model
• Architecture: 3D U-Net with FiLM conditioning
• Input: Meteorological data patches (12 channels + 1 lightning channels, 3D spatio-temporal)
• Output: Generated weather forecast data
• Training data: MeteoLibre meteorological dataset
• Language(s): Python
• License: MIT

Training

The model was trained using:

• Framework: PyTorch with Hugging Face Accelerate
• Optimizer: Adam (lr=5e-4) OR SOAP
• Batch size: 64
• Epochs: 200
• Precision: Mixed precision (bf16)
• Distributed training: Multi-GPU support

And there is different video exemple for the inference.

Performance summary for the first wave of shortcut model:

Performance Summary

Model	Optimizer	Steps	sat_mse	sat_psnr	sat_ssim	light_mae	light_precision	light_recall	light_f1	light_iou
RF (Run 1)	-	128	0.0952	28.5327	0.8042	0.0221	0.5482	0.6535	0.5950	-
RF (Run 2)	-	128	0.1076	27.8870	0.8000	0.0221	0.5157	0.6454	0.5724	-
Baseline	Persistence	baseline	0.2368	24.5138	0.7266	0.0154	0.6714	0.6665	0.6678	0.1023
Shortcut	Adam	16	0.0981	28.3788	0.8106	0.0216	0.6339	0.5192	0.5686	0.0791
Shortcut	Adam	64	0.0983	28.3702	0.8114	0.0207	0.6609	0.5304	0.5860	0.0791
Shortcut	Adam	128	0.0983	28.3581	0.8112	0.0208	0.6518	0.5208	0.5769	0.0791
Shortcut	SOAP	16	0.0601	30.5008	0.8663	0.0156	0.8654	0.6958	0.7710	0.0818
Shortcut	SOAP	64	0.0606	30.4786	0.8661	0.0151	0.8658	0.6879	0.7663	0.0818
Shortcut	SOAP	128	0.0605	30.4848	0.8660	0.0151	0.8635	0.6886	0.7656	0.0818

Metrics from evaluation on 64x20 elements (satellite and lightning data).