michele007rok/StormScope_Ireland

LoRA adapter for NVIDIA's StormScope-Meteosat, fine-tuned to predict Meteosat Third Generation FCI imagery 60 minutes ahead over latitudes [51.2, 55.6] and longitudes [-11.0, -5.2].

The base model predicts 10 minutes ahead from six frames 10 minutes apart. This adapter keeps the base weights frozen and adds rank-8 LoRA matrices to the attention layers (qkv, proj) of blocks 8, 9, 10, 11, 12, 13, 14, 15 in both diffusion experts, plus the final projection to the 16 channels, so the frames it reads and predicts are 60 minutes apart.

Training windows start between 2026-06-01 and 2026-09-01; validation between 2026-09-02 and 2026-09-09. NVIDIA trained the base model on November 2024 to May 2026, so these dates are new to it.

Files

• best.pt: the adapter weights (physicsnemo save_adapter format).
• train.yaml: the exact configuration used for training.

Using it

Load the base model with the same crop and time step, apply LoRA with the same settings, then load the weights with physicsnemo.experimental.peft.load_adapter. The training code is in earth-forecasting, where train.setup_model does the first two steps.

Contains modified EUMETSAT Meteosat FCI Level 1c data. The base model is released under the OpenMDW-1.1 license; check it and EUMETSAT's data terms before reusing this adapter.