Erik-Wikingsson/dawis-checkpoints

DAWIS pretrained checkpoints

Pretrained models for the paper DAWIS: Data Assimilation with Windowed Inverse Sampling via Multitask Interpolants. Code: github.com/Erik-Wikingsson/DAWIS.

The folder layout matches what the code expects under MODELS_ROOT, so download the repository and point MODELS_ROOT at it:

hf download Erik-Wikingsson/dawis-checkpoints --local-dir /path/to/models
# then in .env:  MODELS_ROOT="/path/to/models"

File	Dataset	Used by
DAWIS/models/eta_channel_init_6-FMW-224-06_28_18-3280/last.ckpt	SQG, 64×64	DAWIS window model, --init_states 6
DAWIS/models_SEVIR/SEVIR_FLOWDAS_SPLIT_lr_vil_eta_channel_init_6_flowdas-FMW-224-09_12_05-1441/last.ckpt	SEVIR VIL, 128×128	DAWIS window model, --init_states 6
SQG/models/daisi/daisi_64.pth	SQG, 64×64	DAISI prior
SQG/models/daisi/daisi_sevir_128.pth	SEVIR VIL, 128×128	DAISI prior
SQG/models/flowdas/flowdas_sqg_3hrly.pt	SQG, 64×64	FlowDAS forecaster (FlowDAS baseline, --forward_model flowdas)

The window models are PyTorch Lightning checkpoints reduced to state_dict and hyper_parameters (the training arguments, from which the code reads the architecture); optimizer state is not included. The DAISI priors are plain state dicts. The SQG FlowDAS forecaster is the backbone from DAISI, stored as {"model": state_dict, "step", "val_loss"} without optimizer state.

The SEVIR FlowDAS forecaster is the pretrained checkpoint released by FlowDAS (Chen et al., 2025) and is not redistributed here. Download it into the same layout:

curl -L -o /path/to/models/SQG/models/flowdas/flowdas_sevir.pt \
    "https://www.dropbox.com/scl/fi/5z1bwfdvbztnums9deqhe/latest.pt?rlkey=o5izt721am3hzkcwjmmn7joym&dl=1"

or set FLOWDAS_SEVIR_MODEL_PATH to wherever you saved it.

Citation

@article{wikingsson2026dawis,
  title         = {{DAWIS}: Data Assimilation with Windowed Inverse Sampling via Multitask Interpolants},
  author        = {Wikingsson, Erik and Andrae, Martin and Landelius, Tomas and Lindsten, Fredrik},
  journal       = {arXiv preprint arXiv:2610.03314},
  year          = {2026},
  eprint        = {2610.03314},
  archivePrefix = {arXiv},
  primaryClass  = {stat.ML},
  url           = {https://arxiv.org/abs/2610.03314}
}