Rethinking Forward Processes for Score-Based Data Assimilation in High Dimensions
Paper โข 2604.02889 โข Published
Pretrained score models for MASF: Measurement-Aware Score-based Filter, from Rethinking Forward Processes for Score-Based Nonlinear Data Assimilation in High Dimensions.
Code: https://github.com/tcnllab-oss/masf ยท Data: eunbii1/masf
| file | system | measurement | config |
|---|---|---|---|
checkpoints/kolmogorov_128/grid_mask.pt |
Kolmogorov flow 128ร128 | grid mask (stride 10) | configs/kolmogorov_128/grid_mask.yaml |
checkpoints/kolmogorov_128/center_mask.pt |
Kolmogorov flow 128ร128 | center mask (hole ratio 0.5) | configs/kolmogorov_128/center_mask.yaml |
checkpoints/kolmogorov_128/speed.pt |
Kolmogorov flow 128ร128 | speed | configs/kolmogorov_128/speed.yaml |
checkpoints/kolmogorov_256/grid_mask.pt |
Kolmogorov flow 256ร256 | grid mask (stride 15) | configs/kolmogorov_256/grid_mask.yaml |
checkpoints/kolmogorov_512/grid_mask.pt |
Kolmogorov flow 512ร512 | grid mask (stride 20) | configs/kolmogorov_512/grid_mask.yaml |
checkpoints/lorenz96/cubic.pt |
Lorenz96 (dim 40, F=8) | cubic | configs/lorenz96/cubic.yaml |
Each checkpoint is pretrained on the 1000 prior states in the dataset repo. The model architecture is defined by the matching config in the code repo. During filtering, the model is fine-tuned online at every measurement step.
git clone https://github.com/tcnllab-oss/masf && cd masf
hf download eunbii1/masf --include "checkpoints/*" --local-dir .
hf download eunbii1/masf --repo-type dataset --include "datasets/*" --local-dir .
python main.py -d kolmogorov_128 --measurement_type grid_mask --seed 42
Files land in checkpoints/ and datasets/, where the configs expect them. See notebooks/ for examples.
@article{yoon2026masf,
title={Rethinking Forward Processes for Score-Based Nonlinear Data Assimilation in High Dimensions},
author={Yoon, Eunbi and Chang, Won and Kim, Donghan and Kim, Dae Wook},
journal={arXiv preprint},
year={2026}
}