maxiimilian/otprof-sqfh

OTProf SQF-H (nl)

Vertical profiles of optical turbulence (log10 Cn2) and the WRF MYNN turbulence quantities (QKE, TSQ, EL_PBL) from ERA5 pressure-level data, with the height-aware SqueezeFormer (SQF-H) presented in arXiv:2604.09346. Fried parameter r0 and scintillation index follow from the predicted Cn2 profile.

Use

from otprof.models import sqf_h
from otprof.predict import load_model, predict
from otprof.sources import era5

predictor = load_model(sqf_h.HF_REPO_ID, revision=sqf_h.HF_REVISION)   # pip install "otprof[torch,hub]"
X = predictor.make_inputs(era5.standardize(ds))   # ERA5 from otprof.sources.era5.fetch
y = predict(predictor, X)

Inputs: ERA5 on the 31 pressure levels of otprof.schema (1000 to 20 hPa), from which the model computes its inputs (otprof.models.sqf_h.inputs.make_inputs, features version 1.0.0, schema version 1.0.0; predict refuses inputs of a different MAJOR version). Outputs at the query heights above ground (default: the 99 mean WRF target levels of the training data, 8 to 29043 m); the output quantile mapping (post_y) is applied unless predict(..., oqm=False).

Training

• Domain: Netherlands (nl), WRF 2 km with MYNN; targets: WRF Cn2 and turbulence profiles on the model levels.
• Decoupled training: WRF interpolated to the ERA5 pressure levels as inputs ("WRF-PL"), run on ERA5 at inference.
• Split by year: train 2019-2020, validation 2018, test 2017. Experiment config: experiment.yaml.
• Output quantile mapping (NL-oQM): one map of log10 Cn2 for the whole column, 100 sine-spaced quantiles, fitted on the ERA5-driven predictions of the training years against WRF.
• Training dataset, this model and the code: Zenodo 10.5281/zenodo.23733024.

Performance

The performance of the model is documented in the paper (arXiv:2604.09346). Please refer to it before using the predictions.

Files

config.json (architecture, versions, variable order, default heights), model.safetensors (weights), pipeline.json + steps.json (scalers, target transforms, oQM), experiment.yaml. No pickle.

Citation

arXiv:2604.09346