Skip to content

The Well

The Well is a provider family of large, diverse physics-simulation datasets. NAVIER-CFD integrates it as a first-class data provider while preserving NAVIER-CFD's broader purpose: multi-dataset integration, multi-family model construction, unified training, physical evaluation, evidence-aware recommendation, and chemical-engineering extensions.

Correct access model

The Well is not one monolithic datasets.load_dataset("polymathic-ai/the_well") repository. Its official interface uses the_well.data.WellDataset with a provider base path and an individual dataset name:

from navier_cfd import load_cfd_dataset

dataset = load_cfd_dataset(
    "the_well",
    configuration="active_matter",
    split="train",
    streaming=True,
    n_steps_input=4,
    n_steps_output=1,
    use_normalization=True,
)

Equivalent official provider arguments are:

WellDataset(
    well_base_path="hf://datasets/polymathic-ai/",
    well_dataset_name="active_matter",
    well_split_name="train",
)

Install the optional provider package with:

pip install "navier-cfd[the-well]"

For local data, set streaming=False and pass local_path:

dataset = load_cfd_dataset(
    "the_well",
    configuration="rayleigh_benard",
    split="valid",
    streaming=False,
    local_path="/data/the_well",
)

Canonical adaptation

Official records use fields shaped as:

\[ [T, L_1, \ldots, L_d, F]. \]

By default, NAVIER-CFD converts them to:

\[ [L_1, \ldots, L_d, T F], \]

so time history becomes model input channels while the physical spatial dimension remains correct. The adapter preserves:

  • field names and tensor order;
  • input and output time grids;
  • spatial coordinates;
  • constant scalars;
  • boundary-condition metadata;
  • official split;
  • provider version and access plan;
  • normalization provenance.
sample = dataset[0]
model, plan = load_model(
    "fno",
    dataset="the_well",
    sample=sample,
    return_plan=True,
)

The actual sample determines the number of dimensions, fields, input-history channels, output channels, coordinates, and recommended model configuration.

Official splits and normalization

Use The Well's official train, valid, and test splits. Do not randomly split overlapping windows from the same trajectories. The provider supports official Z-score or RMS normalization; physics metrics should normally be evaluated after restoring physical units.

Scope statement

NAVIER-CFD uses The Well as a high-quality data and benchmark provider. It does not claim ownership of The Well datasets, reproduce the full 15 TB collection, or replace the official package. Cite The Well and the selected dataset source in addition to NAVIER-CFD.

References