NAVIER-CFD 0.6 · executable scientific ML

Pass the dataset. Import the model. Evaluate the physics.

NAVIER-CFD connects 52 executable native reference models, 11 dataset profiles, official provider adapters, unified training and checkpoints, physical metric suites, and evidence-aware recommendation for CFD and chemical-engineering surrogate modelling.

See the Python API Explore metrics View GitHub

52 native reference models 55 catalog models 11 dataset profiles 4 metric suites The Well provider

Dataset-aware import

The dataset determines the model configuration.

The loader resolves dimension, channels, coordinates, representation, spectral modes, graph or attention settings, and forward dispatch. A real sample overrides static defaults.

Official The Well access: NAVIER-CFD uses the_well.data.WellDataset with hf://datasets/polymathic-ai/ and an individual well_dataset_name. It is not treated as one monolithic Hugging Face dataset repository.

from navier_cfd import load_cfd_dataset, load_model

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

sample = well[0]
fno, plan = load_model(
    "fno",
    dataset="the_well",
    sample=sample,
    return_plan=True,
)

Native model coverage

Neural operators, deep learning, and physics-informed ML under one API.

Every native reference model is executable, trainable, shape-tested, backward-tested, and connected to the same dataset, trainer, checkpoint, and evaluation layers.

Operator learning

DeepONet, FNO/PINO, Geo-FNO, GINO, U-NO, GNOT, Transolver, spectral and transformer operators.

Physics-informed ML

PINN, NSFnets, PINNsFormer, PINO, PIBERT, PI-MFM, RiemannONet, DeepM&Mnet and related variants.

Geometry and foundation models

MeshGraphNets, GINO, DoMINO, UPT, DPOT, Poseidon, PROSE-FD, PDEformer, P3D and AeroTransformer.

Generative, correction and UQ

PDE-Refiner, FourierFlow, solver correctors, preconditioners, conformal prediction, adaptation and flow matching.

Scientific scope: these are working NAVIER-CFD reference architectures, not automatic claims of bit-for-bit reproduction of every author repository, private checkpoint or paper table.

Physical evaluation

Accuracy, spectrum, dynamics and physics—not one global error number.

Named suites combine standard field metrics with ideas adopted from The Well and RealPDEBench. Each result records its equation, direction, ideal value, requirements, assumptions and evaluation space.

Data standard

MSE, RMSE, MAE, L∞, relative L1/L2, NMSE, NRMSE, R², Pearson correlation and cosine similarity.

The Well compatible

Variance-scaled errors and Parseval-consistent low-, middle- and high-frequency spectral MSE.

RealPDEBench style

fRMSE, temporal frequency error, turbulent kinetic-energy error, mean velocity-profile error and Update Ratio.

Fluid physics

Divergence, vorticity, kinetic-energy and spectral errors with explicit velocity, grid and axis metadata.

from navier_cfd import MetricContext, MetricSuite

context = MetricContext(
    time_axis=1,
    spatial_axes=(2, 3),
    velocity_channels=(0, 1),
    spacing=(dx, dy),
    profile_axis=2,
)

suite = MetricSuite.combine([
    "data_standard",
    "the_well",
    "fluid_standard",
])
results = suite.evaluate(prediction, target, context=context)

Unified experiment system

One reproducible path from provider data to physically evaluated results.

Provider or raw dataset→Canonical adapter→Model configuration→Trainer→Checkpoint→Metric suites→Manifest

Choose the right model before spending compute.

Use the evidence-aware recommender, then train and evaluate the selected model with the target dataset and physics-aware metric suite.

Launch recommender