Skip to content

Dataset-configured native model suite

NAVIER-CFD 0.5 provides 52 executable native reference models spanning operator learning, geometry-aware deep learning, transformer and graph models, physics-informed machine learning, generative field models, learned correctors, preconditioners, adaptation, and uncertainty.

Dataset as an import argument

from navier_cfd import load_model

model = load_model("fno", dataset="pdebench")
model = load_model("transolver", dataset="airfrans")
model = load_model("gino", dataset="drivaerml")
model = load_model("p3d", dataset="scalarflow")

The dataset profile determines default:

  • dimension;
  • representation type;
  • input and output channels;
  • coordinate dimension;
  • sensor count;
  • Fourier modes;
  • hidden width;
  • number of layers;
  • attention, graph, field, coordinate, or branch/trunk dispatch;
  • normalization description;
  • channel layout.

A real canonical sample overrides dataset defaults:

model, plan = load_model(
    "pibert",
    dataset="realpdebench",
    sample=sample,
    overrides={"hidden_dim": 256, "num_layers": 8},
    return_plan=True,
)

print(plan.to_dict())

Resolution order:

  1. Explicit model keyword arguments.
  2. User overrides.
  3. Actual CFDSample shapes.
  4. Registered dataset defaults.

Dataset profiles

Dataset Representation Default dimension Typical native families
PDEBench structured 2 FNO, PINO, PIBERT, transformer operators
CFDBench structured 2 FNO, PINO, U-NO, PIBERT
RealPDEBench structured 2 PIBERT, PINO, DeepONet, temporal operators
The Well structured 3 FNO, P3D, latent and state-space operators
APEBench structured 2 autoregressive operators and refiners
ScalarFlow structured 3 FNO, P3D, convolutional and generative operators
AirfRANS point cloud 2 Transolver, GINO, MeshGraphNets, Geo-FNO
DrivAerNet++ point cloud 3 GINO, Transolver, AeroTransformer, DoMINO reference
DrivAerML unstructured 3 GINO, MeshGraphNets, geometry transformers
ShapeNet-Car point cloud 3 geometry-conditioned models and preconditioners
EAGLE unstructured 3 graph, transformer, and geometry operators

Defaults are visible and reviewable. Dataset releases can differ in field names, ordering, resolution, and target definitions, so production experiments should provide an actual sample and explicit adapter keys.

Native reference inventory

Operator learning

  • DeepONet
  • MIONet
  • Fourier-DeepONet
  • Nested Fourier-DeepONet
  • Fourier-MIONet
  • Fourier Neural Operator
  • Physics-Informed Neural Operator
  • Geo-FNO
  • Geometry-Informed Neural Operator
  • U-FNO
  • Factorized FNO
  • U-shaped Neural Operator
  • Latent Spectral Model
  • General Neural Operator Transformer
  • Galerkin Transformer
  • Multiwavelet Transformer
  • FactFormer
  • Orthogonal Neural Operator
  • Transolver
  • Laplace Neural Operator
  • State-Space Neural Operator

Physics-informed ML

  • PINN
  • NSFnets
  • PINNsFormer
  • PINO
  • PIBERT
  • PI-MFM
  • RiemannONet
  • DeepM&Mnet

Geometry, graph, transformer, and foundation-style learning

  • MeshGraphNets
  • DoMINO reference
  • Universal Physics Transformer
  • DPOT
  • Poseidon
  • PROSE-FD
  • BCAT
  • PDEformer-1
  • P3D
  • AeroTransformer
  • Tadpole
  • ReViT

Generative, corrector, preconditioning, adaptation, and uncertainty

  • FourierFlow reference
  • PDE-Refiner reference
  • Solver-in-the-Loop corrector
  • Indirect Neural Corrector
  • NeuroSEM corrector reference
  • Neural-operator preconditioned Newton reference
  • Geometry-aware neural preconditioner
  • Conformalized-DeepONet reference
  • TANTE-style adaptation
  • Energy Transformer reconstruction reference
  • FunDiff reference
  • Flow Matching for PDEs reference

PICT, diffSPH, and NeuralDEM remain external integrations because meaningful execution requires their dedicated differentiable CFD or particle runtimes. NAVIER-CFD does not replace those solvers with misleading generic field networks.

Native reference versus official reproduction

A NAVIER-CFD native reference implementation is:

  • importable from the PyPI package;
  • executable with PyTorch;
  • connected to the common dataset API;
  • trainable through CFDTrainer;
  • checkpointable;
  • forward- and backward-tested;
  • covered by adapter conformance tests.

It is not automatically:

  • the exact author repository revision;
  • a reproduction of unpublished preprocessing;
  • a copy of private checkpoints;
  • numerically identical to every paper table;
  • licensed to redistribute third-party weights.

Each object includes a navier_reference_notice and should be benchmarked against the official source before reproduction claims are made.

Conformance testing

from navier_cfd import validate_model_adapter

report = validate_model_adapter(
    "transolver",
    sample,
    dataset="airfrans",
)

assert report.passed

The report checks registry status, dependency availability, dataset-specific construction, parameter count, forward pass, output compatibility, and backward propagation.

Experiment API

from navier_cfd import Experiment, TaskSpec, TrainerConfig

experiment = Experiment(
    dataset_id="cfdbench",
    model_id="pino",
    task=TaskSpec(
        problem="cylinder",
        task_type="forecasting",
        dimension=2,
        mesh_type="structured",
        temporal_mode="autoregressive",
        geometry_mode="fixed",
        physics=("incompressible_navier_stokes",),
    ),
    trainer_config=TrainerConfig(
        epochs=100,
        optimizer="adamw",
        mixed_precision=True,
    ),
)

result = experiment.run(raw_dataset)

The experiment manifest records the dataset profile, actual sample-derived configuration, explicit overrides, model parameters, split seed, trainer settings, checkpoint path, and CFD metrics.