AgentCFD
AI-native computational fluid dynamics for humans and agents.

AgentCFD

AgentCFD is an open-source platform for readable, verifiable, and learning-ready computational fluid dynamics. Its first engineering focus is industrial flow: pipes and ducts, pressure loss, fluid and steam transport, heat transfer, and later reacting flow and combustion.

AgentCFD was initiated by Haoming Luo in September 2026. It follows the product principles established through AgentFEM, but it is a separate codebase with its own fluid-mechanics language, providers, validation evidence, and release cycle.

Project status: pre-alpha. The public workflow and circular-pipe reference solution are executable. Deterministic OpenFOAM circular-pipe generation, execution, mesh checks, patch-history recovery, and pressure-loss validation are experimental capabilities; every run must still earn acceptance.

Why AgentCFD

• AI-native CFD — people, scripts, GUIs, and agents operate the same explicit engineering model instead of hiding scientific intent in generated files.
• Industrial flow first — internal flow, heat transfer, steam and reacting systems come before a broad but shallow solver catalog.
• Results you can check — applicability, conservation checks, provider identity, model fingerprints, and failure evidence travel with the result.
• One run or thousands — the same workflow is designed for a single analysis, parameter campaigns, reproducible datasets, surrogates, and neural operators.
• Open provider boundary — numerical engines can evolve without changing the public model; license and runtime boundaries remain visible.
• FEM–CFD continuity — versioned exchange records prepare pressure, traction, temperature, heat flux, and mesh motion for future AgentFEM coupling.

First executable workflow

from agentcfd import Model, boundaries, fluids, geometry, outputs, procedures, studies

model = Model(
    name="water-pipe",
    study=studies.internal_flow(),
    domain=geometry.circular_pipe(length=10.0, diameter=0.05),
    fluid=fluids.newtonian(
        "water",
        density=998.2,
        dynamic_viscosity=1.002e-3,
    ),
).boundaries(
    inlet=boundaries.mean_velocity_inlet(0.02),
    outlet=boundaries.pressure_outlet(),
    wall=boundaries.no_slip_wall(),
)

result = model.step(
    procedure=procedures.steady(),
    output=outputs.standard(),
).run(provider="reference")

result.require_accepted()
print(result.quantities["flow.pressure_drop"])

The reference provider implements the Hagen–Poiseuille solution, rejects flow outside its declared laminar range, and records mass balance and an independent Darcy–Weisbach identity check.

Install

Install the published alpha from PyPI with Python 3.11 or newer:

python -m pip install agentcfd==0.1.0a2
agentcfd doctor
agentcfd demo pipe

Common pipe-loss screening is available without a CFD runtime:

agentcfd calculate pipe-loss --density 998.2 --viscosity 0.001002 \
  --length 10 --diameter 0.05 --velocity 0.02 --json
agentcfd calculate pipe-flow --density 998.2 --viscosity 0.001002 \
  --length 10 --diameter 0.05 --pressure-loss 2.56512 \
  --regime laminar --json

Gas-model screening and optional CoolProp/IF97 states use equally explicit commands:

agentcfd calculate compressibility \
  --velocity 100 --speed-of-sound 400 --json
python -m pip install "agentcfd[properties]"
agentcfd properties state \
  --fluid IF97::Water --pressure 101325 --temperature 500 --json

Both return structured records rather than presentation-only text. Installed AgentCFD/AgentCAE schemas can be discovered with agentcfd contracts --json.

For editable development from the repository:

git clone https://github.com/haoming-luo/agentcfd.git
cd agentcfd
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
agentcfd doctor
agentcfd demo pipe

Prepare the first OpenFOAM case

OpenFOAM is the primary industrial solver direction. AgentCFD keeps it behind a filesystem-and-subprocess provider boundary: the Apache-2.0 Python core writes an ordinary case, while an OpenFOAM installation already managed by the user does the GPL-licensed numerical work.

Generate a full three-dimensional O-grid pipe case without requiring OpenFOAM:

agentcfd prepare openfoam-pipe openfoam-pipe --json

Or use the public provider from Python:

from agentcfd.providers import OpenFOAMProvider

step = model.step(procedure=procedures.steady(), output=outputs.standard())
case = OpenFOAMProvider().prepare(step, "openfoam-pipe")
print(case.case_sha256)

The generated manifest binds every case file to its content hash and the public model fingerprint. Execution currently targets installations that provide blockMesh, checkMesh, and simpleFoam. AgentCFD recovers inlet/outlet flow, area-averaged pressure, pressure drop, mass imbalance, mesh-quality metrics, convergence evidence, and final native fields. Process completion alone never implies scientific acceptance.

For the canonical fully developed validation case, declare the physical inlet profile instead of silently changing a mean-velocity boundary:

model.boundaries(
    inlet=boundaries.fully_developed_velocity_inlet(0.02),
    outlet=boundaries.pressure_outlet(),
    wall=boundaries.no_slip_wall(),
)

Or run the bundled case end to end against an installed runtime:

agentcfd run openfoam-pipe openfoam-pipe --fully-developed --json

On macOS or another host without native OpenFOAM commands, use the official OpenCFD container directly (Docker remains externally managed):

agentcfd run openfoam-pipe openfoam-pipe \
  --fully-developed \
  --cross-section-cells 16 \
  --axial-cells 400 \
  --container-image opencfd/openfoam-run:2606 \
  --json

Three-grid studies can use agentcfd.verification.grid_convergence_index or consume three serialized results directly:

agentcfd prepare openfoam-pipe-grid pipe-grid \
  --cross-section-cells 8 16 32 \
  --base-axial-cells 40 \
  --json

agentcfd run openfoam-pipe-grid pipe-grid \
  --container-image opencfd/openfoam-run:2606 \
  --json

agentcfd verify grid-convergence coarse.json medium.json fine.json \
  --quantity flow.pressure_drop \
  --json

The preparation workflow creates a 0.5 m by 0.1 m benchmark with one declared fully developed inlet model, isotropic refinement ratios, per-case hashes, and an explicit plan. The grid runner verifies and executes every fresh case and writes the GCI evidence automatically. The result workflow checks that all runs completed, converged, share model and analysis identities, use the same quantity unit, and contain distinct positive dimensionless mesh cell counts before it records Richardson extrapolation, observed order, GCI, the asymptotic ratio, and hashes of all three source files. The rationale for analytical, flow-rate, and boundaryFoam inlets, measured resolution/runtime tiers, and the staged turbulence and steam plan is recorded in the numerical strategy. Common pipe checks are available under agentcfd.engineering: hydraulic diameter, Reynolds number, laminar or iterated Colebrook--White Darcy friction, straight-run pressure loss, local-loss pressure drop, and one auditable pipe_pressure_loss record combining them. The transitional Reynolds range is deliberately rejected rather than silently interpolated.

Architecture

Human / AI agent / script / future GUI
                    |
                    v
Study -> Model -> Domain/Regions -> Fluids -> Boundaries/Sources
                    |
                    v
             Solution Step + Output
                    |
                    v
      Provider lowering and deterministic execution
                    |
                    v
 SimulationResult -> verification -> datasets -> learning
                    |
                    v
      AgentFEM / experiments / NN / PINN / neural operators

The core package has no mandatory third-party runtime dependency or LLM and does not treat successful execution as scientific acceptance. NumPy support is available through the optional arrays extra. AI is a first-class operator of the workflow; fluid mechanics and deterministic numerical computation remain authoritative.

Documentation

• Concepts
• Workflow
• Changelog
• Architecture
• Roadmap
• Product and market strategy
• AgentFEM, CFD, and AI interoperability
• Results, evidence, and AI exchange
• Thermophysical properties and IF97
• Engineering correlations and model screening
• Installation and solver runtime
• Dependency and license policy
• Thermophysical properties
• OpenFOAM provider boundary
• Publishing and PyPI name status
• Validation policy
• Engineering correlations
• Benchmark catalog
• OpenFOAM v2606 execution evidence
• Guide for AI agents

License

AgentCFD is licensed under Apache-2.0. Optional numerical engines retain their own licenses and are connected through explicit provider boundaries.