OpenFOAM is the open source industry standard for computational fluid dynamics. It is a C++ framework and solver collection maintained by OpenFOAM Foundation and OpenCFD Ltd that handles incompressible and compressible flow, heat transfer, turbulence, multiphase flow, and combustion. It is used across automotive, aerospace, energy, and maritime industries. The SciRouter API wraps it so you can run simulations without standing up an HPC cluster.

What is NVIDIA Apollo and why does it matter for CFD?

NVIDIA Apollo is a suite of physics-ML tools announced at SC25. It provides neural network CFD surrogates that achieve 500x or greater speedups over classical solvers on well-defined problem classes by training on large datasets of CFD runs. It is built on the NVIDIA Modulus framework and integrates with OpenFOAM meshes. For parametric studies and design optimization, Apollo-class surrogates turn week-long CFD campaigns into hour-long ones.

What is OpenFOAMGPT 2.0?

OpenFOAMGPT 2.0 is a language-model wrapper that translates natural language descriptions of a simulation problem into a complete OpenFOAM case setup. You describe the geometry, fluid, boundary conditions, and outputs you want, and it generates the solver selection, mesh specification, and configuration files that would otherwise take hours to write by hand. The SciRouter API exposes a hosted version so you can use it without maintaining local LLM infrastructure.

When should I use a neural surrogate vs classical CFD?

Use a classical OpenFOAM run when you need authoritative physics accuracy, when the geometry or conditions are novel, or when the output will be published or used for certification. Use a neural surrogate when you are sweeping a parameter space, optimizing a design, or running active learning loops where throughput matters more than absolute accuracy. The standard pattern is to use the surrogate for exploration and classical CFD for final validation.

Can I run a full aerodynamic optimization through the API?

Yes. The physics API supports parametric geometry, surrogate-based optimization, and classical CFD validation in a single workflow. You define a parameterized shape, specify an objective (minimize drag subject to lift constraint), and the API runs an optimization loop that uses the surrogate for proposal generation and OpenFOAM for high-fidelity checks on top candidates.

What hardware runs the simulations?

Classical OpenFOAM jobs run on CPU clusters sized by problem complexity. Neural surrogates run on GPU workers. The API handles the scheduling automatically: you specify the problem and the accuracy you need, and the router picks the right backend. Billing is per-core-hour for classical runs and per-call for surrogate runs.

OpenFOAM and Neural CFD Surrogates via API

Computational fluid dynamics is the unglamorous workhorse of industrial engineering. Every new airplane wing, every race car aerodynamic package, every wind turbine, every HVAC system, every semiconductor fab airflow analysis goes through CFD at some point. The total addressable market is enormous, and so is the compute bill: a single high-fidelity aerodynamic study for a car can consume tens of thousands of core-hours on an HPC cluster.

Two developments are changing the economics. First, OpenFOAM, the open source CFD framework, has matured to the point where it handles the same workloads as expensive commercial packages for most use cases. Second, neural CFD surrogates (most notably NVIDIA Apollo, launched at SC25) deliver speedups of 500x or more on parametric problems. Put them together behind a hosted API and CFD stops being something that requires an HPC admin to run.

SciRouter's physics API wraps OpenFOAM, Apollo-class surrogates, and OpenFOAMGPT 2.0 behind a single endpoint. You describe a problem; the API picks the right solver; you get back a result.

Note

Neural surrogates are excellent for exploration and optimization but should be validated with classical CFD before any decision that affects safety, certification, or production. The API is designed to make the surrogate-then-classical workflow seamless.

OpenFOAM: the open source standard

OpenFOAM (Open Source Field Operation And Manipulation) has been the industry standard for open source CFD since the mid-2000s. It is a C++ framework that provides solvers for almost every flow regime: incompressible RANS and LES for aerodynamics, compressible and shock-capturing solvers for high-speed flows, VOF multiphase solvers for free surface problems, combustion solvers for engines, and conjugate heat transfer for thermal management.

Most commercial CFD users (in automotive, aerospace, energy, and maritime industries) run OpenFOAM alongside proprietary alternatives because the cost difference is compelling for preliminary design and parametric studies. The bottleneck has historically been operational: keeping an HPC cluster, tuning mesh parameters, and selecting solvers are all expertise-heavy tasks.

NVIDIA Apollo and the neural CFD breakthrough

Neural CFD surrogates have been a research topic for years, but NVIDIA Apollo (launched at SC25 in late 2025) was the first end-to-end production-grade toolkit from a major vendor. It is built on NVIDIA Modulus and provides several components:

Pretrained surrogates for common problem classes (external aero, HVAC, pipe flow, heat exchangers).
Training infrastructure for fine-tuning a surrogate on your own geometry family.
OpenFOAM integration so surrogates read the same meshes and boundary conditions as classical solvers.
Uncertainty quantification through ensemble surrogates, so you know when to trust the prediction.

The reported speedups are impressive: Synopsys announced 500x improvements on chip cooling workflows, and several automotive OEMs reported similar numbers on external aero. Those speedups are what make real-time design exploration tractable.

OpenFOAMGPT 2.0: natural language case setup

OpenFOAMGPT 2.0 is the other half of the modern workflow. It is a language-model wrapper that takes a natural language description of a simulation problem and generates a complete OpenFOAM case directory: mesh generation scripts, solver selection, boundary conditions, turbulence model, time stepping, and post-processing configuration.

This is not a gimmick. Setting up an OpenFOAM case by hand takes hours even for an experienced CFD engineer because the configuration is spread across dozens of files with subtle conventions. OpenFOAMGPT collapses that to a paragraph of description. The API exposes it as a pre-processing step that you can call before running the actual simulation.

A CFD simulation call

cfd-simulate.py

import httpx

API_KEY = "sk-sci-..."
BASE = "https://scirouter.ai/v1"

response = httpx.post(
    f"{BASE}/physics/cfd-simulate",
    headers={"Authorization": f"Bearer {API_KEY}"},
    json={
        "solver": "openfoam",
        "case_description": (
            "External aerodynamics of a generic sedan car body at "
            "30 m/s freestream velocity. Report drag coefficient, "
            "lift coefficient, and surface pressure distribution."
        ),
        "geometry_url": "s3://my-bucket/sedan.stl",
        "mesh_quality": "medium",
        "turbulence_model": "k-omega-sst",
        "use_surrogate": True,
        "surrogate_then_validate": True,
    },
    timeout=3600,
)

result = response.json()
print(f"Drag coefficient (surrogate): {result['surrogate']['cd']:.3f}")
print(f"Drag coefficient (OpenFOAM): {result['classical']['cd']:.3f}")
print(f"Runtime: surrogate {result['surrogate']['runtime_s']}s, "
      f"classical {result['classical']['runtime_s']}s")

The call first runs the neural surrogate (sub-second) and then kicks off a high-fidelity OpenFOAM run as validation. You get both results in the same response so you can see the agreement. Parametric studies use only the surrogate path and complete in minutes rather than days.

Aerodynamic design optimization

The canonical use case for fast CFD surrogates is shape optimization. You parameterize a design (wing camber, spoiler angle, duct cross-section) and use an optimizer to explore the parameter space, evaluating each candidate with the surrogate. What would take weeks on a classical solver completes in hours.

The API includes an optimization wrapper that handles this pattern end to end. You specify the parameterization, the objective (minimize drag subject to lift and stability constraints), and the optimization method (Bayesian optimization, CMA-ES, gradient-based), and the service runs the loop. Top candidates are validated with classical CFD before being returned.

Thermal management and HVAC

Beyond aerodynamics, CFD is ubiquitous in thermal management for electronics and buildings. The API supports conjugate heat transfer for PCB-level cooling analysis, natural and forced convection for room-scale HVAC, and reactor flow for chemical processing. Apollo-class surrogates exist for common thermal workloads, making iterative design practical.

Multiphase and free surface

Multiphase flow (air-water, oil-water, sloshing tanks, marine hydrodynamics) is historically the most expensive CFD regime because it needs fine grids to track interfaces. OpenFOAM handles this well with its VOF solvers. Neural surrogates for multiphase are earlier in development, so the API falls back to classical solvers for these cases while still providing the convenience layer (GPU-accelerated meshing, case setup, post-processing).

Tip

For marine and offshore applications, running a short coupled surrogate-plus-OpenFOAM workflow on early-stage designs can catch fundamental issues before they reach tank testing. The cost of a failed tank test is orders of magnitude more than a CFD campaign.

Case studies and TAM

The industrial users of CFD are concentrated in a few high-value sectors:

Automotive. Every new vehicle program includes dozens of aero and thermal CFD runs. Surrogate-accelerated design loops are a competitive advantage.
Aerospace. Engine and airframe design has always been CFD-heavy; neural surrogates speed up exploration of novel configurations.
Energy. Wind turbines, offshore platforms, and nuclear safety analyses all rely on CFD.
Semiconductors. Chip cooling design is where Synopsys reported Apollo speedups because the problem class is perfectly suited to surrogate training.
Consumer products. Fans, HVAC equipment, kitchen appliances all go through thermal CFD during product development.

Across these sectors, analysts estimate a multi-billion-dollar annual spend on CFD tooling and compute. An API that reduces the cost per simulation by an order of magnitude addresses a real gap.

Getting started

The fastest path is Physics Lab, which provides a web interface for case setup, geometry upload, and result visualization. You can describe your problem in plain English and let OpenFOAMGPT generate the initial case files, then run the simulation with one click.

For production integrations, the Python SDK handles geometry upload, result polling, and VTK download. CFD workflows in industrial R&D groups are increasingly structured as automated pipelines, and an API-first tool fits that pattern naturally.

Open Physics Lab →