The AI engineer that
actually runs your CFD.
IteraSim Eddy is a desktop agent that reads your entire OpenFOAM case — geometry, mesh, boundary conditions, solver logs — then builds it, meshes it and runs it. When the solver diverges at 3 a.m., Eddy diagnoses why, patches the dictionary and restarts it. Unattended.
Free tier forever · No credit card · Runs on your machine, not our cloud
CFD's bottleneck was never compute.
Every plane, reactor, chip and pump is designed with computational fluid dynamics. Yet engineers spend most of a project on setup, debugging and documentation — not engineering. The tooling assumes you already know the answer.
Setup is archaeology
You describe physics in a directory of cryptic text dictionaries. One wrong keyword in fvSchemes and nothing tells you until hour three of a run.
Failure is silent
The solver dies with a floating-point exception and no explanation. You read a log, guess a discretisation scheme, re-run, and lose a day. Then you do it again.
Expertise is the gate
The commercial suites cost tens of thousands per seat. OpenFOAM is free and takes six months to learn. General-purpose AI can't help — it can't see your mesh, your BCs or your logs.
A domain model, not a chatbot wrapper.
Eddy is grounded in the case on your disk and in a retrieval system built specifically over OpenFOAM's source and documentation. It answers with your values — your viscosity, your patch names, your residual history — and it has the shell access to act on them.
From a folder to a converged result.
Eddy drives the real OpenFOAM pipeline through a real shell. Nothing is simulated, nothing is mocked — it runs the same commands you would, and shows you every one of them.
Point it at a case
Any folder with system/, constant/ and 0/ is recognised instantly. Or start from blank — import an STL or STEP and let the Case Builder write the dictionaries.
Describe the physics
"Set up a Boussinesq buoyant case for this geometry with air at 300 K." Eddy drives the builder for you and previews every file change before it writes a byte.
Run the pipeline
blockMesh, snappyHexMesh or Gmsh, then checkMesh, decomposePar, mpirun. Residuals stream live. Locally on your cores, or submitted to your Slurm cluster.
It fixes what breaks
On divergence, Eddy reads the log, identifies the numerical cause, patches the dictionary, and re-runs — with a full history you can revert file by file.
Every step of the workflow, in one window.
Three panes: your case files, the workspace you're working in, and a copilot that can see both. No alt-tabbing between a text editor, a terminal and a browser full of forum posts.
Write a valid case without touching a dictionary.
Nine solvers ship with a complete template. Pick one and Eddy walks you through every dictionary it needs — physics, per-patch boundary conditions matched to what the solver actually accepts, a sane fvSchemes/fvSolution preset, and runtime control. It writes each file in turn with a live counter, and if one fails it stops and tells you exactly which.
Mesh it, then verify the mesh actually contains your geometry.
Import STL or STEP straight from SolidWorks, Fusion or CATIA — STEP is converted in one click through Gmsh. Choose blockMesh for structured boxes, snappyHexMesh for real geometry, or Gmsh for unstructured. The live preview and checkMesh run before you commit a single core-hour to a solver.
- Interactive 3D preview with CAD-style orbit, pan and zoom
- checkMesh parsed into plain English, not a wall of numbers
- Geometry containment check — an empty mesh never reports success
Watch it converge — or watch it fix itself.
Residuals, Courant number and wall-clock stream into the console as the solver runs. Parallel runs are one checkbox: Eddy handles decomposePar with scotch, launches mpirun and reconstructs at the end. If the run dies, the agent reads the log, names the numerical cause, and proposes the patch — you approve it, or let it run unattended.
- Clickable compiler and solver errors that jump to the source line
- Per-case history — revert any AI-made change to a previous version
- Submit to Slurm or LSF over SSH and stream the remote log back live
Results you can defend, not just pictures.
Field plots render in the app the moment a run finishes, and any .foam file opens in ParaView with the case pre-loaded and a sensible view already set. Eddy writes the report too — what was solved, which schemes, which mesh, which residual history — so the documentation isn't the thing you do last and badly.
Everything else in the rail.
Project Navigator
A live file tree over your cases, solvers, meshes and results. Watches the filesystem — edits made by ParaView or a cluster run appear within a second. Right-click an STL for FreeCAD, a .foam for ParaView.
Solver Workshop
Clone any OpenFOAM solver into your workspace, edit the C++ with jump-to-definition, and compile with wmake streaming into the panel below. Highlight a block and ask the copilot to add a source term.
Workflow Wizard
A linear pipeline — geometry, mesh, physics, solve, post — that refuses to let you advance until the previous step validates. The fastest way to learn the real order of operations, and to build repeatable templates.
Compute Engine
Add an SSH profile and run on your cluster instead of your laptop. Rsyncs only what changed, submits Slurm or LSF from a template, streams the remote log home, pulls results back. Inherits your ~/.ssh/config, ProxyJump chains included.
Cloud & versioning
Optional sync to S3, Backblaze B2, Google Drive or your own SFTP/WebDAV — for moving between machines or handing a case to a colleague. Roll an entire case back to how it was three versions ago without leaving the app.
Bring your own model
On Free, plug in a Google Gemini, Anthropic or OpenAI key — or point at a local Ollama model and run the AI entirely offline. Keys are encrypted in your OS keyring, never in plaintext, never on our servers.
Surrogate training
Turn the runs you've already completed into a dataset, then train a neural surrogate — FNO, DeepONet or a PINN — to predict new geometries in seconds instead of hours. Every case you run makes the next one cheaper.
Copilot with case context
The chat on the right can read every file in the open case. Ask why the pressure residual stalls, what a patch type means for your solver, or to explain the dictionary someone else wrote three years ago.
Environment detection
Eddy finds your OpenFOAM installations, compilers and MPI at launch and ranks them by what's actually compiled — not by what the folder is called. The version it picked is always visible in the status bar.
Built for physics, not for autocomplete.
A CFD failure is not a syntax error. The wrong discretisation scheme for your cell Péclet number, a bad mesh patch, an under-relaxation factor that needs to drop — none of that can be inferred from public code. It lives in the heads of people who have debugged thousands of cases.
- Retrieval
- Qdrant + LlamaIndex over a hand-curated OpenFOAM corpus — source, documentation and tutorial cases.
- Knowledge
- A CFD expert base encoding six years of real failure modes and their fixes.
- Execution
- A PTY bridge, so the agent drives genuine shell sessions rather than pretending to.
- Models
- Gemini 2.5 and Claude with automatic routing and multi-provider fallback — one outage doesn't stop your work.
- ML
- PyTorch neural operators (FNO, DeepONet) and PINNs for the surrogate trainer.
- App
- Electron, React and xterm.js on the desktop; Python and FastAPI behind it.
Your geometry stays yours
Cases, meshes and results never leave your machine. Only the specific snippets the copilot needs to answer a question go to the model provider — and the providers we route through are configured not to train on them.
Works offline
The entire local workflow — builder, blockMesh, snappyHexMesh, Gmsh, the solver, the ParaView handoff — runs with no internet at all. Point the copilot at a local Ollama model and even the AI goes offline.
Air-gap ready
For aerospace, defence and energy groups that cannot send anything off-site, Enterprise ships an on-premise or fully air-gapped deployment with a private knowledge base over your own documents.
Validated against physics with a published answer.
An agent that confidently produces a wrong flow field is worse than no agent. So Eddy is checked against benchmarks where the correct answer is already in the literature — and against the failure modes that are notorious for diverging quietly.
Lid-driven cavity at Re = 1000
The standard verification problem in CFD. A correct run reproduces Ghia, Ghia & Shin (1982) along the vertical centreline to well under one percent, with the primary vortex centred near (0.53, 0.56).
- Solver
- icoFoam / simpleFoam
- Agreement
- < 1% vs. Ghia et al. on the centreline profile
- Why it matters
- Run a case with a known answer first, and every later surprise is information about your problem — not your installation.
buoyantSimpleFoam with Boussinesq buoyancy
A configuration notorious for silently diverging — the kind of case that eats an afternoon before you work out it was never going to converge with those settings in the first place.
- Behaviour
- Eddy diagnoses the crash, repairs the case and completes the run without a human in the loop
- Transparency
- Every patch is shown as a diff before it is written, and every change is revertible
- Status
- Verified end to end on real cases during pre-alpha
Wherever a flow decides the design.
Aerodynamics
External flow, drag and lift studies, wind loading — RANS and LES on real CAD geometry.
Multiphase & thermal
Boiling, condensation, free surfaces, conjugate heat transfer and buoyancy-driven convection.
Process engineering
Mixers, reactors, pipe networks, porous media and particle-laden flow in plant equipment.
Energy & nuclear
Thermal-hydraulics, breeder-blanket and coolant-loop studies on air-gapped infrastructure.
The app is free. You pay for the brain.
IteraSim never charges you to run Eddy itself — the whole local OpenFOAM workflow is free forever. Paid plans cover managed AI routing and the curated OpenFOAM knowledge base, so you don't have to wire up API keys or budget tokens.
The full desktop app, running entirely on your machine.
- Runs entirely on your machine
- Your geometry never leaves your network
- Full OpenFOAM workflow automation
- Copilot with your own API key or Ollama
- Meshing — blockMesh, snappyHexMesh, Gmsh
- ParaView integration
- 50,000 managed AI tokens / month
- Managed AI keys (no setup)
- Team seats & shared cases
Managed models and the curated OpenFOAM knowledge base.
- Everything in Local
- Curated OpenFOAM knowledge base (RAG)
- 5M managed tokens / month
- Priority Gemini 2.5 / Claude routing
- No API-key setup — we handle it
- Cancel anytime, one click
Hosted knowledge base and metered keys for a whole group.
- Everything in Local
- Curated OpenFOAM knowledge base (hosted)
- Managed AI keys — no setup, metered
- Priority model routing
- Email support, 1 business day
- Seat management & billing Roadmap
- Shared team case library Roadmap
- Usage analytics per seat Roadmap
For groups whose data cannot leave the building.
- Everything in Team
- On-premise / air-gapped deployment
- Dedicated onboarding & training
- Commercial use with indemnification
- Private knowledge base, your own docs Roadmap
- SSO / SAML Roadmap
- Audit log & SLA Roadmap
All plans include the full desktop app and commercial use. Cancel anytime. Academic and student pricing available — just ask.
We taught this before we automated it.
IteraSim runs an international workshop series on OpenFOAM and CFD–machine-learning integration: forty hours of curriculum, delivered to engineers and researchers who use it the next day. The product exists because the same questions kept coming up — and almost none of them were about physics.
Built by the people who write the solvers.
IteraSim's domain layer isn't borrowed from anywhere. Behind it sit four OpenFOAM solvers written from scratch in C++: a coupled CFD–DEM framework validated against our own experiments, a Volume-of-Fluid solver with microscale phase change for Leidenfrost dynamics in non-Newtonian fluids, a heterogeneous LTNE porous-media solver with radiation, and a multi-region tritium-transport solver. The work is published in Physics of Fluids, and an earlier paper was selected for the front cover of ACS Industrial & Engineering Chemistry Research.
The company exists because writing those solvers meant six years of spending more hours fixing broken simulations than doing engineering. The failure mode never changed: the solver dies, you read a log, you guess which scheme to swap, you re-run, you lose a day. Eddy is that loop, closed — with the tacit knowledge that normally takes six months to acquire written into the product instead of into a notebook.
Research continues at ETH Zürich: foundation models for multiphase physics using neural operators, trained on a thousand-case high-fidelity CFD dataset generated on HPC clusters. That is where Eddy's surrogate layer comes from, and it is why every case a customer runs makes the next one cheaper.
One file. No root. No dependencies.
Eddy is the interface; OpenFOAM is the engine. Install any recent OpenFOAM (v2306 or newer) and Eddy detects it at launch — the version it picked is shown in the status bar.
A single AppImage — nothing is installed system-wide, and your settings and cases live outside the file so upgrades never touch them.
Download for Linux · 172 MB- Download the AppImage above.
- Make it executable and run it — that is the whole install.
- Eddy finds your OpenFOAM on first launch.
A single .exe installer that runs the backend inside WSL2. On first launch the environment wizard finds your OpenFOAM installation and your compilers for you.
Download for Windows · 113 MB- Download the installer above.
- Double-click it and follow the prompts.
- The wizard sets up WSL2 and finds OpenFOAM for you.
Not officially supported yet — it's on the roadmap. Leave us your address and we'll tell you the day the first build ships.
Notify meBoth builds are the full app — the free tier needs no licence key. Verify your download with sha256sum iteraSim-Eddy-1.0.1.AppImage and compare it against the hash above. All releases and release notes →
Requirements
A 64-bit Linux or Windows machine, 8 GB RAM or more, and OpenFOAM v2306 or newer (the ESI-OpenCFD fork is fully supported; the Foundation fork works for most cases). ParaView and FreeCAD are optional — Eddy can install them for you.
First launch
Create an account, and the environment wizard detects your OpenFOAM installations, compilers and MPI — ranking them by what is actually compiled, not by what the folder is called. Then open a case, or build one.
Start with a benchmark
Run the lid-driven cavity first. It has a published answer, so it tells you whether a surprising result later is your setup or the solver. The user guide walks it through screen by screen.
The things everyone asks first.
Is my case data ever sent to IteraSim's servers?
Do I need OpenFOAM installed separately?
Which OpenFOAM versions are supported?
Does Eddy need an internet connection?
Can I use Eddy without paying?
The copilot changed a file and now my case won't run. How do I undo it?
Can I run in parallel on multiple cores?
How do I import geometry from SolidWorks, Fusion 360 or CATIA?
Does it work on macOS?
Can I use Eddy for commercial and consulting work?
What happens to my work if I stop paying?
Let the agent run the simulation.
Download Eddy, point it at a case you already have, and ask it what the case actually does. That first answer takes about thirty seconds.
Free forever on the Local tier · Linux & Windows · info@iterasim.in