SIMD Agent

AI-native physics simulation agent.

🎯 What it does

Describe the physics you want to understand — flow, heat transfer, buoyancy, pressure drop, mixing — together with a mesh of the physical domain, and the agent picks the governing equations and turbulence model, configures discretization schemes, properties, and boundary conditions, submits the case to an OpenFOAM solver, streams residuals and flow fields back, and self-heals on solver failure. The result: design-decision support — questions that usually take a CFD specialist a week ("does this U-bend overheat at 500 K?", "what's the pressure drop through the regasifier?") become a paragraph of intent plus a geometry.

✨ Features

• 🌊 Natural-language input — describe the physics you want to understand in plain English
• 🧭 Auto solver selection — picks the right OpenFOAM solver from your prompt + mesh
• 🩹 Self-healing loop — diagnoses solver failures and retries with focused fixes
• 📡 Live streaming — residuals and 3D fields delivered live over WebSocket
• 🧠 Multi-LLM — Gemini, Vertex AI, or local Ollama
• 🔓 Open source — AGPL v3, run anywhere

🛠 Capabilities

Feature	Description
Compressible & incompressible	simpleFoam family + rhoSimple variants
Turbulence regimes	k-ε, k-ω SST, k-ω, Spalart-Allmaras
Conjugate heat transfer	solid-fluid coupling (chtMultiRegion)
Multiphase flows	VOF (experimental — interFoam family)
LLM providers	Gemini · Vertex AI · Ollama (local)
Object storage	local filesystem · Google Cloud Storage
Authentication	Neon Auth · open (no auth)

See Documentation/solvers for the full list of supported solvers.

🚀 Quick start

The agent is one of three components — the other two live in their own repos:

• simd-ai/agent (this repo) — the orchestrator
• simd-ai/runner — the OpenFOAM runner (separate process)
• simd-ai/ui — the Next.js frontend (drive the agent)

You'll need:

• Docker + Docker Compose, or Python 3.11+ for bare-metal mode
• One LLM credential: a Gemini API key, OR a Vertex AI service-account JSON, OR a local Ollama install

Clone and run the installer:

git clone https://github.com/simd-ai/agent
cd agent
./install.sh

install.sh is an interactive wizard that writes .env and either brings up the stack (Docker mode) or sets up a venv (bare-metal mode).

Preview the install wizard

Docker mode

Brings up agent + postgres + frontend + OpenFOAM runner from a single docker compose file. Once the wizard finishes, the agent is at http://localhost:8000 and the frontend at http://localhost:3000. Everything stops with docker compose down.

Bare-metal mode

Installs only the agent in a local Python venv. The wizard prints the uvicorn command at the end. You bring your own runner and (if you want a UI) frontend:

# OpenFOAM runner — clone & start separately
git clone https://github.com/simd-ai/runner
cd runner && ./run.sh
# (or point ``SIMULATION_SERVER_URL`` at a remote one)

# Frontend — clone & start separately (optional)
git clone https://github.com/simd-ai/ui
cd ui && npm install && npm run dev

Or skip the UI entirely and drive the agent through the WebSocket / HTTP API at http://localhost:8000 (see Documentation/api/).

⚙️ How it works

A FastAPI service orchestrates per-file OpenFOAM codegen with an LLM, validates the output with deterministic plugin-side rules, ships the case to a service running OpenFOAM, and streams residuals and post-processed VTK back through a WebSocket. When the solver fails, the agent diagnoses the error with a smaller LLM call and retries with focused fixes — up to seven attempts by default. This is the self-healing loop.

See Documentation/architecture for the full design, Documentation/self-healing for a walkthrough of one real failure.

🧪 Examples

Four end-to-end cases ship under examples/. Each carries its mesh, its prompt, and the generated OpenFOAM case files — so you can run the simulation directly with OpenFOAM, or watch the agent regenerate it from the prompt.

Case	Physics	Solver
u-shape-pipe	compressible inverted-U duct	rhoSimpleFoam + kOmegaSST
z-bend	transient turbulent water pipe	pimpleFoam + kOmegaSST
inner-outer-pipe	2D LN2/water counter-flow regasifier	chtMultiRegionSimpleFoam
cylindrical-cht	natural convection around a heated cylinder	buoyantBoussinesqSimpleFoam

🗺 Roadmap

See ROADMAP for what's next and what's in the idea bin. Tracking happens in GitHub Issues.

🤝 Contributing

See CONTRIBUTING. New solver plugins drop into simd_agent/solvers/<name>/ and are auto-discovered; new LLM providers drop into simd_agent/llm/<name>/. No registry edits needed.