F1 Car Aerodynamic Design - AI-Assisted Pipeline

An open-source, AI-assisted workflow for designing efficient Formula 1 cars using parametric geometry (Blender), multi-backend CFD simulation, and ML surrogate models.

Baseline reference: Red Bull RB19/RB20 design philosophy, FIA 2024 regulations.

Architecture

Parameter --> Blender --> STL --> Backend Router --> Force Extraction --> Results JSON
Selection     Geometry    Export      |                                       |
                                     +---> Omniverse Flow (GPU CFD)          |
                                     +---> OpenFOAM (CPU CFD)                +---> ML Training
                                     +---> Modulus PINN (~1ms)               |     (Autoresearch)
                                     +---> ML Surrogate (GP/MLP)             |
                                     +---> Empirical Estimate                +---> Active Learning

Directory Structure

f1-design-ai/
├── blender/
│   └── f1_car_generator.py       # Parametric F1 car geometry (Blender 4.x/5.x)
├── scripts/
│   ├── run_pipeline.py           # Pipeline orchestrator (5 backends + ML + active learning)
│   ├── omniverse_sim.py          # NVIDIA Omniverse Flow GPU CFD backend
│   ├── convert_to_usd.py         # STL -> USD conversion for Omniverse
│   ├── modulus_surrogate.py      # NVIDIA Modulus PINN surrogate (F1AeroNet)
│   ├── openfoam-docker.sh        # OpenFOAM Docker wrapper (Apple Silicon compatible)
│   └── install.sh                # macOS tool installer (Homebrew-based)
├── ml/                           # ML surrogate pipeline
│   ├── surrogate.py              # 4 model types: Linear, MLP, GP, PINN
│   ├── data_prep.py              # Data loading, normalization, train/test split
│   ├── experiment.py             # Experiment runner + model comparison
│   ├── autoresearch_config.py    # Karpathy's autoresearch pattern
│   ├── active_learning.py        # GP uncertainty-driven parameter proposal
│   └── models/                   # Trained model artifacts (.pkl, .pt)
├── openfoam/
│   └── f1_baseline/              # Complete OpenFOAM case (simpleFoam, k-omega SST)
│       ├── 0/                    # Initial/boundary conditions
│       ├── constant/             # Physical properties & mesh
│       └── system/               # Solver settings
├── autoresearch/                 # Autoresearch experiment tracking
│   ├── seed_paper.md             # Research problem description
│   ├── run.py                    # Experiment entry point
│   └── experiment_log.json       # Results from 11 experiments
├── results/
│   └── sweep_all.json            # 32 empirical parameter sweep results
├── config.yaml                   # Central configuration (backends, physics, ML, paths)
├── requirements.txt              # Python dependencies
├── CLAUDE.md                     # Architecture & conventions (source of truth)
├── DESIGN.md                     # Architecture decisions & rationale
├── EXPERIMENT.md                 # Parameter sweep results & ML experiment tracking
├── RESEARCH.md                   # Literature references & research directions
└── NEXT-TO-DO.md                 # Prioritized backlog

Quick Start

# 1. Install tools (macOS)
chmod +x scripts/install.sh && ./scripts/install.sh
pip install -r requirements.txt

# 2. Generate baseline F1 car geometry
blender --background --python blender/f1_car_generator.py

# 3. Run pipeline (choose a backend)
python3 scripts/run_pipeline.py --estimate-only              # Empirical (instant, no tools needed)
python3 scripts/run_pipeline.py --backend openfoam           # OpenFOAM CPU CFD
python3 scripts/run_pipeline.py --backend omniverse          # NVIDIA Omniverse GPU CFD
python3 scripts/run_pipeline.py --backend surrogate          # Best trained ML model

# 4. Parameter sweeps
python3 scripts/run_pipeline.py --sweeps all --estimate-only

# 5. Train ML surrogates (Karpathy's autoresearch pattern)
python3 scripts/run_pipeline.py --train-surrogate

# 6. Active learning (propose + simulate new data points)
python3 scripts/run_pipeline.py --active-learn 5 --backend omniverse

# OpenFOAM via Docker (if not installed locally)
./scripts/openfoam-docker.sh blockMesh
./scripts/openfoam-docker.sh simpleFoam

Simulation Backends

The pipeline supports 5 backends with automatic fallback:

Priority	Backend	What it does
1	Omniverse Flow	NVIDIA GPU-accelerated CFD
2	OpenFOAM	CPU CFD (simpleFoam, k-omega SST, Docker fallback)
3	Modulus PINN	Physics-informed neural network (~1ms inference)
4	ML Surrogate	Best trained model from ml/models/
5	Empirical	Correlation-based estimates (always available)

ML Surrogate Models

Model	Strengths	Test R² (32pts)
PINN	Physics constraints, best generalization	0.85-0.92
GP	Uncertainty estimates for active learning	0.51-0.99 (uneven)
MLP	Nonlinear, PyTorch + CUDA/MPS	0.02-0.93
Linear	Fast, interpretable	-1.83-0.76

PINN is the best overall surrogate at small dataset sizes. See EXPERIMENT.md for full results.

Parametric Design Space

Variable	Default	Range	Unit
ride_height	0.030	0.020 - 0.050	meters
front_wing_angle	14.0	10 - 20	degrees
rear_wing_angle	16.0	10 - 22	degrees
diffuser_angle	12.0	6 - 18	degrees
sidepod_undercut	0.15	0.08 - 0.20	meters

Tool Stack

Tool	Role
Blender 4.x/5.x	Parametric geometry via Python API
OpenFOAM v2406+	CFD simulation (simpleFoam, k-omega SST)
NVIDIA Omniverse	GPU CFD via Flow extension
NVIDIA Modulus	Physics-informed neural network surrogates
PyTorch	ML model training (CUDA/MPS/CPU)
GPyTorch / scikit-learn	Gaussian Process surrogates
Docker	OpenFOAM containerized fallback

Key Research References

• Ravelli & Savini (2018) — F1 CFD with OpenFOAM (5.7% drag error vs wind tunnel)
• DrivAerNet++ (NeurIPS 2024) — 8,150 car designs + CFD dataset
• F1 Front Wing PINNs — Neural network CFD surrogate (R²=0.968)
• Karpathy's Autoresearch (2026) — Autonomous ML research pattern
• AI Design Agents (ArXiv 2025) — Multi-agent aero design framework

Documentation

File	Contents
CLAUDE.md	Architecture, commands, conventions (source of truth)
DESIGN.md	Architecture decisions with rationale
EXPERIMENT.md	Parameter sweep results, ML experiment tracking
RESEARCH.md	Literature references, research directions
NEXT-TO-DO.md	Prioritized backlog (P1-P7)

Current Status

• 32 empirical data points (no CFD validation yet)
• 4 trained ML surrogates (PINN best at MSE=0.0038)
• 11 autoresearch experiments completed
• OpenFOAM case configured with coarse mesh for first run
• Pipeline end-to-end validation in progress