Generative Physics & Industrial Tomography
This repository contains the implementation of Generative Physics-Informed Flow Matching, a framework for solving inverse problems in fluid dynamics. By treating physics as a generative process, we move from simple geometric morphing to complex 3D industrial tomography.
The repository is structured into 6 hierarchical experiments, validating the transition from standard Computer Vision to Physics-Informed Neural Networks (PINNs).
| ID | Model Name | Description | Physics/Math Basis |
|---|---|---|---|
| 01 | Toy Shape Morphing | Morphs a Circle distribution into a Square. | ODE Flow Matching |
| 02 | MNIST Digits Gen | Generates handwritten digits (0-9) via flow. | Convective Velocity Fields |
| 03 | Fashion MNIST Gen | Generates clothing items (T-shirts, Boots). | High-Dim Image U-Net |
| 04 | 2D Laminar/RANS | (Core Thesis) Detects non-convex obstructions (Stars) using Turbulence. | RANS |
| 05 | 2D Unsteady Video | Locates objects via vortex shedding frequency in video data. | Spatiotemporal PINNs |
| 06 | 3D Volumetric Wake | (Final Engine) 3D Inverse Tomography for pipe flows. | Galilean Wake Tracking |
To run the foundational "Toy Morphing" model:
# 1. Clone the repo
git clone [https://github.com/piyushpiku/convective-genai-engineering.git](https://github.com/piyushpiku/convective-genai-engineering.git)
cd convective-genai-engineering
# 2. Install dependencies
pip install -r requirements.txt
# 3. Run the model
python 01_Toy_Shape_Morphing/morphing_model.py
01. Shape Morphing (Generative)
|
04. RANS Star Detection (Physics-Informed)
|
02. MNIST Digits (Latent Space)
|
06. 3D Wake Tracking (Temporal)
|