hybrid_ml_coupling.md

layout	default
title	Hybrid PyTorch-SU2 Coupling
description	Examples of coupling SU2 with external ML libraries (PyTorch) via the Python Wrapper.
group	Workflow Features

Hybrid Physics-ML Coupling with PyTorch

This tutorial demonstrates how to establish a bidirectional coupling between the SU2 solver and PyTorch using the standard Python Wrapper (pysu2).

Unlike offline training (where data is saved to disk and trained later), this workflow allows for online (in-situ) learning, where a neural network is trained on flow physics as the simulation progresses.

Goals

1. Execute the SU2 solver step-by-step via Python.
2. Extract flow variables (e.g., RMS_DENSITY) from memory in real-time.
3. Train a PyTorch surrogate model inside the MPI loop.
4. Demonstrate the "Physics → ML" data bridge.

Resources

The complete scripts and configuration files for this tutorial are available in the SU2 Tutorials Repository.

Prerequisites

• SU2 compiled with MPI and Python Wrapper support:
  • -Dwith-mpi=enabled
  • -Denable-pywrapper=true
• Python packages: mpi4py, torch, numpy

The Coupling Loop

The core of the workflow is the CSinglezoneDriver, which allows Python to control the C++ solver iteration-by-iteration.

# Initialize Driver with MPI
driver = pysu2.CSinglezoneDriver("inv_NACA0012.cfg", 1, MPI.COMM_WORLD)
driver.Preprocess(0)

# Training Loop
for i in range(n_iterations):
    # 1. Run Physics (SU2)
    driver.Run()
    
    # 2. Extract Data (C++ -> Python)
    physics_state = driver.GetOutputValue("RMS_DENSITY")
    
    # 3. Train Model (PyTorch)
    # ... standard pytorch forward/backward pass ...
    
    # 4. (Optional) Feedback
    # Update flow boundary conditions or deformations based on ML prediction.