[Resource] Hybrid PyTorch-SU2 Coupling with Python Wrapper (Example & Docs)

[ayush4874]

Hi everyone,

I've been working on setting up an online (in-situ) coupling between SU2 and PyTorch using the Python Wrapper, and I noticed there wasn't a standalone example for this workflow in the tutorials yet.

I've successfully prototyped a bidirectional loop where SU2 runs within an MPI process, and a PyTorch model extracts flow physics (like RMS_DENSITY) in real-time.

To help others getting started with Physics-ML coupling, I've submitted the scripts and documentation to the official repositories:

• Tutorial Code: su2code/Tutorials#71 (Adds train_online.py and configs)
• Documentation: su2code/su2code.github.io#189 (Explains the CSinglezoneDriver loop)

Current Limitation & Next Steps:
Currently, this workflow is excellent for extraction (Physics -> ML). However, I've noticed that injecting updated weights back into the C++ MLPCpp module during runtime is tricky due to encapsulation.

I am currently prototyping a C++ patch to allow dynamic weight injection (Online Learning), which I hope to propose for GSoC 2026.

I'd love any feedback from those working on data-driven surrogates!

Best,
Ayush

[EvertBunschoten]

Hi Ayush.
Thank your for your enthusiasm in the topic! Using SU2 for the training of physics-informed neural networks is a really nice opportunity.
I ran your tutorial case and for my understanding, here you train a model to evaluate the value of the density residual based on the flow iteration? It's a bit unclear to me what you want to achieve there, so maybe you can provide an example of what the target application will be.

[ayush4874]

Hi Evert, thanks for running the case!

You are completely right, mapping Iteration -> Residual has no physical value. I just used that as a minimal "Hello World" example to prove the software architecture works (ensuring the Python driver can extract data from SU2 over MPI without crashing).

My actual target application is Data-Driven Turbulence Modeling (Closure Modeling). The goal is to use the neural network to predict closure terms (like eddy viscosity) and update them on-the-fly to match experimental data (Inverse Design/Data Assimilation).

This is the PR I just opened on MLPCpp (PR #4). The tutorial showed we can read from SU2; the PR enables us to write (update the model) to do the actual turbulence work.