What is the best way to deploy a trained neural network as an FMU for co-simulation?

[eymenipek]

I have a trained feedforward neural network (built in PyTorch) that serves as a surrogate model for a physical component and it is available in ONNX.

My questions

Is there a recommended toolchain for converting a trained ONNX model into a standalone C/C++ FMU without embedding a full inference runtime? (e.g., ONNX-to-C code generators)
For simple architectures, is the "export weights + hand-code forward pass in C" approach considered production-viable, or is there a more robust pattern?
Has anyone successfully deployed an ONNX Runtime–based FMU on a HiL target and, if so, how did you handle the runtime dependency?
Are there commercial or open-source tools that solve this cleanly?

Any guidance on best practices, toolchains, or pitfalls would be greatly appreciated.

[chrbertsch]

There are several ways to deploy ONNX NNs as FMUs, including some in-House solutions that I cannot share here.

Here are some publicly available ways to do it:

• Do you know ONNX2FMU? https://github.com/HyRES-FBK/onnx2fmu
• if you have MATLAB Simulink, you could import the FMU there https://de.mathworks.com/help/reinforcement-learning/ug/import-existing-networks-using-onnx.html, https://de.mathworks.com/videos/integrate-tensorflow-model-into-simulink-for-simulation-and-code-generation-1649142078929.html and create a standalone FMU from there
• if you have a Modelica tool, you could import the ONNX model via https://github.com/xrg-simulation/SMArtIInt and create an FMU from there

Please note that with CS-FMUs you will get a timestep delay (if you are not in event mode in FMI 3.0 FMUs), so perhaps ME FMUs would be more suitable for static models.