Implementation of a Neural Network Model using Jaxmat Library and Export to MFront

[marius-duvillard]

We implement a Neural Network model with the newly developped Jaxmat librairy. Jaxmat is based on JAX and Python, and we intend to use it in conjunction with MFront.

Since JAX-based models are not directly exportable to C++, one approach is to convert the model to TensorFlow and utilize the TensorFlow C++ API or the cppflow library. Initially, we will test a non-linear elastic model by exporting the predicted stress and tangent operators.

[marius-duvillard]

Here is a repository with a simple tutorial on how to integrate a Neural Network for elastic behavior:

https://github.com/marius-duvillard/jax2mfront

Enjoy!

More work needs to be done to include Neural Generalized Standard Material (NGSM) models.

[marius-duvillard]

I have just added a new commit introducing the training of an NGSM for viscoelasticity using relaxation loading paths. At this stage, the associated MFront implementation relies on an explicit time integration scheme. Although this is a simplified approach, the results obtained so far are already quite satisfactory.