Using a Neural Network to Replace the Return-Mapping Algorithm in Crystal Plasticity Models

[maddymak9956]

Hello everyone,

I am working on integrating a neural network–based constitutive model into a finite element solid mechanics solver (MOOSE framework). The goal is to replace the conventional iterative stress update / return-mapping procedure used in crystal plasticity models with a neural network surrogate.

In classical crystal plasticity, the constitutive update at each integration point requires solving nonlinear equations for slip rates across multiple slip systems, typically using an iterative Newton-type procedure. This can become computationally expensive, especially for simulations with many integration points and complex hardening laws.

My idea is to train a neural network (using PyTorch) on data generated from a conventional crystal plasticity model. The network would take strain increment and ISV's at previous time step as inputs, and predicts the slip increment thereby bypassing the iterative solver.

Can you please suggest how can I integrate/use neural network in MOOSE solid mechanics app for this?

[GiudGiud]

@dschwen @dewenyushu @sapitts any suggestions?

[GiudGiud]

You'll want to look into the libtorch integration in MOOSE. I think you may want to train the neural networks during regular return mapping solves, feeding it the inputs and the solved output, and at some point switch to the trained NN instead of the return mapping solves

[chaibhave]

NEML2 is probably the way to go, I think? @hugary1995

[sapitts]

I concur with the suggestion to explore NEML2

[hugary1995]

I'm not sure about that. @maddymak9956 I suggest you to first follow through the moose tutorials and examples to learn to implement custom Materials. What you described is a simple end-to-end feed-forward NN surrogate which does not benefit from using NEML2. The training can be done independent of MOOSE.

Putting that aside, as a general advice, I don't think your idea would work. The CP model is too stiff for you to represent it as a feed-forward NN. However, I would not discourage you from first testing out your idea. I would cheer for you if it turns out to be working for your case. If it doesn't work we can revisit this.