Recurrent neural network based model discovery of nonlinear viscoelasticity

Date

2024-7-22

Author

Masood, Saım

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

110
views

31
downloads

Cite This

This thesis introduces a novel framework for automated model discovery in the context of nonlinear viscoelasticity. The framework leverages a recurrent neural network (RNN) model representing the stress update procedure that inherently satisfies necessary physical constraints. We trained the model on data comprising temporal sequences of applied deformation and resulting stresses. We use gradient-based optimization for the parameter identification, with gradients evaluated analytically via a recurrent derivative update algorithm. The loss function comprises two terms: one quantifying the accuracy of the predictions and another inducing sparsity. The sparsity term encourages some learnable parameters to be zero, resulting in interpretable models with a few meaningful parameters. We demonstrate the ability of the framework to produce interpretable sparse models on synthetically generated data and experimental datasets of VHB 4910 and HNBR50 polymers. Further validation and applicability of the framework are illustrated through a finite element simulation using the model learned from the experimental dataset of HNBR50 polymer.

Subject Keywords

Recurrent neural network, Model discovery, Nonlinear viscoelasticity

URI

https://hdl.handle.net/11511/110389

Collections

Graduate School of Natural and Applied Sciences, Thesis