E-Thesis 123 views
A Neural Network-Based Finite Element Framework for the Modelling of Inelastic Solids / EUGENIO ZAVALA
Swansea University Author: EUGENIO ZAVALA
DOI (Published version): 10.23889/SUThesis.69253
Abstract
A neural network-based surrogate model for inelastic solid materials simulations is presented. The network architecture incorporates the elastoplasticity equations and has been proven to be exact in one-dimensional elastoplasticity with hardening. This strategy provides an alternative to the complex...
| Published: |
Swansea University, Wales, UK.
2024
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Perić, D., and Dettmer, W. G. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa69253 |
| Abstract: |
A neural network-based surrogate model for inelastic solid materials simulations is presented. The network architecture incorporates the elastoplasticity equations and has been proven to be exact in one-dimensional elastoplasticity with hardening. This strategy provides an alternative to the complex task of formulating constitutive equations for new materials due to the capability of learning directly from data. The strategy has a wide range of applications un- der the two main scenarios that generate the necessary data: (a) by performing physical experiments, and (b) by numerical homogenisation within the multi-scale analysis. The network is constructed based on the concept of internal states and utilises only observable variables (strains and stresses); hence, a recurrent structure is required to account for the deformation history. Due to the challenging task of training recurrent networks, a supervised parallel optimisation framework has been developed, which combines the exploration and exploitation capabilities of one (or several) meta-heuristic algorithms. The proposed optimisation strategy is a general-purpose optimisation framework. Therefore, this development offers a tool for a wide range of applications. Several analyses have been conducted, including hyperparameter analysis, algorithmic complexity and parallel message communications. The obtained performance of the proposed methodology surpasses current state-of-the-art algorithms. The trained network- based constitutive model is embedded as a surrogate model into a finite element code on a Gauss point level, replacing a traditional library of algorithmic constitutive models. The stress update is obtained from the network output, and the consistent tangent modulus is derived from the network architecture utilising the prescribed functional dependencies. Numerical experiments are presented using network surrogates trained with the Von Mises elastoplasticity model. |
|---|---|
| Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information. |
| Keywords: |
Data-driven computational mechanics, Neural Networks, Optimisation, Constitutive Modelling |
| College: |
Faculty of Science and Engineering |
| Funders: |
EPSRC / UKAEA |