Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 149893

149893 - Microvoiding and Constitutive Damage Modeling With Artificial Neural Networks 

Continuum models of porous media have significantly advanced computational fracture mechanics for traditional ductile materials, but their inherent assumptions often limit their applicability to a broader range of target materials and varying loading conditions. In this study, we introduce a novel data-driven approach utilizing artificial neural networks (ANNs) to predict both the microscopic voiding characteristics (void shape, porosity) and the macroscopic stress-strain constitutive response of porous elasto-plastic materials under diverse deformation states.

Our methodology involves training ANNs on a comprehensive dataset derived from finite element models of 3D representative volume elements (RVEs), each featuring a discrete spherical void. These RVEs were subjected to various combinations of loading states to generate the training data. The trained ANNs demonstrated remarkable capability in making accurate interpolative and extrapolative predictions across a wide spectrum of initial porosities (ranging from 0% to 20%) and loading conditions that were not included in the training dataset. Notably, these predictions remained robust even at high deformation strains that induce significant material softening and void growth.

A key aspect of our approach is the implementation of transfer learning, allowing the ANNs initially trained on a specific porous material dataset to be efficiently adapted to other porous materials with substantially different properties. This adaptation requires a significantly reduced training dataset, showcasing the versatility and efficiency of our machine learning model. In contrast to the Gurson constitutive model, which assumes a rigid-plastic response along with spherical symmetric void growth and axisymmetric loading, our ANNs are only limited by the assumptions inherent in the finite element models used to generate the dataset for training. By encompassing low to very high initial porosities in the training dataset, our ANNs are applicable to simulating the deformation response of a wide range of porous material structures.

Our results highlight the potential of data-driven approaches to revolutionize the field of computational fracture mechanics. The ANNs accurately quantify the stress-strain constitutive behavior and the void evolution response under complex loadings, providing quantitative measures of the evolving shapes of the deformed voids. This study underscores the transformative impact of machine learning in advancing our understanding and modeling of porous elasto-plastic materials. It offers a promising path forward for future research and practical applications in material science, extending the applicability of these neural networks to general porous material systems without the need for significant re-training. Ultimately, the neural network approach bridges scales to achieve rapid constitutive modeling of materials, offering a powerful tool for elucidating the micromechanics of material damage and failure in fracture and fatigue problems.

 

Presenting Author: Ning Li Department of Aerospace Engineering, University of Illinois at Urbana-Champaign

Presenting Author Biography: Ning Li is a fifth-year Ph.D. student in the Department of Aerospace Engineering at the University of Illinois at Urbana-Champaign. Ning's research focuses on computational multiscale mechanics of materials, with a specialization in the mechanical properties of fiber-reinforced materials. His work involves constitutive damage modeling using artificial neural networks, aiming to enhance the predictive capabilities of material behavior.

Authors:

Huck Beng Chew Department of Aerospace Engineering, University of Illinois at Urbana-Champaign
Ning Li Department of Aerospace Engineering, University of Illinois at Urbana-Champaign