Decoding material networks: exploring performance of deep material network and interaction-based material networks
Journal
Journal of Mechanics
Journal Volume
40
Start Page
796
End Page
807
ISSN
1811-8216
Date Issued
2024
Author(s)
Abstract
The deep material network (DMN) is a multiscale material modeling method well-known for its ability to extrapolate learned knowledge from elastic training data to nonlinear material behaviors. DMN is based on a two-layer building block structure. In contrast, the later proposed interaction-based material network (IMN) adopts a different approach, focusing on interactions within the material nodes rather than relying on laminate composite structures. Despite the increasing interest in both models, a comprehensive comparison of these two computational frameworks has yet to be conducted. This study provides an in-depth review and comparison of DMN and IMN, examining their underlying computational frameworks of offline training and online prediction. Additionally, we present a case study where both models are trained on short-fiber reinforced composites. We trained each model using elastic linear datasets to evaluate their performance and subjected them to multiple loading tests. Their performance is closely compared, and the possible factors that cause differences are explored. The superiority of IMN in offline training and online prediction is found. © The Author(s) 2024. Published by Oxford University Press on behalf of Society of Theoretical and Applied Mechanics of the Republic of China, Taiwan.
Subjects
deep material network
interaction-based material network
mechanistic machine learning
Publisher
Oxford University Press (OUP)
Type
journal article