Abstract
Despite the increasing importance of strain localization modeling (e.g.,failure analysis) in computer-aided engineering, there is a lack of effectiveapproaches to consistently modeling related material behaviors across multiplelength scales. We aim to address this gap within the framework of deep materialnetworks (DMN) - a physics-based machine learning model with embedded mechanicsin the building blocks. A new cell division scheme is proposed to track thescale transition through the network, and its consistency is ensured by thephysics of fitting parameters. Essentially, each microscale node in the bottomlayer is described by an ellipsoidal cell with its dimensions back-propagatedfrom the macroscale material point. New crack surfaces in the cell are modeledby enriching cohesive layers, and failure algorithms are developed for crackinitiation and evolution in the implicit DMN analysis. Besides single materialpoint studies, we apply the multiscale model to concurrent multiscalesimulations for the dynamic crush of a particle-reinforced composite tube andvarious tests on carbon fiber reinforced polymer composites. For the latter,experimental validations on an off-axis tensile test specimen are alsoprovided.