| Contents | In order to bridge the gap between microstructural complexity and homogenized macroscopic behavior, recent developments in computational mechanics necessitate effective and accurate multiscale modeling tools. For reduced-order modeling of heterogeneous materials, Deep Material Networks (DMNs) offer a robust, physics-informed machine learning framework that makes accurate and timely predictions under challenging loading scenarios. By integrating mechanics-based building blocks with hierarchical network designs, DMNs provide a scalable method for learning and extrapolating constitutive behavior from high-fidelity simulations. This presentation introduces the core concepts, along with its mathematical formulation, offline training techniques, and online prediction extrapolation capabilities. Particular attention is paid to the framework's physical consistency and interpretability, distinguishing it from traditional black-box surrogate models. Recent developments from the presenter and collaborators are also highlighted. These advancements demonstrate the potential of DMNs to accelerate computational workflows and deepen understanding of multiscale material behavior. |
