Data-driven approaches and advances in machine-learning algorithms are emerging techniques sought out to speed up the computational modeling and time-to-solution predictions of microstructure and materials behavior. In computational mechanics, these techniques bypass computationally expensive direct numerical simulation (DNS) solvers, such as the finite element method, fast Fourier transform, or mesh-free solvers, by approximating the effective constitutive response of materials microstructures with surrogate models trained on stress-strain datasets. For example, recent studies have put forward such surrogate models capable of discovering unknown constitutive laws or rapidly predicting the effective nonlinear material’s behavior as well as path-dependent behavior of microstructures. These surrogate models are based on a variety of methods, including artificial neural networks, two-dimensional and three-dimensional image-based convolutional neural networks, cluster-based reduced-order models, and Gaussian process regression models. However, all of these methods rely heavily on computationally expensive microstructure-level simulations for a given and fixed material constitutive relation. In other words, these machine-learning models need to be retrained whenever the constitutive relation changes, limiting their extrapolation performance. In this work, a group led by Prof. R mi Dingreville from the Center for Integrated Nanotechnologies, Sandia National Laboratories, USA, demonstrated an approach to visualize the network parameters as an analogous unit cell and use this visualization to “quilt” patches of shallower networks to initialize deeper networks for a recursive training strategy. The result is an improvement in the accuracy and calibration performance of the network and an intuitive visual representation of the network for better explainability. The authors suggested that this visualization and recursive training strategy could be extended to other deep material network (DMN) architectures, which allows for the material orientation at the nodes to vary. This approach could also be used for other classes of microstructures, notably polycrystalline materials. Although the approach used in this study could be improved with a better refinement of the visualization of the network, such as the consideration of multiple microstructural length scales or the representation of the network as an actual microstructural representation of the DNS periodic unit cell, this work is a step towards better strategies for developing explainable DMN with improved accuracy.