Demand-driven predictive tailoring of anisotropic yield surfaces in origami metamaterials via machine learning

Abstract

The yield surface defines the elastic-to-plastic transition in materials. However, accurately capturing the multiaxial yield surfaces of anisotropic metamaterials remains challenging with conventional criteria, and active tailoring of yield surfaces is also underdeveloped. Here, a novel machine learning framework (Q-TEncMLP) is proposed for predicting and on-demand tailoring anisotropic yield surfaces in origami metamaterials. First, a predictive deep learning model (TEncMLP) is trained on limited data to achieve end-to-end mapping from topologies to multiaxial yield surfaces. Through transfer learning with frozen parameters, the model generalizes to new yield surfaces using only 20% of additional data, enhancing efficiency across different stress states and geometric variations. Beyond prediction, attention-weight analysis provides mechanical interpretability by revealing the roles of key parameters in anisotropic yielding. Furthermore, TEncMLP is embedded into reinforcement learning as a digital twin environment, where mechanics-informed reward functions facilitate demand-driven tailoring of yield surfaces. This allows tailored yield surfaces for various objectives, including max/minimization, target matching, and lightweighting while preserving mechanical performance. Overall, this work not only clarifies the key mechanisms governing anisotropic yield in origami metamaterials, but also provides a general paradigm for intelligent constitutive modeling, shifting from experience-driven to demand-driven.