Time-Resolved Regression Digital Twin for Continuous Inference of Encoded Mechanical Heterogeneity in Architected Materials

High Resolution Image Download MS PowerPoint Slide Architected materials achieve programmable mechanical behavior through spatially encoded compositional heterogeneity, yet inferring encoded stiffness fractions from structural appearance alone remains unreliable when mechanically distinct configurations are visually indistinguishable prior to loading. Here we show that deformation sequences acquired during tensile loading constitute a physically meaningful information channel for continuous design-space inference, and introduce a design-space digital twin, a single learned model spanning a continuous material design space. A spatiotemporal framework integrating convolutional feature extraction, long short-term memory temporal learning, and regression-constrained representation is trained on six discrete stiffness fractions (G0–G50) at strains up to 60%. An intermediate condition (G25), withheld entirely from training, is predicted with a mean value of 25.3 and a mean absolute error of 1.47, more than three times lower than the minimum error achievable by discrete classification, confirming genuine interpolation within a continuous and navigable design space. Systematic variation of sequence starting frames reveals a deformation-dependent learnability regime, wherein mechanically informative morphology emerges progressively during loading and an optimal onset balances morphological contrast against temporal evolution. These findings reframe digital twinning for architected materials as a design-space inference problem, establishing deformation evolution as the information channel and regression-constrained temporal learning as the enabling strategy.