Machine Learning-Enabled Prediction of Spatiotemporal Strain Evolution in Stereolithography-Fabricated Architected Hydrogels

This study developed a machine-learning framework for predicting the spatiotemporal evolution of full-field strain in stereolithography-fabricated architected hydrogels. NIPAm-based hydrogel films were fabricated by grayscale digital light processing stereolithography with stiff re-entrant honeycomb cellular domains embedded within a softer continuous matrix. Spatial variation in UV exposure produced local differences in crosslink density and stiffness, creating mechanically heterogeneous structures with programmed deformation behavior. During tensile loading, digital image correlation revealed that deformation evolved from relatively uniform strain at low loading to pronounced strain localization along pathways governed by the underlying cellular geometry and stiffness mismatch between the stiff domains and compliant matrix. The experimentally measured strain maps were used to train a sequence-to-sequence deep-learning model combining convolutional encoders with ConvLSTM layers. A total of 120 time-resolved strain maps were divided into 80% training and 20% validation data, with three consecutive strain fields used to predict the subsequent deformation state. The trained model accurately reproduced both global strain distributions and localized deformation pathways and achieved a validation mean squared error of 5.49 × 10^-3. Prediction error showed only weak dependence on strain magnitude but increased substantially in regions with high spatial strain gradients, indicating that sharp localization boundaries and spatial complexity were the principal challenges for prediction. The model also produced smoother strain fields than the raw DIC data while preserving physically meaningful localization patterns, effectively reducing high-frequency measurement noise. Overall, the work demonstrates how experimental mechanical testing, full-field DIC, and deep learning can be integrated to predict deformation evolution and improve understanding of structure-property relationships in architected soft materials.