Multiscale inverse design of mechanical cloaks via deep generative modeling

Mechanical cloaks offer a promising route for mitigating stress concentrations induced by defects; however, their design is hindered by the prohibitive cost of navigating high-dimensional topology spaces. Here, we propose a data-driven multiscale inverse design framework that integrates generative modeling with evolutionary optimization to realize heterogeneous truss-based cloaks. We first construct a comprehensive structure-property library of truss unit cells and train a conditional variational autoencoder (CVAE) to map target effective stiffness tensors to specific microstructural topologies. A high-throughput computational pipeline is subsequently established that couples finite element analysis with an SE-ResNet18 surrogate model to predict the cloaking performance from stiffness distributions. By embedding this surrogate within a genetic algorithm, we efficiently explore the design space to identify optimal stiffness fields, which are then reconstructed into physical structures via the CVAE. The optimized designs achieve a mean relative displacement-field error of 2.74 ± 0.18% (standard deviation) across ten independent optimization runs, with finite element validation confirming that the reconstructed cloaks retain cloaking performance, achieving errors as low as 1.91% for the best-performing design. This work establishes a systematic AI-accelerated framework for the data-driven design of heterogeneous mechanical metamaterials.