Generative Design of Energy‐Absorbing Auxetic Metamaterials via Variational‐Autoencoder Latent‐Space Exploration and Nonlinear Finite‐Element Screening

Abstract

Auxetic cellular structures, which exhibit a negative Poisson's ratio, are conventionally designed manually, one topology at a time. This paper presents a generative design framework in which a convolutional variational autoencoder (VAE) with a low computational cost, trained on 31,920 solid isotropic material with penalisation (SIMP) topologies, automatically generates a large variety of new auxetic cellular structures. The generated topologies are typically blurred or carry defects such as disconnected regions and thin struts, so the framework does not stop at generation. A rule‐based repair protocol completes each raw cell into a manufacturable, simulation‐ready structure, and 78% of the sampled cells become valid while about one third are novel. To demonstrate that every generated cell can be fully evaluated, the structures are then verified by quasi‐static explicit finite‐element analysis (FEA) under periodic boundary conditions, which measures the finite‐strain Poisson's ratio together with the specific energy absorption (SEA) and the plateau stress (σpl). The three best verified structures summarise the outcome: cand_333 holds the deepest auxetic response (ν = −0.38) with a low flat plateau, while cand_98 and cand_49 (ν = −0.19 and −0.08) have the highest energy absorption (SEA up to 15.6 kJ/kg) and the flattest plateau, respectively. The FEA model is validated against a published auxetic‐lattice compression experiment within 3.2%–7.7% curve deviation. Future work will further optimise the VAE generation and extend these property calculations systematically in follow‐up studies; the framework is released as a reusable tool for auxetic and other cellular metamaterials.