Mechanical property prediction of complex-configuration mechanical metamaterials based on multimodal CNN

Abstract

The macroscopic mechanical properties of mechanical metamaterials are strongly influenced by their microstructural geometries. However, efficient and accurate prediction of these properties remains a significant challenge, especially for large-scale datasets featuring highly diverse configurations. This study introduces a novel image-based multimodal deep learning framework that combines structural images with auxiliary indices to enable the simultaneous prediction of discrete performance indices and complete force-displacement (F-D) curves. To achieve this, we have constructed a high-resolution database comprising 64,000 unit-cell images (porosity Ф ∈ [0.3,0.8]) and developed an automated Python–Abaqus platform capable of completing the full finite element analysis process for a single configuration in approximately three seconds, thus significantly enhancing both modeling and computational efficiency. Systematic evaluations using three representative CNN backbones demonstrate that the multimodal Xception model surpasses both ResNet50 and VGG16, achieving considerably higher accuracy compared to unimodal input. High prediction accuracy is attained for performance indices (R² ≈ 0.97–0.99). Moreover, by integrating structure-sensitive auxiliary indices and an attention mechanism, the prediction accuracy of Poisson’s ratio is enhanced from R² = 0.88 to 0.93, effectively capturing the microstructure-sensitive response characteristics. For curve prediction, the average value of the area ratio metric on the validation set is 1.01, accurately reflecting nonlinear response behaviors. The proposed framework significantly advances the development of performance prediction and intelligent design methodologies for mechanical metamaterials.