Mechanical performance prediction of bio-inspired metamaterials based on multimodal attention mechanism

Abstract

This study presents a bio-inspired metamaterial design and prediction framework driven by image data and enhanced by a multimodal attention fusion mechanism. Drawing inspiration from the irregular and heterogeneous vein architectures of leaves and insect wings, we extract representative topological features through image processing and filtering, which are then converted into truss-like configurations to build a large, diverse image database. A fully automated modeling and simulation platform is developed, enabling high-throughput finite element analysis (FEA) of 46,880 structures, from which naturally optimized configurations with negative Poisson’s ratio and high energy absorption capacity are identified. To enable rapid and accurate performance prediction, we construct a deep learning framework that integrates image features with structural auxiliary indexes—porosity (Φ), minimum strut width (L_min), and central filling ratio (R_O). An attention mechanism is employed to adaptively weight these modalities, significantly enhancing the model’s representational power. The proposed model achieves a prediction accuracy of 0.98 for energy absorption, significantly outperforming traditional unimodal models. Furthermore, robustness and generalization capabilities are verified through extensive structural perturbation experiments, including configuration variation, thickness modulation, and local topology editing. The results demonstrate the model’s strong adaptability to microstructural changes, offering a powerful tool for bio-inspired mechanical metamaterial design and evaluation.