Performance-Driven Precise Design Method for Mechanical Metamaterials Based on Pixels and Deep Learning

Abstract

This study proposes a mechanical metamaterial design method based on image processing and deep learning, enabling forward prediction of mechanical properties and data-driven inverse design of configurations. First, mechanical metamaterial configurations are converted into pixel representations, and image processing techniques are used to generate a large dataset of complex and diverse mechanical metamaterial configuration images. Then, a mechanical metamaterial performance computation platform (PAI) is developed through secondary development in Python to enable precise extraction of complex structural features from configuration images. This platform automates the characterization of discrete mechanical performance indexes and compression response curves. Next, a CNN-based model is trained on the image-performance database, achieving efficient and highly accurate predictions of multiple mechanical performance indexes and response curves, with an accuracy of up to 97%. Furthermore, this study introduces an inverse design method that combines performance prediction and efficient retrieval, its core lies in preconstructing and characterizing a diverse configuration database, transforming the inverse design problem into a regression-retrieval task. This approach eliminates the reliance on expensive high-performance computing, overcomes the limitations of existing studies that constrained by simpler configurations. More importantly, this approach enables controlled-error generation, ensuring the absolute accuracy of the generated configurations. By adopting image-based data-driven mechanical metamaterial design, this study introduces a new paradigm in mechanical metamaterial research. Compared to traditional approaches that rely on parametric and complex modeling processes, this method is more intuitive, concise, and highly efficient.

Keywords: Mechanical metamaterials, Inverse design, Performance prediction, Pixels, Deep learning