Deep Learning Accelerated Design of Bézier Curve-Based Cellular Metamaterials with Target Properties

Abstract

Machine learning has sparked significant interest in the realm of designing mechanical metamaterials. These metamaterials derive their unique properties from microstructures rather than the constituent materials themselves. In this context, we introduce a novel data-driven approach for the design of an orthotropic cellular metamaterials with specific target properties. Our methodology leverages a Bézier curve framework with strategically placed control points. A machine learning model harnesses the positions of these control points to achieve the desired material properties. This process consists of two main steps. Initially, we establish a forward model capable of predicting material properties based on given designs. Then, we construct an inverse model that takes material properties as inputs and produces corresponding design parameters as outputs. Our results demonstrate that the dataset generated using the Bézier curve-based strategy shows a wide range of elastic distributions. Describing the geometry in terms of design parameters, rather than pixel-based figures, enhances the training efficiency of the networks. The dual-network training approach helps avoid contradictions where specific elastic properties may correspond to various geometric designs. We verify the prediction accuracy of the inverse model concerning elastic properties and relative density. The presented approach holds promise for accelerating the design of cellular metamaterials with desired properties.