SD-DAL: Structure Discovery of Elastic Metamaterials via Deep Active Learning

By mimicking some natural structural features such as honeycombs and plant xylem, architected materials can achieve extremely high material utilization through ordered or disordered microstructures, thereby obtaining properties beyond conventional materials. However, the rational design of these metamaterials generally relies on experts' prior knowledge and requires painstaking effort. Here, we present an effective method to discover novel designs of elastic metamaterials via a machine learning (ML) cycle consisting of the finite element method (FEM) and deep neural networks (DNN). With moderate amount of initial data, high-quality structure-property data points will be actively added to the database as the model learns. Using this structure discovery method via deep active learning (SD-DAL), it provides a data-efficient paradigm for the fast and intelligent design of architected materials with tailored mechanical, physical, and chemical properties. Specifically, we apply the method to in-plane structure designs which undergo out-of-plane loads. The optimal generated elastic metamaterials can exhibit 40% higher elastic modulus compared to the initial rational design with the same weight. And the load-bearing capacity is enhanced by the discovered new design, which interestingly shows a similar effect to the distribution and orientation of leaf veins in plants. The proposed SD-DAL is promising in the lightweight and advanced materials and structures design.