Associate Professor Dr. Peng Wen and his collaborative team at the Department of Mechanical Engineering, Tsinghua University, developed an innovative architected metamaterial design methodology using a generative architecture design and multiobjective active learning loop (GAD-MALL) combined algorithm. This novel approach has proven effective for the design and optimization of 3D-printed porous metal bone implants, presenting a versatile solution toward solving high-dimensional multi-objective problems with data sparsity and enormity of the search space in metamaterial design.
Architected materials are one of the most widely adopted engineering materials. Due to their excellent mechanical performance and adaptable properties, architected materials are very popular in many fields, such as those of lightweight structures, tissue engineering, battery electrodes, and electromagnetics. Moreover, recent progress in 3D printing has further enabled the customized and inexpensive fabrication of complex material geometries. Despite the broad applicability and immense potential of architected materials, designing them is particularly difficult. The traditional method generally relies on numerical simulation and theoretical analysis which are usually exhausting and time-consuming. Recently, machine learning (ML) has emerged as a promising technique to circumvent this problem and find the optimal solution without any prior knowledge requirements. Furthermore, active learning that combines ML and simulations or experiments to tackle optimization problems is an emerging topic at the frontier of science. However, some of these methods mainly focus on 2D-structure-related problems, while others use Bayesian optimization to solve low-dimensional problems or focus on an unconstrained single objective. The efforts toward solving high-dimensional multi-objective problems are often obfuscated by the data sparsity, the enormity of the search space, and stringent external constraints. For example, porous bone implants should align with the elastic modulus of bone tissue, while also providing adequate compressive strength and biocompatibility; meanwhile, a porous structure comprising 3*3*3 units, with each unit varying across seven porosity levels, can yield an immense 727 potential designs.
The research introduces an inventive integration of a generative model, three-dimensional convolutional neural network (3D-CNN), and numerical simulations within the GAD-MALL algorithm for designing architected metamaterial structures. The process begins with unsupervised learning on a dataset of over 18,000 porous geometries using an autoencoder model to distil high-dimensional data into a low-dimensional latent space representation. Gaussian mixture models are then employed for sampling in the latent space, with the decoder reconstructing the sampled data to new architected metamaterial designs. 3D-CNN models are then used to predict the performance of these new designs and select the designs with high predicted performance, which are subsequently validated through numerical simulations. Through iterative cycles of generation, prediction, and simulation, the GAD-MALL algorithm effectively optimizes the multi-objective performance of the structures. Optimal designs are 3D-printed and experimentally validated, demonstrating significant improvements over conventional designs and confirming the efficacy of this AI-driven design paradigm in addressing the challenges of high-dimensional multi-objective optimization in designing architected metamaterials.
