It was said that Artificial Intelligence (AI) is the new electricity. During the dawn of the 20th century, alternating current and its induction motor affected every industry and every part of human life and changed it forever. Nowadays, we start seeing AI doing the same thing. Modeling based optimization in the novel material design is no exception, with a few recently published papers using machine/deep learning methods to successfully assist and improve the established numerical methods. Mechanical metamaterials are artificial structures with counterintuitive mechanical properties that originate in the geometry of their unit cell instead of the properties of each component. In recent times, materials with enhanced properties have been intensely sought after in the scientific and industrial communities. In essence, such materials are usually optimized through the choice of the constituent materials, the volume fraction, and the architecture. To date, designing the architectures of such materials is contemporary, as it efficiently only allows one to obtain materials with unprecedented properties not found in natural materials. These properties include but not limited to very high stiffness-to-weight ratio and negative Poisson’s Ratio. Topology optimization offers a systematic method to design materials given a set of loads boundary conditions, and constraints. This approach aims to identify the optimal material distribution that provides the highest performance of the system while keeping the design constraints satisfied. Coupled by recent developments in additive manufacturing, numerous intricate geometries can now be achieved. While more advanced fabrication techniques allow for new opportunities for making such metamaterials, the optimization problem remains challenging. Worse off, numerical analyses, including topology optimization problems, can be extremely computationally expensive due to the many required iterations.
In general, data-driven models are rising as an auspicious method for the geometrical design of materials and structural systems. Nevertheless, existing data-driven models customarily address the optimization of structural designs rather than metamaterial designs. To address this, University of Illinois researchers: Mr. Hunter T. Kollmann, Dr. Diab W. Abueidda, Professor Seid Koric, Dr. Erman Guleryuz and Dr. Nahil A. Sobh developed a new deep learning model at the National Center for Supercomputing Applications (NCSA) and its Industry Program and the Center for Artificial Intelligence Innovation . The deep learning model is based on a sophisticated convolutional neural network (CNN) that predicts optimal metamaterial designs. Their work is currently published in the research journal, Materials and Design.
In their approach, they first presented an overview of the topology optimization problem along with the homogenization and periodic boundary conditions. The team then outlined the design space and training, validation, and testing datasets the data a needed for training. They then scrutinized the architecture of the CNN model, hyperparameters, and the loss function and metrics utilized in assessing the performance of the CNN model. Overall, the model was generated using the energy-based homogenization method and periodic boundary conditions.
The authors reported that the developed deep learning model non-iteratively optimized metamaterials for either maximizing the bulk modulus, maximizing the shear modulus, or minimizing the Poisson’s ratio (including negative values). More so, the researchers mentioned that the data was generated by solving a large set of inverse homogenization boundary values problems, with randomly generated geometrical features from a specific distribution.
In summary, the study developed a CNN model in order to predict with accuracy the optimized metamaterial design for either maximizing the bulk modulus, maximizing the shear modulus, or minimizing the Poisson’s ratio. Remarkably, the researchers realized that after properly training their deep learning model on high-end computers, they were able to instantly inference have quality topology optimization results on a low-end computing platform, such as laptops, and without high-end computers or modeling and optimization software In a statement to Advances in Engineering, Dr. Diab W. Abueidda explained that their model paves way for amazingly accelerated topology optimization algorithms for multiscale metamaterial systems and other computationally expensive topology optimization problems.
Hunter T. Kollmann, Diab W. Abueidda, Seid Koric, Erman Guleryuz, Nahil A. Sobh. Deep learning for topology optimization of 2D metamaterials. Materials and Design 196 (2020) 109098