Design of architectured composite materials with an efficient, adaptive artificial neural network-based generative design method

In recent years, machine learning methods have been applied increasingly in material design and discovery. While these approaches have demonstrated promising performance and produced novel materials that are unachievable previously, most existing machine learning methods are supervised methods and need millions of labeled training data. The construction of the training data requires massive computational resource and is extremely time-consuming, forming a major bottleneck of machine learning approaches. In this work, an efficient artificial neural network-based inverse design method is developed for the design of architectured composite materials with novel properties. By adopting an adaptive learning and optimization strategy, the design space can be effectively explored, thereby greatly reducing the number of required labeled training data. In addition, a generative adversarial network is used to generate design candidates which drastically reduces the number of design variables and speeds up the optimization process. The excellent performance of the method is demonstrated on the design of several novel composite materials such as materials with high toughness, high stiffness near theoretical upper bound. Compared with some existing machine learning based methods, a two-order-magnitude reduction in the number of labeled training data has been achieved while maintaining the same level of design performance.