Abstract: Composites are widely used to create tunable materials to achieve superior mechanical properties. Brittle materials fail catastrophically in the presence of cracks. Incorporating softer constituents into brittle materials can alleviate stress concentration, leading to tougher and stronger composites. However, searching for the optimal designs of composites is extremely challenging due to the astronomical number of possible material and geometry combinations. Here, we apply machine learning to a composite system and demonstrate its capacity to accurately and efficiently predict mechanical properties including toughness and strength. The method we used incorporates machine learning techniques to generate optimal designs with orders of magnitude better than the mean properties of the input training data, and at a much lower computational cost compared to exhaustive methods. Additionally, the method can further rebuild the detailed performances of the designs, even when this information is lost in the training process. The results demonstrate the ability of machine learning to search for optimal designs with very limited training data. To demonstrate the application of machine learning to composite design, we optimize a large-scale system not tractable by an exhaustive brute force approach and show that it is a promising tool towards composite design. This work offers a new perspective in the exploration of design spaces and accelerating the discovery of new functional, customizable composites.
Full paper: GX Gu, C-T Chen, MJ Buehler, De novo composite design based on machine learning algorithm, Extreme Mechanics Letters, 2018, DOI: 10.1016/j.eml.2017.10.001