High-throughput design of fiber reinforced cement-based composites using deep learning

Abstract As the combinatorial space of a composite is virtually infinite and cannot be explored completely, a deep-learning method was proposed for high-throughput fiber-reinforced cement-based composites (FRC) design. First, a deep hierarchy network was developed to measure the relationship between the experimental variables and the FRC properties. A gradient-based high-throughput method based on the deep hierarchy network was then proposed to design FRCs, which were expected to have one or more certain properties. At last, a fine-tuning method was employed to guarantee its transferability for all types of FRCs. The results showed that the proposed method was able to design cement-fiber-water-curing-aging systems for carbon fiber reinforced cement-based composites (CFRCs). The fine-tuning method could transfer the CFRC model to design other FRCs. Thus, the proposed method showed promise for releasing the composite material property optimization from labor-consuming and low-efficiency laboratory tests.