PREDICTION OF MECHANICAL BEHAVIOR OF WOVEN COMPOSITE VIA DEEP NEURAL NETWORK

Abstract

The mechanical behavior of CFRP was trained by deep-neural-network (DNN). For an accurate analysis of composite properties, micromechanics of failure based multi-scale simulation method was introduced for progressive damage analysis of composite materials. The meso-scale and micro-scale representative volume was used for multi-scale simulation, and stress transfer between meso-micro scale model, was performed by applying stress amplification factor (SAF). With the developed simulation method, stress-strain curves of CFRP were derived depending on constituent properties and yarn structures. The databases of mechanical behavior were trained by deep-neural-network, which use stress-strain curves as training output, and mechanical, geometrical properties as training input, respectively. As a result, mechanical behavior of CFRP could be predicted by the developed method in a very fast time with high accuracy.