Prediction of Macroscopic Compressive Mechanical Properties for 2.5D Woven Composites Based on Artificial Neural Network

The complex modeling and computational cost are unavoidable in analysis of finite element models (FEMs) when mechanical properties of woven composite materials are predicted. To overcome the drawbacks of FEMs, two different artificial neural network models (ANNMs) based on quasi-static axial compression experimental data of 2.5D woven composite plates (2.5DWCPs) are constructed: (1) The direct strength prediction model (DSPM) is a non-destructive way to predict strength, which is meaningful in engineering; (2) The indirect strength prediction model (ISPM) is based on stress-strain curves, which firstly proposes a simplified data processing method including the state variables (SVs). The SVs are introduced to modify the experimental stress-strain curves, which not only reduces training data size but also significantly improves prediction accuracy. Then, the performance of the DSPM and the ISPM has been evaluated by numerical examples. The results indicate that the DSPM has simple and direct expressions of input parameters (IPs) and output parameters (OPs), which makes it easier to construct and train ANNMs. The ISPM not only utilizes sufficient training data from experiments but also performs well in predicting stress-strain curve and failure strain. In short, two proposed ANNMs have ability to fast and accurately predict compression strength, which are more suitable for engineering than FEMs. To reduce experimental costs, the DSPM is proposed to produce reasonable results. If a lot of experimental data are prepared, the ISPM can produce more accurate results.