Integrated Finite Element Analysis-Artificial Neural Network Predictive Models for Hybrid Composites Failure under Biaxial Tension: Effects of Data Sizes

Abstract

Hybrid composite laminates have wide usage in aerospace and automotive sectors due to the high strength-to-weight ratio that they have; however, predicting their biaxial tension failure load is challenging. In addition, the size of input datasets for developing an accurate predictive model has always been debatable. Thus, this study develops an integrated Finite Element Analysis (FEA)-Artificial Neural Network (ANN) based predictive model with different datasets for hybrid determining failure of boron/glass epoxy laminates under biaxial tension. ANN models were trained with 27 and 100 datasets using a Multilayer Perceptron (MLP) structure by the Levenberg–Marquardt algorithm. Varying fiber angle, thickness and layup, the FEA results obtained for failure loads range from 384.83 N to 8281.81 N. For the ANN27 model, the data normalization approach resulted in 60.98% reduction in percentage error, signifying improvement in prediction. Statistical t-tests further confirmed that the 27 and 100 dataset models were not significantly different from each other, thus highlighting similar predictive capabilities. To conclude, this study formulates a valid FEA–ANN model for composite laminate failure load prediction with good insight regarding dataset size effect and the improvement achieved by the normalization approach.