Prediction of stress-strain behavior of carbon fabric woven composites by deep neural network

In this work, a novel deep neural network was proposed for predicting the mechanical behavior of plain carbon fabric reinforced woven composites. The deep neural network was trained by a pre-simulated stress-strain curve database of woven composites depending on yarn structures and the mechanical properties of the fiber and matrix. Micro-mechanics-based multi-scale analyses of woven composites were conducted for progressive damage analysis. These analyses utilized the stress amplification factor to transfer stress between the micro-scale and meso-scale simulations and the respective failure criteria were applied for micro-scale stresses of the fiber and matrix, respectively. The database of stress-strain curves under tensile, compressive and shear loading was acquired for different yarn geometries and constituent properties. These variables were used as training input and the resulting stress-strain curves were used as training output of the network. To optimize the network, hyper parameters of the neural network, such as the number of layers and nodes, were determined by the Hyperband optimization algorithm. The train and test of deep neural network model was performed by TensorFlow backend using the Keras library in Python. Mechanical tests were performed to validate the predicted mechanical behavior from both simulation and the deep neural network. As a result, the stress-strain curves under tensile, compressive and shear loading of arbitrary woven carbon composites can be successfully predicted in several seconds by the deep neural network with high accuracy.