Design of Artificial Neural Networks for damage estimation of composite laminates: Application to delamination failures in ply drops

This work proposes a data driven approach which utilizes Artificial Neural Networks (ANN) in conjunction with parametric non-linear finite element analysis. The aim is to provide a low cost numerical counterpart to the expensive experimental testing of advanced composite laminates. The training data of ANN are obtained from physical based modeling of the damage evolution and associated delamination failures of ply drops. In contrast to a black-box ANN modeling approach, the core of this study concerns the development of a method for determining an optimal neural architecture. More specifically, we employed a random search handcrafted methodology for the neural net topology learning based on exploration and experimentation. This methodology is enhanced by a detailed statistical analysis used to make inferences about the procedural generation of architectures. In the same context, a series of experiments are performed to obtain an optimal set of hyperparameters to achieve a good performance in the training dataset. Furthermore, a visualization of the respective manifolds of the ANNs hidden layers is provided using two popular dimensionality reduction techniques, namely PCA and t-SNE, so as to transform the network layer output data into 2D representations. Additional tests, among others, regarding the network ability to generalize to unseen data, showed that the optimal well-trained neural network is accurate and robust enough for near real-time predictions of the various damage evolution patterns, and outperforms other data driven methods under comparison.