A data-driven approach to full-field nonlinear stress distribution and failure pattern prediction in composites using deep learning

An image-based deep learning framework is developed in this paper to predict damage and failure microstructure-dependent composite materials. The work motivated by the complexity computational cost of high-fidelity simulations such proposed predicts post-failure full-field stress distribution crack pattern two-dimensional representations composites based on geometry microstructures. material interest selected be a high-performance unidirectional carbon fiber-reinforced polymer composite. contains two stacked fully-convolutional networks, namely, Generator 1 2, trained sequentially. First, learns translate microstructural distribution. Then, 2 output pattern. A physics-informed loss function also designed incorporated further improve performance facilitate validation process. In order provide sufficiently large data set for training validating framework, 4500 are synthetically generated simulated an efficient finite element framework. It shown that approach can effectively composites' pattern, most complex phenomena simulate solid mechanics.