Deep neural networks based predictive-generative framework for designing composite materials

Designing composite materials as per the application requirements is fundamentally a challenging and time consuming task. Here we report the development of a deep neural network based computational framework capable of solving the forward (predictive) as well as inverse (generative) design problem. The predictor model is based on the popular convolution neural network architecture and trained with the help of finite element simulations. Further, the developed property predictor model is used as a feedback mechanism in the neural network based generator model. The proposed predictive-generative model can be used to obtain the micro-structure for maximization of particular elastic properties as well as for specified elastic constants. One of the major hurdle for deployment of the deep learning techniques in composite material design is the intensive computational resources required to generate the training data sets. To this end, a novel data augmentation scheme is presented. The application of data augmentation scheme results in significant saving of computational resources in the training phase. The proposed data augmentation approach is general and can be used in any setting involving the periodic micro-structures. The efficacy of the predictive-generative model is demonstrated through various examples. It is envisaged that the developed model will significantly reduce the cost and time associated with the composite material designing process for advanced applications.