Rapid cure simulation using artificial neural networks

Abstract

Process simulations based on physical models describing the governing process phenomena have played an important role in improving the fundamental understanding and development of composites fabrication techniques over the last decade. In the several ongoing research efforts directed at improving manufacturing affordability of composite materials, simulation models offer great potential for tasks such as process optimization, robust process design, total quality manufacturing, and model-predictive on-line process control. However, the computational tedium associated with a rigorous numerical process simulation hampers the realization of the complete potential of simulation-assisted materials manufacturing, which calls for the simulation time scales to be comparable to those of the fabrication processes. Towards addressing this need, this paper presents the use of artificial neural networks trained using a physical process model as an effective substitute for rigorous numerical simulations in the simulation-assisted materials manufacturing endeavor. Considering the case of thermosetting-matrix composites fabrication, the use of neural networks trained and validated using the physical models is shown to increase the computational speed by several orders of magnitude. Further, network training in terms of non-dimensional groups formed of the process and product parameters is presented as an effective means of better incorporating the physical relationships among the parameters, generalizing the network training across material systems and product specifications, and reducing the number of training parameters. The accuracy of the neural network-based simulations is assessed for a wide range of practically relevant processing conditions.