Abstract

Physics-based process simulation is a cost-effective alternative to trial and error when determining suitable process parameters for the manufacture of a composite structure. A critical component of the manufacturing process is the temperature cycle imposed on the structure during cure. High fidelity 3D finite element analyses can be used to accurately predict the temperature history and subsequent material properties for a given cure cycle, but at a significant computational cost. Determining the optimal cure cycle which minimizes potential manufacturing defects and meets specifications using high fidelity finite element simulations can be prohibitively time-consuming as typical gradient-based optimization techniques may require thousands of model evaluations. In many common manufacturing scenarios, a 3D structure can be approximated by one or more 1D slices which significantly reduces the time required to run a simulation, but it still may take hours to complete the optimization. This work demonstrates how recurrent neural network (RNN) surrogate models can be used as a computationally inexpensive alternative to finite element analyses for cure cycle optimization. RNN surrogate models have been trained for predicting process parameters such as temperature, degree of cure, and resin viscosity of 1D ?drill-through? stacks of larger assemblies consisting of a composite laminate on a tool. Leveraging the speed of these surrogate models, they have been subsequently used in an optimization algorithm to determine the shortest duration cure cycle that meets the manufacturing requirements. Multiple 1D stacks can be considered in parallel to ensure the cure cycle is valid for the whole assembly and not just a single location. The manufacturing requirements are defined by the process specification, which typically changes from process to process and this method is shown to be capable of handling process specifications of ?real-world? complexity. Ultimately, it is shown that the time required to perform the optimization using the surrogate model is considerably less than the time required to run the optimization using the finite element model and both methods converge to comparable optimized cure cycles.