Composite temperature profile and tooling optimization via Deep Reinforcement Learning

Abstract

Carbon Fibre Reinforced Plastic (CFRP) manufacturing cycle time is a major driver of production rate and cost for aerospace manufacturers and manufacturers in other composite heavy industries. One major thread in optimizing liquid moulding manufacturing processes such as VARTM involves modifying variables affecting heat transfer while the composite is being processed. Modelling and control of these curing processes is known to be challenging as the temperature response during cure is dependent on many variables including resin cure kinetics, carbon perform thickness and material properties, tooling thickness and material properties, and autoclave or oven heat transfer coefficients. In this work, we introduce a novel optimization method for composite cures taking into account both air temperature and tooling. Framing oven air profile optimization as optimal control, we are able to learn an adaptive and generalizable temperature controller via Deep Reinforcement Learning in simulation. Using 5 1D heat diffusion models at different locations, this controller learns from experience to anticipate the thermochemical dynamics of composite curing, and makes use of both heating and cooling to reduce cycle time while adhering to soft constraints on the maximum temperature of the system and avoid thermal runaway. This adaptive controller is then used in a Bayesian Optimization loop to optimize tooling geometry, adding and removing thermal mass in different locations to ensure the resultant cure adheres to hard constraints and has minimal overall cycle time. On two realistic aerospace parts with complex geometry and varying thicknesses we are able to reduce ramp to cure time by 27 to 40%.