Deep reinforcement learning for stacking sequence optimization of composite laminates

Abstract

Fiber reinforced polymer (FRP) composite laminates are increasingly used in a wide range of safety–critical products due to their excellent material properties. The stacking sequence of FRP composite laminates plays a critical role in the resulting part’s mechanical properties. Despite this, composite engineers still commonly fabricate parts with “classic” ply layups (e.g. four ply laminate with fiber orientation angles: [-45/0/45/90°]), which may have sub-optimal mechanical performance in the expected loading/use cases. Finding the composite stacking sequence that achieves the best material and mechanical properties possible is a challenging optimization problem characterized by the large domain space involved in solving an inverse design problem. This paper introduces a novel approach to optimizing stacking sequence for composite plate stiffness by applying off-policy deep reinforcement learning (DRL). We formulate the problem as a sequential decision making process. The state of the system is based on the stiffness of the composite for the currently selected stacking sequence and the action is to select a new stacking sequence using our reward function, formulated as the offset, normalized stiffness modulus. We compare our DRL model to two classical stacking sequences, a randomized baseline model, and a competitive genetic algorithm (NSGA-II). For maximizing longitudinal composite plate stiffness, the DRL model finds the optimum solution for all ply thicknesses and the genetic algorithm comes within 0.5% of the optimum. The DRL model determines the optimum stacking sequence significantly faster and is less sensitive to parameter tuning. The DRL model and the genetic algorithm both outperform the random baseline algorithm by over 5.7 standard deviations in the most conservative case. This research demonstrates the ability of DRL to effectively and efficiently optimize composite laminate stacking sequence.