Deep Reinforcement Learning for Composite Material Optimization

Abstract

Structural optimization has long been a complex and time-consuming task as it requires many designs to be verified through finite element method (FEM) prior to achieving an optimal solution. This often necessitates a surrogate method that could extract key features in a structure and determine suitable designs more efficiently. In this paper, we introduced deep reinforcement learning (DRL) by combining deep Convolutional Neural Network with Reinforcement Learning algorithm. To validate the optimization accuracy, the deep Convolutional Neural Network is trained based on four different types of geometric models that are commonly used in the analysis of composite materials: plate with circular cutout model, square packed fiber reinforced, hexagonal packed fiber reinforced and hollow particle reinforced composite models. Our result will show the DRL framework could accurately and efficiently propose the optimal design for different composite models