Deep Reinforcement Learning for Digital Materials Design

Designing composites has been a research topic of interest in the field of materials science. As an elegant mathematical representation for composites, the concept of digital materials (DMs) was developed to express structures with complex geometries and various material distributions. DMs have a vast design space to achieve targeted physical properties, which makes it challenging for solving inverse problems. Here, a deep reinforcement learning (DRL) scheme is utilized to automate the DM design process without the designer’s prior knowledge. Based on the reward signal of structural mechanical property changes, DRL algorithms can initiate new design patterns in a self-updating process. As a demonstration example, a DM system composed of two different materials are selected as testing environments with three different levels of design space sizes. The collaborative deep Q network (DQN) architecture is developed to comprise two cooperative agents for two types of element-level modification operations to satisfy the design constraints, such as material fraction. The quality of each composite pattern is calculated through the finite element analysis (FEA) simulation. Results show the proposed approach can effectively handle the complex state-action space problems for the digital material design process with significant computation advantages, compared with those of the genetic algorithm with a 15.9% final design quality enhancement. As such, this new class of DRL scheme could be a powerful tool to enable the autonomous discovery process for next-generation free-form DM designs.