Deep reinforcement learning for optimal design of compliant mechanisms based on digitized cell structures

Metamaterial mechanisms are micro-architectured compliant structures that operate through the elastic deformation of specially designed flexible members. This study introduces an efficient design methodology for compliant metamaterial mechanisms using deep reinforcement learning (RL). In this approach, design domains are digitized into finite cells with various hinge connections, reformulating the design problem as a combinatorial optimization problem. To tackle this intricated optimization problem, we unfold the domain to transform the design problem into a Markov decision process where the deformation behaviors of the designed compliant mechanisms are computed through finite element analysis (FEA). The digitized cell structures are modeled using 1-dimensional (1D) beam elements, significantly reducing the computational load of FEA. The FEA results are utilized in the deep RL framework to optimize compliant mechanism designs based on specific functional requirements. This methodology is applied to the design of compliant gripper and door-latch mechanisms, exploring the effects of cell tiling direction and penalization strategies for disconnected hinges. The optimized designs generated by deep RL outperform human-guided designs, achieving a 56.3% improvement in rotational compliance for the gripper mechanism and a 2.7-fold improvement in linear compliance for the door-latch mechanism, compared to human-guided designs. The optimized compliant mechanisms are fabricated using additive manufacturing, and their performance as compliant mechanisms is experimentally validated. These findings highlight the potential of RL-based design optimization using digitized cell structures, demonstrating its capability to efficiently design high-performance compliant metamaterial mechanisms while maintaining computational efficiency.