Machine learning-assisted shape morphing design for soft smart beam

Programming the shape of soft smart materials is a challenging task due to the enormous design space involved. In this study, we propose a novel approach to determine applied stimuli that enable the desired actuated shapes of soft smart materials. Our approach combines finite element methods (FEM), deep neural networks (DNN), and particle swarm optimization (PSO). By employing beams made of dielectric elastomer (DE) as models, we partition a DE beam into multiple actuating units, allowing independent actuation through the application of paired electrical stimuli. FEM is subsequently employed to compute the actuated shapes in response to various applied stimuli. The selection of these stimuli is accomplished through the utilization of Latin hypercube sampling. By utilizing the FEM-derived data, we have developed a machine learning (ML) surrogate model that integrates a long short-term memory (LSTM) network with a fully connected neural network (FCNN). Finally, PSO is employed to determine the optimal applied stimulus that yields the desired actuated shape. The LSTM-FCNN surrogate model is utilized to evaluate the fitness of PSO. The ML-PSO framework exhibits remarkable performance and efficiency in the context of the inverse design of soft smart beams.