Hybrid Wrap-Based Shape Optimization of a Scroll Compressor with Deep Reinforcement Learning

Designing the profile of a scroll compressor to maximize its capacity within a given maximum diameter of a scroll compressor is crucial for the performance enhancement of the refrigeration cycle. To maximize the volume of suction chambers, a hybrid wrap that consists of nodes connected with multiple curves is applied for the scroll profile. However, the shape of the hybrid wrap is hard to optimize because of the many degrees of freedom. Therefore, this research aims to automate and optimize the shape of the scroll compressor with proper wall thickness, using deep reinforcement learning which is a proper optimization method with unknown constraints and nonlinear objective functions. During the optimization process, the nodes are displaced by the agent of the learning process and connected with smooth curves automatically to reconstruct the scroll profile. As the environment, an analytical pressure model based on mass and energy balance equations is established to derive leakage and structural stress on the scroll compressor. The proximal policy optimization method is adopted to maximize capacity and avoid structural damage, by adopting a penalty on a reward function. The reward evolution of the training and the transition of scroll shape are presented to show optimization results. The optimization method proposed in this study is expected to greatly reduce the difficulty and cost of designing the hybrid wrap-based scroll compressor.