Design optimization of axial slot casing treatment and blade in an axial compressor based on deep learning and reinforcement learning

Abstract

Casing treatments have been identified as a promising approach to broaden the operational stability range of compressors by influencing the flow field and delaying the onset of rotating stall. In this study, an integrated optimization of axial slot casing treatment and blade is employed to improve the stall margin without peak efficiency penalty. The casing treatment is defined by 2 B-spline curves, and the blade is parameterized by free form deformation. A multi-objective optimization platform, leveraging machine learning techniques, is developed to facilitate this process. Stall margin improvements and efficiency are predicted using a transformer encoder model with an embedded multi-head self-attention mechanism. The optimization process, driven by reinforcement learning algorithms, aims to maximize stall margin improvement, with policy updates implemented using Proximal Policy Optimization (PPO) algorithms. The performance of the optimal design is further validated through numerical simulations, demonstrating a 13.1 % increase in stall margin without any penalty on peak efficiency. Detailed analysis of the flow field reveals a reduction in the intensity of the tip leakage flow, accompanied by an enhancement in the axial momentum of the main flow. As a result, the interface between the main flow and the tip leakage flow shifts toward the trailing edge. By reducing the influence of the vortex core, additional losses induced by the casing treatment are effectively counteracted, thereby preserving peak efficiency.