APPLICATION OF REINFORCEMENT LEARNING IN 1-D AERODYNAMIC DESIGN OF AXIAL COMPRESSOR

The traditional aerodynamic design of compressors depends on the experiences of the designer and many optimization calculations. As a result, the performance of the compressor depends on the design level of the designer. A new design method based on reinforcement learning is proposed for the 1-Dimensional aerodynamic design of axial compressors. Deep Deterministic Policy Gradient (DDPG) algorithm is used to extract design experience from the 1-D experiences design program HARIKA. After training, the efficiency was increased by 2.0% relative to the initial value, and the design margin was 23%, which met the design requirements. Under the same design time, the efficiency of new method optimization results was 0.2% lower than the traditional method's results following by better pressure ratio performance at the design point, which proved the effectiveness of the new design method. Furthermore, the approach gained the redesign ability of after training. Compared to the Differential Evolution (DE) optimization results (16000 steps), the efficiency of DDPG redesign results was 0.3% lower and the number of steps is 20, which verified the quick redesign ability of the new design method.