Turbine blade optimization considering smoothness of the Mach number using deep reinforcement learning

The Mach number distribution is an important factor in turbine design. Smooth Mach number distribution is an important feature of blades with small aerodynamic loss. However, existing gradient-based methods, such as adjoint methods, are problematic when the Mach number distribution's smoothness is considered an objective function. In the present study, we propose an optimization method based on deep reinforcement learning (DRL) to maximize the smoothness of the Mach number distribution. In addition, the task was assigned repeatedly with slightly different conditions rather than viewing the turbine design task as a single optimization task. A situation such as this requires each optimization problem to be solved in a short time. For this purpose, we propose an optimization method based on DRL. The computation time for training the model is long; however, once trained, the model can be reused to solve different problems, and the computation time to solve each optimization problem is short. The numerical examples show that the trained agent successfully improves the smoothness, and the model can be reused for various tasks under different conditions.