Shape Optimization of a Hydraulic Turbine Using Deep Reinforcement Learning

ABSTRACT

The present study explores the potential of deep reinforcement learning for the geometric optimization of a hydraulic turbine and compares its performance with that of an evolutionary algorithm. Shape optimization in fluid mechanics often leads to high‐dimensional, highly non‐linear problems, requiring efficient methods to approximate global optima. Evolutionary algorithms have widely been established, but they require numerous function evaluations. In turbine design, each evaluation entails a resource‐intensive computational fluid dynamics simulation. The optimization approach based on deep reinforcement learning is grounded in policy gradient‐methods, where the agent learns a policy to generate turbine geometries that maximize a reward derived from the simulation results. The algorithm is designed to iteratively improve the geometry by sampling new configurations based on learned policies, guided by feedback from the computational fluid dynamics evaluation. A comprehensive axial turbine model, parameterized by 30 design variables, is employed as the test case, capturing the key aspects of the turbine's design and functionality. The optimization targets two main objectives: increasing the efficiency of the turbine and minimizing the cavitation volume, with head as side condition. The study reveals a critical limitation: the deep reinforcement learning‐method converges toward regions in the parameter space associated with high average rewards, yet these regions do not consistently contain the highest‐performing turbine geometries. In contrast, the evolutionary algorithm generates designs with superior fitness values, indicating that the deep reinforcement learning‐method favors regions of the design space associated with robustness to variations in design parameters but does not consistently identify the highest‐performing solutions.