Low-cost and generalizable optimization for the centrifugal pump using deep reinforcement learning

Designing a centrifugal pump inevitably involves geometry choices that, if set by experience or random sampling, leave hydraulic performance uncertain. Optimization is a solution, but current optimization methods rarely combine low computational cost with generalization ability to unseen geometries. This paper proposes a deep-reinforcement-learning-based (DRL-based) optimization method for a centrifugal pump that meets both requirements. First, to reduce computational cost, the three-dimensional centrifugal pump is simplified to a two-dimensional impeller that preserves its essential flow characteristics. The feasibility of this simplification is confirmed by the almost identical variations of head and efficiency as the flow rate of both the three-dimensional and simplified models. Then, after training the DRL algorithm, the optimized design delivers 23.58 m head and 81% mechanical efficiency at its design point due to flattened blade loading and smoother pressure and velocity gradients around the blades. Compared with the conventional genetic algorithms, the DRL-based method can achieve the same hydraulic performance at a lower computational expense. More importantly, once trained on sufficient data, the DRL-based method generalizes to unseen geometries and consistently converges to the global optimum.