Machine-learning-based indirect aeroacoustic optimization of a marine axial ventilation fan with embedded low-noise constraints

The aerodynamic noise of axial fans has become as critical as their traditional performance metrics such as flow rate and efficiency, yet its inclusion in optimization is hindered by the high computational cost of high-fidelity aeroacoustic simulations. This study proposes a machine-learning-based indirect aeroacoustic optimization framework for a marine axial fan that embeds low-noise constraints derived from empirical aeroacoustic knowledge, without directly incorporating computationally expensive aeroacoustic simulations in the optimization process. First, an automated co-simulation framework is developed by integrating parametric geometry modeling with steady-state computational fluid dynamics to generate the aerodynamic performance dataset. Then, a neural-network-based surrogate model is trained and coupled with the genetic algorithm to perform aerodynamic optimization under low-noise design constraints. To demonstrate the effectiveness of the proposed framework, a commercial marine axial ventilation fan is optimized and manufactured. Experimental results demonstrate an average noise reduction of approximately 1.4 dB and a 10.4% decrease in power consumption relative to the baseline design. Further flow-field analysis reveals that, although no explicit aeroacoustic source mechanisms are incorporated during optimization, the optimized design effectively suppresses vortical structures associated with noise sources. The proposed strategy provides an efficient design framework for integrated aerodynamic–aeroacoustic optimization of axial fans.