Machine learning-assisted axial fan blade optimization for electronics cooling

In an air-cooling system for electronic devices, the axial fan critically determines energy use, heat dissipation, and noise. To address shrinking cooling spaces and noise challenges, this study integrated CFD methods with Bayesian optimization (BO) to design fan blades achieving high airflow with low noise. A CFD model combining the fan with a finned heat pipe cooler simulated CPU cooling scenarios, obtaining flow rate 𝑄 at 2000 rpm. Aerodynamic and acoustic simulations further provided pressure fluctuations and sound pressure level (SPL). Starting from a commercial blade profile, parameterization using blade wrap angle and quadratic functions reduced dimensionality. These parameters and 𝑄 formed a high-dimensional black-box function. Therefore, a Gaussian process regression (GPR) machine learning model was employed to rapidly identify the optimal Q combination from the explored data. Prototype experiments validated the CFD models. Results showed a 24% increase in Q compared to the original design, reducing CPU surface temperature by over 2 °C under a 280 W heat load. Among blades with >20% 𝑄 improvement, combinations exist where SPL match those of the prototype fan, achieving the optimization of conflicting parameters. Sensitivity analysis clarified the effect of each parameter on 𝑄 , static pressure rise 𝑃 , and SPL, guiding key parameter optimization for axial fans in electronic cooling.