AI-Prediction of Pressure Characteristics for Centrifugal Fans

Abstract The flow in a centrifugal fan is in general three-dimensional, viscous, unsteady, turbulent and therefore not easy to predict. Traditional design methods are based on analytical, physics-based approaches and empirical design rules. Designed machines are always tested with experiments. Since the beginning of this century, more and more digital experiments (numerical simulations) have been carried out and physical measurements on the test rig have only been performed on individual prototypes. With modern methods of computational fluid dynamics (CFD) and the simultaneous increase in computing resources, it is possible to obtain results faster and at the same time more and more accurately. In the present contribution, due to the application of machine learning methods, a systematically generated data set of CFD results should lead to a modern data-driven design method. For this purpose, a parameterized model of the centrifugal fan with circular arc blades was designed where the geometric parameters outlet diameter d2, outlet width b2 and outlet blade angle beta2 could be systematically varied. Subsequently, a targeted variation was carried out using the design of experiments (DOE) method and the pressure characteristic (Δp – qV-diagram) was simulated using CFD. These CFD simulations were performed using the commercial solver ANSYS CFX 2024R2, whereby the CFD model was previously developed based on a detailed grid independence study and validated with real measurement data. This database was then used as a training data set for the AI-based pressure characteristic prediction. This was because each pressure characteristic (output parameter) contains a specific set of parameters (input parameter) that was evaluated using machine learning methods. A commercial ML-tool was used for this analysis. This tool uses a combination of neural networks and Gaussian processes. The aim was to predict the resulting pressure characteristic for a specific new parameter set that was not part of the training data. This procedure was performed for several different parameter configurations and then recalculated and validated by CFD simulations. The theory and the numerical results are explained and shown in detail.