A region-segmentation combinational loss model based on data-driven machine learning for a boundary layer ingestion fan

Abstract

The boundary layer ingesting (BLI) fan is continuously working under distorted inflows. The rotor blade moves across the distorted and undistorted regions and the resultant off-design condition makes the loss present a notable non-uniform distribution feature. Accurately capturing the loss is very important for evaluations of BLI fan performances in preliminary design stage. However, unsteady Reynolds-Averaged Navier-Stokes (URANS) simulations are extremely computational expensive even though they can well predict the loss, and the traditional empirical or physics-based loss models are usually fade when predicting loss variation trends and absolute magnitudes under complex inflow conditions. This paper develops a data-driven machine learning based region-segmentation combinational loss model to realize a fast and accurate prediction of the aerodynamic loss in a BLI fan rotor. According to the distinctions of loss sources in different spanwise fractions, a region segmentation idea is first used to create local loss model in each sub-region, and then a full-span combinational prediction model is constructed through a data-driven based neural network approach. Results show that the data-driven based region-segmentation combinational loss model presented in this work performs better than the traditional empirical or physics-based loss models in terms of predicting both the loss variation trends and absolute magnitudes. Compared with the high-fidelity URANS simulations, the averaged loss prediction errors in a BLI fan rotor are kept within 10 % under different blade load and inflow distortion conditions.