Abstract Radial pumps and compressors are used in various engineering applications, including rocket turbopumps, automotive turbochargers, and refrigeration systems. Several physical effects, including viscous losses, flow separation, compressibility, and rotational dynamics, dominate radial impeller flow, making flow field prediction very difficult and requiring computationally expensive CFD. However, designers typically only require information at specific positions, resulting in most simulation data being unused. Accurately predicting the impeller exit flow field is often key to impeller design. Recent advances in reduced-order modeling and machine learning show promise for a priori flow field prediction. In this study, nine different reduced-order models (ROMs) were created to predict the dimensionless exit flow field of radial flow impellers in real time. The ROMs consist of various linear and nonlinear dimensionality techniques paired with different regressors. Inputs include parameterized, dimensionless impeller geometry based on Bezier control points, number of blades, and dimensionless operating conditions. The ROMs were trained using over 1800 flow fields from high-fidelity CFD simulations representing a large radial flow compressor design space. ROMs were evaluated on relative error, training time, and evaluation time. Principal component analysis coupled with Gaussian process regression (PCA-GPR) emerged as the preferred ROM, training within two seconds, evaluating hundreds of cases in real time, and matching the accuracy of computationally demanding nonlinear alternatives. PCA-GPR predictions show pressure, density, and velocity profiles within 5% average of CFD results. The ROM was validated through four test cases probing robustness across different operating conditions and geometries.
Oliphant et al. (Fri,) studied this question.