Machine learning-enhanced throughflow method for the improved performance prediction of radial impellers

In aerospace applications which include a hybrid compressor consisting of multiple axial stages and one final radial stage, challenges arise in the preliminary design phase in matching the efficiencies of the axial and radial stages across a broad operating range. Low-level calculation tools, such as two-dimensional (2D) throughflow methods, require empirical correlations and usually fail to accurately predict the performance of the radial compressor, especially beyond the design point. The complex flow profile hinders the viability of the empirical correlations. For this reason, the preliminary design of radial compressors is often based on three-dimensional (3D) Computational Fluid Dynamics (CFD) simulations instead. However, CFD simulations are time-consuming, and for routine preliminary design, a rapid - and yet accurate - calculation tool would be advantageous. The present dissertation aims to improve the design and off-design predictive accuracy of throughflow methods for radial compressors, specifically for the radial impeller. A methodology is proposed which calibrates 2D calculations to benchmark CFD results by means of machine learning methods. Relevant performance metrics which are traditionally prescribed through correlations in a throughflow method for radial impellers are the slip factor, the loss coefficient, and the aerodynamic blockage. As a first step, the proposed approach included baseline throughflow calculations by combining established empirical correlations for these three metrics. Subsequently, the baseline values of the metrics were calibrated against reference values from CFD simulations, for which a neural network regression model was used. Applied to two representative test cases, the new throughflow calculations with the calibrated metrics yielded substantially improved predictions of both total-total pressure ratio and isentropic efficiency compared with the initial baseline throughflow results using empirical correlations. The proposed methodology aligned the predicted speed lines for pressure ratio and efficiency with the CFD benchmarks for the majority of operating points considered. The method was also capable of replicating the curvature of the reference speed lines near choke conditions. Still, the most significant disparities in prediction were observed at the limits of the operating range, i.e., in the vicinity of choke and near surge. Between these two limit operating conditions, the relative prediction errors in pressure ratio remained within -2.0% and 0.5%. Across all speed lines, the relative errors in efficiency prediction ranged between -1.0% and 0.5%. To summarize, the proposed machine learning-enhanced 2D throughflow method appears to be sufficiently accurate for a wide range of operating conditions and thus offers a viable alternative to CFD simulations in the preliminary design of radial impellers. Despite the relatively limited number of test cases considered, this approach shows promise for broader generalization and future integration of additional geometries and operating conditions.