Using AI for Loss Prediction in a Hybrid Meanline Modelling Method to Deliver Improved Turbocharger Map Prediction

Abstract

Meanline modelling approaches remain attractive due to an unrivalled ability to predict full turbine performance maps quickly compared to high fidelity approaches such as CFD, especially in the preliminary design process. As improvements in performance on a component level approach a point of diminishing returns, the ability to efficiently optimise the complete charging system for a given duty is a topic that is attracting significant research interest. In order to achieve this aim, the current piece of work seeks to combine the advantages of machine learning techniques and physical meanline modelling to facilitate faster, more accurate predictions of complete turbocharger maps. This paper presents a novel methodology for turbocharger turbine rotor performance prediction based on hybrid modelling. The radial turbine rotor for turbocharger applications is parameterized in order to conduct CFD simulations for a variety of turbine geometries. The results of the CFD simulations make up a database to be fed into an artificial neural network, by which the rotor losses can be predicted when a specific geometry and operating condition are given. The predicted losses are then utilized in the meanline code, substituting the conventional loss models to predict the turbine rotor performance. One novel aspect of this work is that the losses occurring in the rotor are split into two portions, before and after the throat respectively. Another novel aspect is that the blockage level was implemented as a variable value in this meanline model to reflect the changing secondary flow field at the rotor throat at different operating conditions. This hybrid meanline modelling method is evaluated on several unseen test cases and shows good agreement with the CFD results from the perspective of total-to-static efficiency and mass flow rate. The hybrid meanline modelling method displays great potential in wide range radial turbine performance prediction with enhanced accuracy in comparison to traditional approaches.