| description abstract | Abstract. Machine learning on large datasets allows physical structures that are present in data to be discovered. This provides an opportunity to develop a new generation of compressor preliminary design tools, which have a more physically accurate underlying structure. In this article, a new loss model for preliminary design has been developed, using a data-centric approach, with a more physically accurate loss decomposition. This new loss model is compared to existing preliminary design loss models using a large dataset of Reynolds-averaged Navier–Stokes (RANS) computational fluid dynamics (CFD) solutions. It is shown that the new, physically decomposed, loss model provides more accurate loss predictions at the preliminary design stage, over a wider range of the design space. For instance, the new model is shown to be able to capture the effect on loss when 3D blade design is used, stage loading is changed and the trailing edge thickness relative to the maximum thickness is allowed to vary. The new model is shown to be accurate, over this design space, to within ±9% compared to the accuracy of the model of Wright and Miller, which is only accurate to within ±22%. Furthermore, the physical decomposition of the new model means the model can easily be applied to different datasets and enhance understanding of how design changes influence the sources of loss, giving designers better guidance at the preliminary stages of design. | |
