Towards the Automatic Generation of Transonic Centrifugal Impellers by Deep Learning and Feature Importance Techniques

Abstract In recent years, the use of artificial intelligence (AI) has transformed the field of turbomachinery, enabling the development of advanced tools for optimizing and designing more efficient geometries. Artificial Intelligence techniques, particularly Artificial Neural Networks (ANNs), enable the handling of vast simulation datasets, assisting the design process and improving machine performance. This paper presents an innovative method that exploits databases of CFD solutions to generate optimal geometries for the design. By leveraging the ANN’s ability to capture complex, nonlinear relationships between geometric features in multi-objective problems, this method explores a broad design space efficiently. A feature importance analysis is conducted on the databases, which span the entire operating range of the compressor, to evaluate how geometric parameters influence key performance metrics and to identify those with the least significant impact. Starting from these databases, a general procedure for generating a “family” of transonic impellers is outlined. The procedure is conceived to automatically generate geometries that meet specific flow coefficient values while maintaining a constant relative Mach number at the impeller eye. The resulting family members exhibit similar transonic behaviour and performance characteristics, allowing machines to be tailored to specific operational requirements. This automated design approach provides a valuable tool for improving turbomachinery performance across different applications.