Exploiting Deep Learning for the Optimization of Transonic Centrifugal Impellers

Artificial intelligence is a powerful tool across numerous scientific disciplines. Within turbomachinery, the application of Artificial Neural Networks (ANNs) stands out for its capacity to facilitate the optimization of stage components. ANNs can reveal complex nonlinear relationships between geometric features, which can pose challenges even for experienced designers to discern. This paper explores impeller optimization, leveraging the capabilities of ANNs. Thanks to a parameterization scheme tailored for 3D geometries, neural network training utilizes a richly diverse database of geometric configurations. The investigation of the resulting Pareto front results in the discovery of highly efficient solutions across diverse geometrical configurations. Crucially, the utilization of ANNs yields multiple solutions with comparable performance, affording designers the flexibility to select solutions tailored to specific requirements, such as achieving a balance between maximum efficiency and operating range. Notably, the analysis of flow fields reveals how artificial intelligence uncovers diverse solutions to maximize efficiency. The paper introduces three optimized solutions for a high-Mach number impeller, each highlighting different blade geometries that impact flow fields and blade loadings within the transonic region. While sharing certain geometrical features, the three cases demonstrate notable variations in blade design. Performance is evaluated by analyzing the polytropic efficiency, work coefficient, and polytropic head coefficient curves, which allow a comparative evaluation of the maximum efficiency and operating range of the three impeller models.