Axial Compressor Map Generation Leveraging Autonomous Self-Training AI

Abstract

The paper describes the study performed by SoftInWay in the scope of the Phase I SBIR project funded by NASA. The project was dedicated to a study of optimization of the variable geometry reset angle schedules with the use of innovative autonomous AI technology. In the scope of the project, an automated compressor performance data generation workflow was developed. Three highly loaded multistage axial compressors were designed. The developed workflow was used to generate the training, validation, and test data sets for all three compressors. Multiple different architectures of artificial neural networks were studied, and parametric models for the representation of performance speedlines were developed. Utilizing the developed approaches, the artificial neural networks were trained for all three compressors to predict their performance with a relative error below 3 %. The trained neural networks were successfully used in the optimization of the variable inlet guide vanes and variable stator vanes reset angle schedules with a relative error of total-to-total pressure ratio prediction below 2 % for most of the points and relative error of total-to-total efficiency prediction below 1 % for all the points of the operational line. The capability of the developed AI models to accurately predict the optimal combination of reset angles and efficiency of the axial compressor with multiple vanes controlled independently allowed doing quick evaluations of efficiency and stability margins. The availability of such information enables the opportunity to make technical-economical decisions about the reasonability of implementation of independent variable vanes and their number during engine system analysis.