Axial Compressor Map Generation Leveraging Autonomous Self-Training Artificial Intelligence. Phase 2

Abstract This paper discusses a study performed by SoftInWay as part of a Phase II SBIR project funded by NASA. In contrast with the Phase I project (published in paper GTP-22-1328) where three discrete compressors were considered, the Phase II study was focused on addressing the problem of axial compressor long development time and cost with the use of AI models capable of predicting the geometry and performance of various multi-stage axial compressors with multiple variable vanes. The applicability of the AI models to various compressors enables the opportunity to avoid iterations between engine cycle analysis and compressor design. In this paper, automated compressor design and performance generation workflows are described. The approach for autonomous selection of the architectures and hyperparameters of Machine Learning (ML) models is explained. The uncertainty quantification techniques are considered. The developed ML-powered methods for compressor geometry prediction are discussed. The ML models' accuracy values and representations of typical geometry and performance predictions are given. The utilization of the ML models in engine cycle analysis is discussed.