Session: 36-04 Neural-Network based approaches (3)

Paper Number: 129190

129190 - Centrifugal Compressor Performance and Flow Path Generation With Artificial Intelligence 

Centrifugal compressors are widely used turbomachines in various fields, such as automotive, refrigeration, aerospace, HVAC, and others. All compressors usually represent a component of some system, e.g., part of a turbocharger in an internal combustion engine, part of a refrigeration cycle, etc. Every system has to satisfy the system requirements and constraints. The compressor parameters are determined as a result of system analysis with 0D or 1D solvers. At the same time, the compressor performance as well as other components influence the results of system analysis. The system analysis usually starts with compressor performance taken from some prototype or with some rough assumption based on the engineer’s experience and prior domain knowledge. Often, such assumptions are far from the actual performance of the compressor. As a result, many iterations are often required to get a satisfactory agreement between system parameters and component parameters meeting the requirements and constraints of the entire system. Therefore, the design and development projects of modern centrifugal compressors are long in time and extremely expensive.

The authors' previous work was devoted to the solution of an essentially similar problem for multistage axial compressors with multiple variable guide vanes and airbreathing engines for aircraft [GTP-23-1376]. However, concerning centrifugal compressors, there is still a gap, because previous works in this field were either focused on the development of some kind of simplified performance models, or machine learning models for specific compressors (where it is required to train new models for each new compressor), which lacked accuracy and insufficient level of details about the geometry. The goal of the research described in this paper is to bridge the gap and develop highly accurate single-stage centrifugal compressor performance and flowpath geometry generation models leveraging artificial intelligence technology and enable the opportunity to avoid expensive design-build-test iterations and dramatically reduce the time and cost of developing modern centrifugal compressors.

In this paper, the various configurations of centrifugal compressors are listed, and the configurations of interest are described. The selection of input and output variables that provide sufficient information about compressor design along with the respective ranges is justified.  The automated workflows for centrifugal compressor design and performance data generation for neural network training are described.  The approaches for data preprocessing that enable high-accuracy predictions are provided. The peculiarities of the application of the AI technology developed by authors for axial compressors and its adaptations to centrifugal compressors are discussed. The architectures of the final geometry and performance models obtained with the optimal model search and training algorithm are provided. The analysis of the trained model accuracy as well as the technique for quantitative assessment of prediction reliability is provided. The utilization of the centrifugal compressor AI model integrated into a gas turbine engine simulation environment is demonstrated.

Presenting Author: Leonid Moroz SoftInWay, Inc.

Presenting Author Biography: Leonid Moroz, Ph.D., is the founder and CEO of SoftInWay, Inc.. Upon graduation from Kharkiv Polytechnic Institute (Ukraine) in 1982, Dr. Moroz held various positions at NPO TURBOATOM before founding SoftInWay to design, analyze, and optimize axial and radial turbines, pumps, and compressors. He possesses over 30 years of engineering experience including the development of turbopumps, gas turbine engines, and the design and modeling of turbines for mechanical drives and nuclear power plants. Under his leadership, the AxSTREAM® platform for design, analysis, and optimization of turbomachinery and related subcomponents and systems was developed and implemented globally by turbomachinery OEM and R&D organizations. He is an expert in the internal aerodynamics/hydrodynamics, and rotor dynamics of rotating machinery, and has successfully led multiple projects focused on the design of turbomachinery for aerospace and ground applications. Dr. Moroz holds master's and Ph.D. degrees from the National Technical University “KhPI” (Kharkiv, Ukraine).

Authors:

Valentyn Barannik SoftInWay Switzerland GmbH
Maksym Burlaka SoftInWay, Inc.
Bohdan Lysianskyi SoftInWay Switzerland GmbH
Leonid Moroz SoftInWay, Inc.