Session: 31-02 Compressor Design

Paper Number: 152689

Efficient ML-Based Prediction of Turbomachinery Blade Performance With B-Spline Surface Representation 

ABSTRACT FOR ASME TURBO EXPO 2025

 

EFFICIENT ML-BASED PREDICTION OF TURBOMACHINERY BLADE PERFORMANCE USING A CAD-CENTRIC FREEFORM SURFACE DESCRIPTION

Hamun Bertram¹, Christian Janke², Robert Flassig¹, Peter Flassig¹

¹Brandenburg University of Applied Sciences

²Rolls-Rolls Deutschland Ltd & Co KG

 

Rapid and accurate prediction of turbomachinery blade performance is crucial for accelerating the blade design process and enhancing optimization efficiency. Recent machine learning (ML) approaches, such as those referenced in [1] and [2], still require a significant number of high-fidelity Computational Fluid Dynamics (CFD) simulations, which are computationally expensive and limit scalability.  One reason for the high demand for training data in data-driven AI methods, as proposed by [1] and [2], may be the extensive number of features, given that blade geometries are represented by dense surface point-clouds.

To address this challenge, we propose a novel approach that utilizes a B-Spline surface representation for defining blade geometries. Using control points of B-Spline surfaces as input features significantly reduces the dimensionality compared to surface-point-cloud methods, which require dense and computationally expensive datasets, leading to a streamlined and compact geometric description. This reduction in complexity aims to decrease the training time required for ML models, thereby accelerating the overall blade design process and enhancing its feasibility for industrial application.

In this study, we explore the blade design space using Design of Experiments (DoE) by systematically varying key geometric parameters, such as chord length, stagger angle, camber-line, thickness distribution, and span-wise stacking. The open-source ParaBlade tool with its unified parametrization method, resulting in a compact yet precise representation of the blade surface, is employed [3]. Required CFD simulations are conducted using ANSYS CFX to evaluate critical performance metrics, such as polytropic efficiency, across various design configurations, with the NASA Rotor 67 serving as the benchmark case [4].

A systematic analysis is performed to determine the number of training points required to achieve a specified level of prediction accuracy for key aerodynamic performance metrics, generative design efficiency, and optimal Convolutional Neural Network (CNN) architectures. The results from the autoencoder-based CNN approach are compared with a classical Principal Component Analysis (PCA) method [5, 6, 7]. Finally, a CNN-based optimization is carried out, and the results are evaluated against existing optimization outcomes reported in the literature [1, 2].

 

 

References

[1] Gouttiere, A., Alessi, G., Wunsch, D., Zampieri, L., & Hirsch, C. (2023). Realtime CFD Based Shape Optimization Using Geometric Deep Learning for Families of Turbomachinery Applications. In Proceedings of ASME Turbo Expo 2023. GT2023-101938.

[2] Beqiraj, K., Perrone, A., Sanguineti, M., & Ricci, G. (2023). Advantages of Machine Learning Methods in Aerodynamic Blade Optimization. In Proceedings of ASME Turbo Expo 2023. GT2023-102481.

[3] Agromayor, R., Anand, N., Müller, J.-D., Pini, M., & Nord, L. O. (2020). A Unified Geometry Parametrization Method for Turbomachinery Blades. Computer-Aided Design, 133, 102987.

[4] Strazisar, A. J., Wood, J. R., Hathaway, M. D., & Suder, K. L. "Laser Anemometer Measurements in a Transonic Axial-Flow Fan Rotor," NASA Technical Paper 2879, NASA Lewis Research Center, 1989. Available at: https://ntrs.nasa.gov/citations/19900001929.

[5] Shlens, J. (2014). A Tutorial on Principal Component Analysis. arXiv. Available at: https://ar5iv.labs.arxiv.org/html/1404.1100

[6] Wold, S., Esbensen, K., & Geladi, P. (1987). Principal Component Analysis. Chemometrics and Intelligent Laboratory Systems, 2(1–3), 37–52. https://doi.org/10.1016/0169-7439(87)80084-9

[7] Mishra, S., Sarkar, U., Taraphder, S., Datta, S., Swain, D., Saikhom, R., Panda, S., & Laishram, M. "Comparative Study of LBPH, CNN, and PCA Algorithms for Image Analysis and Recognition," 2023 7th International Conference on Computation System and Information Technology for Sustainable Solutions (CSITSS), IEEE, Bangalore, India, pp. 02-04 November 2023. Available at: https://ieeexplore.ieee.org/abstract/document/10334195

 

 

Presenting Author: Peter Flassig Brandenburg University of Applied Sciences

Presenting Author Biography: Prof. Dr.-Ing. Peter Flassig holds a Diploma in Aerospace Engineering from Technical University of Berlin. For several years he did research at Brandenburg University of Technology on multi-disciplinary design optimization, uncertainty quantification, robust design and reliability in the field of jet engines. At the aero engine manufacturer Rolls-Royce he worked ten years in the Design Systems Engineering department with a secondment to the United Kingdom supporting development of aerothermal methods and tools, application and realisation of research and development projects together with academic partners, students doing their degree theses and world-wide engine projects. Since April 2020 he works as Professor at Brandenburg University of Applied Sciences with the main areas of teaching Engineering Design and Machine Elements. His research interests are Industriy 4.0 related and, thus, focused on machine learning, numerical methods in general, probabilistic methods, reverse engineering and the concepts of digital twin and digital thread.
https://www.linkedin.com/in/peter-flassig/

Authors:

Hamun R. Bertram Brandenburg University of Applied Sciences
Christian Janke Rolls-Royce Germany
Robert Flassig Brandenburg University of Applied Sciences
Peter Flassig Brandenburg University of Applied Sciences