Session: 26-01 Probabilistic and Machine Learning Methods Development and Applications

Paper Number: 129046

129046 - A Hybrid Surrogate Modeling Approach for Data Reduction and Design Space Exploration of Turbine Blades 

The conventional iterative geometry optimization process for turbine blades using computer aided engineering (CAE) simulations is cost and time consuming and has limitations in terms of computational requirements and data management. A particularly problematic aspect is that the generated multi-disciplinary computational fluid dynamics (CFD) and finite element analysis (FEA) 3D simulation data cannot be fully reused for future optimizations. The data for multiple geometry variations of the turbine blade, generated in a Design of Experiments (DoE) study, can range from several hundred gigabytes to multiple terabytes. This leads to difficulty in storing and accessing them for a longer period. To address this challenge, we introduce a methodology based on machine learning models for data reduction and forecast of 3D surface data, e.g., temperature and pressure data. Specifically, we build a Variational Autoencoder (VAE) that combines two models: the Encoder and the Decoder. The Encoder transforms the input data into a latent representation of reduced dimensionality in a latent space, and the Decoder reconstructs the input data from the given latent representation in its original space. The latent representation of the input data and the trained VAE together result in much smaller data amounts and file sizes, solving the issue of data storage for future use. Furthermore, we train a feed-forward neural network that learns the mapping of the geometry parameters to the latent space. This additional approach predicts the surface data results for new and unseen blade geometry variations and generates their latent representations without the need for their mesh representation and new simulation runs.

Keywords: Surrogate modeling, Conjugate Heat Transfer, Turbine Blade, Variational Autoencoder, Neural Networks

Presenting Author: Sazeed Sufian Ali Siemens Energy AG

Presenting Author Biography: Name: Ali, Sazeed Sufian

Date of Birth: 07.06.1993

Education:

- Bachelor of Science (B.Sc.) in Transport and Traffic Engineering from TU Berlin.
- Master of Science (M.Sc.) in Aerospace Engineering from TU Berlin.

Work:

- Enrolled in PhD program at Siemens Energy AG in cooperation with the Chair of Data Analysis and Modeling of Turbulent Flows, TU Berlin.
- Department for Tools and Data Application.
- Exploration of 3D data reduction and acceleration techniques for turbine blade optimization with machine learning and AI.

Authors:

Sazeed Sufian Ali Siemens Energy AG
Vikas Shivprasad Yadav Chair of Data Analysis and Modeling of Turbulent Flows, TU Berlin, Berlin, Germany
Behnam Nouri Siemens Energy AG
Abdulla Ghani Chair of Data Analysis and Modeling of Turbulent Flows, TU Berlin, Berlin, Germany