Session: 36-07 Machine Learning & Artificial Intelligence Methods - Part 2

Submission Number: 175935

Aerodynamic Optimization of Turbine Blades for Wide-Condition Based on Physics-Informed Deep Learning Method 

Traditional turbine blade optimization is often hampered by complex parameterization, high computational costs, and poor variable operating condition adaptability. In this study, an AI-driven optimization framework that integrates generative design, deep learning prediction, and multi-objective evolutionary algorithm is introduced, and applied to the wide-condition aerodynamic performance improvement of NASA E3 high-pressure turbine blade.

Firstly, a Generative Adversarial Network is used to create an automated parameterization method. This approach reduces the original high-dimensional control parameters to just five latent variables with clear physical meanings, enabling a efficient representation of blade profile. Nextly, a physics-informed T-shaped Convolutional Neural Network (TCNN) is innovatively proposed. It models the complex physical relationships between input geometry and operating parameters, intermediate flow field, and output aerodynamic performance using a parallel-serial hybrid structure. Two parallel sub-networks process the pressure distributions on the blade’s pressure and suction sides, respectively. And pressure field information is then fed into a third sub-network to predict performance metrics like total pressure loss coefficient (C_pt). Compared to the models based on Artificial Neural Network (ANN) and serial Convolutional Neural Network (SCNN) trained on the same data, TCNN shows superior accuracy. Its average relative prediction error in high-loss condition is 35% lower than ANN and 50% lower than SCNN. Furthermore, the TCNN maintains a prediction error below 9% for all samples, while the other models exceed 10% for some samples and consistently underestimate the true losses. In the final optimization stage, TCNN model serves as a fast evaluator. With Non-dominated Sorting Genetic Algorithm, a global search of design space is performs to minimize the weighted loss across several typical operating conditions. The optimized blade achieves a 0.63% to 20.8% reduction in the C_pt, with the most significant improvement occurring in high-loss condition. This study provides an effective pathway for the rapid, intelligent optimization of turbine blades for frequent load changes.

Presenting Author: Yiran Li Department of Energy and Power Engineering, Tsinghua University

Presenting Author Biography: Li Yiran (2001– ) is a PhD student at Tsinghua University, China. Her research interests focus on gas turbine blade variable-condition optimization based on AI methods.

Authors:

Yiran Li Department of Energy and Power Engineering, Tsinghua University
Yiyang Wu Department of Energy and Power Engineering, Tsinghua University
XUEYING LI Department of Energy and Power Engineering, Tsinghua University
Jing Ren Department of Energy and Power Engineering, Tsinghua University
Hongfen Tang China Datang Technology Innovation Co., Ltd
Lin Yang China Datang Technology Innovation Co., Ltd
Hongsheng Chen China Datang Technology Innovation Co., Ltd