Search

Aiming at the problems of a long design period and imperfect surrogate modeling in the field of airfoil design optimization, a convolutional neural network framework for airfoil design and performance prediction (DPCNN) is established based on the deep learning method. The airfoil profile parameterization, physical field prediction, and performance prediction are achieved. The results show that the DPCNN framework can generate substantial perfect airfoil profiles with only three geometric parameters. It has significant advantages such as good robustness, great convergence, fast computation speed, and high prediction accuracy compared with the conventional machine learning method. When the train size is 0.1, the predicted results can be obtained within 5 ms. The prediction absolute errors of physical field of most sample points are lower than 0.002, and the relative errors of aerodynamic performance parameters are lower than 2.5%. Finally, the optimization attempt of operating parameters is completed by gradient descent method, which shows good stability and convergence. Overall, the DPCNN framework in this paper has outstanding advantages in time cost and prediction accuracy. [ABSTRACT FROM AUTHOR]

The thermal-fluid-structure interaction coupling method is adopted in this paper to analyze the sealing performance of spiral-grooved dry gas seal. Effect of grooved ring rotation on sealing performance is discussed in detail. The real gas model is applied to obtain the accurate sealing performance under different inlet temperature and pressure with air and SCO 2 as working fluid respectively, which provides a vital reference for SCO 2 seal application. The results obtained with the thermal-fluid-structure interaction coupling method are reliable. It is found that grooved ring rotation has a significant influence on sealing performance especially for large film thickness. The pressure and temperature of the model which is grooved on the rotating ring are higher than those of the model which is grooved on the stationary ring, but the open force and leakage are opposite. In addition, SCO 2 model and air model show a significant difference on the sealing performance and the sealing performance of SCO 2 model is easier to be affected by the inlet temperature and pressure compared with air model. [ABSTRACT FROM AUTHOR]

Aerodynamic design optimization of the blade profile is a critical approach to improve performance of turbomachinery. This paper aims to achieve the performance prediction with deep learning method and realize fast design optimization of a turbine blade. Two parameterization methods based on geometric relationships (PGR) and neural network (PNN) are proposed, which can generate smooth and complete blade profiles. A dual convolutional neural network (DCNN) is constructed to predict the physical fields and aerodynamic performance. The implementations of DCNN are accomplished based on the datasets generated by the two parameterization methods respectively, which are called PGR-DCNN and PNN-DCNN model. Results show that the prediction accuracy increases and then keeps stable as train size increases. The two models can offer the detailed physical field distribution within 3 ms and accurately predict the aerodynamic performance. The prediction errors of performance parameters for 99% samples in validation set are less than 0.5% with PGR-DCNN model, which are significantly better than conventional machine learning methods. Finally, based on the accurate predictive models, the gradient-based design optimization for rotor blade profile is completed in 38 s. The efficiency of the two optimal blades reaches 89.29% and 88.92% respectively, which verifies the feasibility of our method. • Two parameterization methods based on geometric relationships and neural network are proposed. • The DCNN model is constructed to reconstruct physical fields and predict performance. • High prediction accuracy and fast calculation speed are achieved by DCNN model. • The gradient descent method is adopted to conduct the optimization of the turbine blade profile. [ABSTRACT FROM AUTHOR]

End wall contraction is an effective approach to reduce the substantial secondary loss in the cascade with low aspect ratio. This paper proposed a series convolutional neural network architecture (SCNN) to improve the design optimization problem of end wall profile of a turbine stator blade. The specific implementation of the architecture is discussed in detail. The effect of the train size and sensitivity analysis of design variables are carried out. Finally, the optimization is completed by the gradient descent method, and the aerodynamic performance of end wall profile before and after optimization is compared. It shows that the SCNN architecture performs outstandingly with 30% training data. It can quickly and accurately provide rich flow field information and performance parameters within 3 ms after training. When the train size is 0.3, the mean prediction errors of mass flow rate and efficiency of each sample are lower than 0.1%, which performs significantly better than the Artificial Neural Network and Gaussian Process Regression model. When the stator blade adopts optimized contracted end wall profile, the power and efficiency are raised by 4.43% and 1.39% respectively with the mass flow rate only changing by 1.48%, which verifies the feasibility of the SCNN architecture. • The SCNN model is established to reconstruct flow field and predict performance. • High prediction accuracy and fast solving speed are achieved by SCNN model. • The sensitivity analysis of the design variables is carried out. • The optimization of end wall profile is completed with gradient descent method. [ABSTRACT FROM AUTHOR]

Obtaining the off-design characteristics of core components such as turbines and compressors is the basis of off-design analysis for energy systems. However, the characteristics are difficult to accurately acquire in the initial design stage of turbomachinery. Based on deep learning techniques, an accurate and rapid field reconstruction and off-design aerodynamic performance prediction method is proposed. First, a Generative Adversarial Network with the added Bezier layer is employed to establish a database of turbine blade profiles. Then, a Dual Convolutional Neural Network (Dual-CNN) is established to reconstruct the pressure and temperature fields as well as predict the off-design performances of different profiles and working conditions. Based on the above two kinds of neural networks, a turbine in a solar-based supercritical carbon dioxide Brayton cycle is taken as an example. The field reconstruction and off-design performance prediction are conducted on the basis of the established rotor blade profile database. The accuracy of field reconstruction is guaranteed. The off-design performance prediction of the established Dual-CNN indicates that the example blade profile is suitable for operation with larger mass flow rate. Compared with the traditional method, Dual-CNN can reduce the off-design analysis time of one blade geometry from 38.4 h to 7.68s. • A Dual-CNN is established to predict the off-design performance and reconstruct the fields for turbomachinery. • A Bezier-GAN and the geometric transformation are employed to generate smooth profile database. • The designed model can accurately reconstruct the fields. • The time cost of off-design analysis is reduced from 38.4 h to 7.68s. [ABSTRACT FROM AUTHOR]

Based on the deep learning method, the physical information of Reynolds equation is introduced into the neural network, and a deep learning network frame (ReF-nets) is built to predict the flow fields and aerodynamic characteristics of gas bearing. The prediction results of neural network have high enough accuracy comparing with those from finite difference method (FDM). Furthermore, a comprehensive investigation of unsupervised learning, supervised learning and semi-supervised learning in PDE (partial differential equation) solution prediction is carried out, which is the first time these three learning strategies are fully compared in PDE solving field to our best knowledge. The advantages and disadvantages of the three learning methods are illustrated in detail from the perspectives of physics interpretability and prediction accuracy. It is found that the unsupervised learning method has the strongest physics interpretability in predicting the flow field. The supervised learning method has the highest prediction accuracy, but it has nearly no physics interpretability. While semi-supervised learning method can take into account the advantages of both methods, and perform well when there are some known flow field data. • A physics informed neural network of Reynolds equation (ReF-nets) is built. • Comparison of three learning strategies in predicting PDE solution is carried out. • MC-ReF-nets realize the function to solve the PDE under multiple cases. • This study aims to solve the actual gas bearing problem, not merely in the theory. [ABSTRACT FROM AUTHOR]