C(NN)FD -- a deep learning framework for turbomachinery CFD analysis

Overview

• This paper presents a novel deep learning framework for real-time prediction of manufacturing and build variations on the performance of axial compressors in gas turbines.
• The focus is on predicting the impact of tip clearance variations, which can significantly affect efficiency and CO2 emissions.
• The proposed C(NN)FD architecture achieves real-time accuracy comparable to traditional computational fluid dynamics (CFD) simulations.
• The methodology is designed to be scalable to industrial applications by filtering only the relevant parts of the CFD solution.

Plain English Explanation

The paper describes a new deep learning approach to modeling the performance of gas turbine compressors. Gas turbines are used in power generation and aircraft, and their compressors play a crucial role in their efficiency. However, small variations in the manufacturing process can affect the clearance between the compressor blades and the surrounding casing, which can significantly impact the overall efficiency and CO2 emissions of the turbine.

Traditionally, computational fluid dynamics (CFD) simulations have been used to model these effects, but they are computationally intensive and can't be run in real-time. The authors of this paper have developed a deep learning model called C(NN)FD that can predict the impact of these manufacturing variations on compressor performance just as accurately as the detailed CFD simulations, but much faster.

By focusing the deep learning model on only the relevant parts of the flow field, the authors have created a scalable approach that can be applied to real-world industrial problems. This allows manufacturers to quickly assess the impact of their production processes on the efficiency and emissions of the gas turbines they produce, which is crucial for meeting environmental regulations and improving the overall sustainability of power generation and aviation.

Technical Explanation

The paper presents a novel deep learning framework called C(NN)FD for real-time prediction of the impact of manufacturing and build variations on the performance of axial compressors in gas turbines. The focus is on predicting the effects of tip clearance variations, which can significantly affect the overall efficiency and CO2 emissions of gas turbines.

The C(NN)FD architecture combines convolutional neural networks (CNNs) with computational fluid dynamics (CFD) simulations. The CNN component learns to predict the flow field around the compressor blades based on the input geometry and tip clearance variations. This flow field information is then used to calculate the overall compressor performance, such as efficiency.

By filtering only the relevant parts of the CFD solution, the authors are able to make the methodology scalable to industrial applications. The real-time accuracy of the C(NN)FD model is shown to be comparable to traditional, computationally intensive CFD simulations.

The authors demonstrate the effectiveness of their approach on a case study involving an axial compressor for a gas turbine. They show that the C(NN)FD model can accurately predict the impact of tip clearance variations on compressor efficiency, which is crucial for understanding the effects of manufacturing tolerances and optimizing the production process.

Critical Analysis

The paper presents a promising approach to bridging the gap between detailed CFD simulations and the needs of industrial applications. By leveraging deep learning, the authors have created a model that can predict compressor performance in real-time with accuracy comparable to the CFD benchmark.

One potential limitation is the reliance on CFD simulations for training data. While the authors mention that the CFD simulations are validated against experimental data, there may still be some discrepancies between the simulated and real-world performance. It would be interesting to see how the C(NN)FD model performs when evaluated against physical measurements.

Additionally, the paper focuses on a specific case study involving an axial compressor. It's unclear how well the methodology would generalize to other types of turbomachinery or more complex geometries. Further research may be needed to assess the broader applicability of the C(NN)FD approach.

The authors also acknowledge that their approach relies on accessing relevant parts of the CFD solution, which may not always be feasible. Alternative strategies for reducing the computational burden of CFD, such as modal decomposition or GPU acceleration, could potentially be combined with the C(NN)FD framework to further improve its scalability and efficiency.

Conclusion

This paper presents a novel deep learning framework, C(NN)FD, for real-time prediction of the impact of manufacturing and build variations on the performance of axial compressors in gas turbines. The focus on tip clearance variations is particularly relevant, as these can significantly affect the efficiency and CO2 emissions of gas turbines.

The authors have successfully demonstrated the ability of the C(NN)FD model to achieve real-time accuracy comparable to traditional, computationally intensive CFD simulations. By filtering only the relevant parts of the CFD solution, the methodology is designed to be scalable to industrial applications.

This research represents an important step towards bridging the gap between detailed simulations and the needs of industry, allowing for rapid assessment of the effects of manufacturing processes on gas turbine performance. The potential to improve the efficiency and sustainability of power generation and aviation through this approach is compelling and worth further exploration.