Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Giuseppe Bruni, Sepehr Maleki, Senthil K. Krishnababu

TL;DR

This work addresses the industrial need to quickly assess how manufacturing variations, especially tip clearance, affect axial compressor performance. It introduces C(NN)FD, a deep learning surrogate based on a 3D U-Net with residual connections that predicts flow fields from tip clearance and blade-geometry inputs, enabling real-time generation of radial profiles and overall performance metrics. The model achieves CFD-level accuracy in a 1.5-stage compressor context, with low MAE and high R^2 on both local flow fields and derived quantities like mass flow and polytropic efficiency, while delivering inference times under 1 second. The approach promises scalable integration into manufacturing/build processes and can be extended to additional variations and multi-stage configurations in future work, reducing computational costs and environmental impact without sacrificing essential physics-informed outcomes.

Abstract

Deep Learning methods have seen a wide range of successful applications across different industries. Up until now, applications to physical simulations such as CFD (Computational Fluid Dynamics), have been limited to simple test-cases of minor industrial relevance. This paper demonstrates the development of a novel deep learning framework for real-time predictions of the impact of manufacturing and build variations on the overall performance of axial compressors in gas turbines, with a focus on tip clearance variations. The associated scatter in efficiency can significantly increase the CO2 emissions, thus being of great industrial and environmental relevance. The proposed C(NN)FD architecture achieves in real-time accuracy comparable to the CFD benchmark. Predicting the flow field and using it to calculate the corresponding overall performance renders the methodology generalisable, while filtering only relevant parts of the CFD solution makes the methodology scalable to industrial applications.

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

TL;DR

This work addresses the industrial need to quickly assess how manufacturing variations, especially tip clearance, affect axial compressor performance. It introduces C(NN)FD, a deep learning surrogate based on a 3D U-Net with residual connections that predicts flow fields from tip clearance and blade-geometry inputs, enabling real-time generation of radial profiles and overall performance metrics. The model achieves CFD-level accuracy in a 1.5-stage compressor context, with low MAE and high R^2 on both local flow fields and derived quantities like mass flow and polytropic efficiency, while delivering inference times under 1 second. The approach promises scalable integration into manufacturing/build processes and can be extended to additional variations and multi-stage configurations in future work, reducing computational costs and environmental impact without sacrificing essential physics-informed outcomes.

Abstract

Deep Learning methods have seen a wide range of successful applications across different industries. Up until now, applications to physical simulations such as CFD (Computational Fluid Dynamics), have been limited to simple test-cases of minor industrial relevance. This paper demonstrates the development of a novel deep learning framework for real-time predictions of the impact of manufacturing and build variations on the overall performance of axial compressors in gas turbines, with a focus on tip clearance variations. The associated scatter in efficiency can significantly increase the CO2 emissions, thus being of great industrial and environmental relevance. The proposed C(NN)FD architecture achieves in real-time accuracy comparable to the CFD benchmark. Predicting the flow field and using it to calculate the corresponding overall performance renders the methodology generalisable, while filtering only relevant parts of the CFD solution makes the methodology scalable to industrial applications.
Paper Structure (9 sections, 16 figures, 1 table)

This paper contains 9 sections, 16 figures, 1 table.

Table of Contents

  1. Introduction
  2. Tip Clearance variation background
  3. Machine Learning applications to turbomachinery
  4. Novel solution to machine learning scalability
  5. Methodology
  6. Data Generation
  7. The C(NN)FD framework
  8. Results
  9. Conclusion

Figures (16)

  • Figure 1: Overview of the 1.5 stage compressor gas path considered
  • Figure 2: Overview of the CFD domain and locations selected for post-processing, with corresponding axial velocity contours
  • Figure 3: $V_x$ contours - smallest and largest clearance comparison
  • Figure 4: $T_0$ contours - smallest and largest clearance comparison
  • Figure 5: $P_0$ contours - smallest and largest clearance comparison
  • ...and 11 more figures