Generative Design of a Gas Turbine Combustor Using Invertible Neural Networks

P. P. Krueger, Hanno Gottschalk, Bastian Werdelmann, Werner Krebs

TL;DR

This paper introduces a novel generative design approach for gas turbine combustors using Invertible Neural Networks (INNs). By training an INN on a dataset of simulated combustor designs and their performance characteristics, the method can generate multiple valid design alternatives that meet specific performance requirements, significantly accelerating the design process for challenging applications like hydrogen combustion.

What Problem Does This Paper Address?

The transition to cleaner energy sources, particularly the combustion of 100% hydrogen in gas turbines, necessitates a complete redesign of combustion systems to ensure stable operation and prevent issues like flashback, while also minimizing NOx emissions. Traditional design methods, which rely on iterative simulations and parameter sweeps, are time-consuming and often yield only a single satisfactory design, failing to explore the full design space or transfer knowledge effectively between different engine classes. This paper addresses the need for a more efficient and comprehensive design methodology that can rapidly generate a diverse set of viable combustor designs meeting stringent performance criteria.

Methods

The core of this work is the application of Invertible Neural Networks (INNs) for generative design. INNs are a class of deep learning models that learn a bijective mapping between a complex data distribution (in this case, combustor designs and their performance labels) and a simpler, known distribution. This invertibility allows the model to perform both forward (design to performance) and backward (performance to design) transformations.

The authors train an INN on a dataset generated from CFD simulations of parameterized combustor geometries. The INN is trained in both a "normalizing direction" (predicting performance labels from design parameters) and a "generative direction" (generating design parameters from performance labels and a latent noise vector). This dual training approach, using a combination of supervised (MSE) and unsupervised (MMD) losses, ensures that the INN accurately captures the relationship between design and performance while also learning to generate diverse and realistic designs. The architecture of the INN is based on coupling blocks, which are designed to be invertible and computationally efficient.

Datasets

The primary dataset used in this study was generated through Computational Fluid Dynamics (CFD) simulations of parameterized gas turbine combustor geometries. This dataset consists of 6 independent geometrical parameters and 3 performance labels (pressure loss, unmixedness, and thermoacoustic growth rate). The authors also employed surrogate models (fully connected neural networks) trained on this dataset for faster prediction of performance labels, which were crucial for hyperparameter tuning and data augmentation. The augmented dataset, created using these surrogate models, contained 200,000 data points for training the final INN.

Key Insights and Ideas

  • Bridging the Design-Performance Gap with Invertibility: The key innovation is using the invertibility of neural networks to directly generate designs from desired performance metrics, overcoming the information loss inherent in traditional forward-only modeling.
  • Generative Design for Complex Systems: The paper demonstrates that generative AI, specifically INNs, can be effectively applied to complex engineering design problems like gas turbine combustors, moving beyond typical image or text generation tasks.
  • Accelerated Design Exploration: By generating multiple design alternatives rapidly, INNs enable a more thorough exploration of the design space, allowing engineers to consider a wider range of options and trade-offs.
  • Knowledge Transfer Potential: The generative approach has the potential to learn and transfer design knowledge across different engine classes, reducing the effort required for redesigns.
  • Data Augmentation for Robustness: The use of surrogate models for data augmentation significantly improved the robustness and performance of the INN, allowing it to learn from a much larger dataset than would be feasible with direct CFD simulations.

Key Results

  • The trained INN was able to generate diverse combustor designs that met specified performance targets for pressure loss, unmixedness, and thermoacoustic stability.
  • The INN approach significantly outperformed a Gaussian Process (GP) baseline in terms of accuracy for generating designs that meet target performance labels, achieving substantially lower mean absolute errors across all labels. For example, the relative difference in error for UM was up to 94.85% in favor of the INN.
  • The generative process with INNs is nearly instantaneous for inference, allowing for interactive design exploration, whereas the GP baseline required seconds per design.
  • The study successfully generated designs targeting both stable (G < 0) and unstable (G > 0) thermoacoustic conditions, demonstrating the method's versatility.

Limitations and Future Work

  • Dataset Size and Simulation Cost: While the paper demonstrates the effectiveness of data augmentation, the initial dataset generation through CFD is computationally intensive and time-consuming.
  • Surrogate Model Accuracy: Although the surrogate models provided sufficient accuracy for tuning and augmentation, their inherent errors can influence the final INN performance, particularly for labels with higher prediction uncertainty like thermoacoustic growth rate (G).
  • Future Work: The authors plan to apply this methodology to the design of premixing passages for burners capable of 100% H2 combustion. They also intend to incorporate flashback propensity as an additional performance parameter, potentially using experimental data alongside simulations. The potential for extending the methodology to include more design parameters and labels is also highlighted. The authors also suggest that the generated design solutions can be used to further extend the training dataset, establishing an active learning approach.

Summary updated

Abstract

Abstract The need to burn 100% H2 in high efficient gas turbines featuring low NOx combustion in premix mode requires the complete redesign of the combustion system to ensure stable operation without any flashback. Since all engine frames featuring a power range from 4 MW up to 600 MW are affected, a huge design effort is expected. To reduce this effort, especially to transfer knowledge between the different engine classes, generative design methods using latest AI technology will provide promising potential. In this work, this challenge is approached utilizing the current advances in generative artificial intelligence. We train an invertible neural network (INN) on an expandable database of geometrically parameterized combustor designs with simulated performance labels. Utilizing the INN in its inverse direction, multiple design proposals are generated, which fulfill specified performance labels.
BibTeX
@article{krueger2024generative,
  author = {P. P. Krueger and Hanno Gottschalk and Bastian Werdelmann and Werner Krebs},
  title = {{Generative Design of a Gas Turbine Combustor Using Invertible Neural Networks}},
  journal = {Journal of Engineering for Gas Turbines and Power},
  year = {2024},
  doi = {10.1115/1.4066294},
}
Figure