Deep Operator Learning-based Surrogate Models with Uncertainty Quantification for Optimizing Internal Cooling Channel Rib Profiles

Guang Lin, Izzet Sahin, Christian Moya, Amirhossein Mollaali, Guang Lin 0001, Guillermo Paniagua

TL;DR

This paper introduces Deep Operator Networks (DeepONets) and their Bayesian extension (B-DeepONets) to create highly accurate surrogate models for predicting heat transfer and pressure drop in internal cooling channels. These models enable rapid, uncertainty-aware optimization of complex rib geometries, a crucial task for improving the efficiency of gas turbine blades.

What Problem Does This Paper Address?

Advanced gas turbine engines rely on efficient internal cooling channels to manage extreme temperatures. Rib turbulators are commonly used to enhance heat transfer in these channels, but optimizing their complex 3D shapes to balance heat transfer gains with pressure losses is a computationally intensive challenge. Traditional methods require extensive simulations for each design iteration, making comprehensive exploration of the design space impractical. Furthermore, existing surrogate models often lack robust uncertainty quantification, which is vital for avoiding "adversarial" designs that appear optimal but perform poorly in reality. This paper aims to bridge this gap by developing data-efficient, uncertainty-aware surrogate models that can accelerate the optimization of these cooling channels.

Methods

The core of this work lies in the application and extension of Deep Operator Networks (DeepONets). DeepONets are a novel class of neural networks designed to learn mappings between infinite-dimensional function spaces, making them ideal for approximating complex physical phenomena like fluid flow and heat transfer.

  • DeepONet: The paper utilizes DeepONets to learn the mapping from a continuous rib geometry (defined by control points) to the detailed distributions of pressure and heat transfer (Nusselt number) within the cooling channel. This involves two sub-networks: a "Branch Net" that encodes the input rib geometry and a "Trunk Net" that encodes the output location.
  • Bayesian DeepONet (B-DeepONet): To address the limitations of small or noisy datasets and to provide uncertainty quantification, the authors extend DeepONets into a Bayesian framework. B-DeepONets provide credible intervals for predictions, allowing for the identification of potentially adversarial designs and a more robust optimization process. This is achieved using a stochastic gradient replica-exchange Langevin diffusion (SG-reLD) method to sample the posterior distribution of model parameters.
  • Optimization Framework: A gradient-free, constrained cross-entropy method is employed to optimize the rib profiles. This method iteratively samples candidate designs, evaluates their performance using the DeepONet or B-DeepONet surrogates, and refines the search towards optimal solutions. The reward function is based on the thermal performance factor, which balances heat transfer enhancement and pressure drop.

Datasets

The datasets used for training and testing the DeepONet and B-DeepONet models were generated through 2D numerical simulations of a rib-roughened internal cooling channel.

  • Generated Dataset: A total of 275 distinct rib profiles were generated by randomly adjusting control points within predefined constraints. These profiles were then simulated to obtain pressure and Nusselt number distributions.
  • Training/Testing Split: The dataset was split into 90% for training and 10% for testing, using a simple random selection to ensure independence between data points.
  • B-DeepONet Training: For B-DeepONet, a smaller subset of sampled locations (q=100) was used for training to simulate scenarios with even scarcer data.

Key Insights and Ideas

  • Data Efficiency of DeepONets: DeepONets can learn complex operator mappings from relatively small datasets, significantly reducing the need for extensive high-fidelity simulations compared to traditional surrogate models.
  • Uncertainty Quantification for Robust Optimization: B-DeepONets provide crucial uncertainty estimates, enabling the identification and avoidance of "adversarial" designs that might otherwise be selected by deterministic models, especially with limited training data.
  • Accelerated Design Exploration: The combination of DeepONets/B-DeepONets with a gradient-free optimization method allows for rapid exploration of the rib design space, leading to improved thermal performance.
  • Bridging 2D and 3D: While simulations were 2D, the approach is designed to be applicable to 3D problems, with the 2D simulations serving as a validated foundation.

Key Results

  • High Accuracy: DeepONet models achieved mean L₁ trajectory errors of less than 6% for pressure and 8% for Nusselt number distributions on unseen test profiles.
  • Significant Computational Speedup: DeepONet-based surrogates reduced simulation time from hours (numerical approach) to milliseconds (DeepONet), a speedup of several orders of magnitude.
  • Improved Thermal Performance: The optimization process using both deterministic DeepONets and B-DeepONets successfully identified rib profiles that showed enhanced Nusselt number ratios and reduced pressure drop compared to a baseline square rib profile. Figure 12 shows percentage improvements in both metrics.
  • Credible Intervals: B-DeepONet successfully generated confidence intervals that captured the true distributions of pressure and Nusselt number, even for profiles not seen during training.

Limitations and Future Work

  • Limitations:
    • The current study uses 2D simulations, and extending to full 3D would require more computational resources.
  • While data-efficient, obtaining even the generated dataset involved significant numerical simulations.
  • The current optimization is gradient-free; gradient-based methods could potentially offer further improvements.
  • Future Work:
    • Develop gradient-based optimization methods for rib design.
  • Implement Langevin and replica-exchange MCMC sampling for more refined uncertainty quantification.
  • Explore multi-fidelity DeepONets to leverage vast amounts of low-fidelity data alongside a few high-fidelity points, further reducing data requirements.
  • Develop a continual learning Bayesian multi-fidelity framework for ongoing optimization and uncertainty quantification.
  • Investigate the application of these methods to more complex geometries and conditions.

Summary updated

Abstract

This paper designs surrogate models with uncertainty quantification capabilities to improve the thermal performance of rib-turbulated internal cooling channels effectively. To construct the surrogate, we use the deep operator network (DeepONet) framework, a novel class of neural networks designed to approximate mappings between infinite-dimensional spaces using relatively small datasets. The proposed DeepONet takes an arbitrary continuous rib geometry with control points as input and outputs continuous detailed information about the distribution of pressure and heat transfer around the profiled ribs. The datasets needed to train and test the proposed DeepONet framework were obtained by simulating a 2D rib-roughened internal cooling channel. To accomplish this, we continuously modified the input rib geometry by adjusting the control points according to a simple random distribution with constraints, rather than following a predefined path or sampling method. The studied channel has a hydraulic diameter, Dh, of 66.7 mm, and a length-to-hydraulic diameter ratio, L/Dh, of 10. The ratio of rib center height to hydraulic diameter (e/Dh), which was not changed during the rib profile update, was maintained at a constant value of 0.048. The ribs were placed in the channel with a pitch-to-height ratio (P/e) of 10. In addition, we provide the proposed surrogates with effective uncertainty quantification capabilities. This is achieved by converting the DeepONet framework into a Bayesian DeepONet (B-DeepONet). B-DeepONet samples from the posterior distribution of DeepONet parameters using the novel framework of stochastic gradient replica-exchange MCMC.
BibTeX
@misc{lin2023deep,
  author = {Guang Lin and Izzet Sahin and Christian Moya and Amirhossein Mollaali and Guang Lin and Guillermo Paniagua},
  title = {{Deep Operator Learning-based Surrogate Models with Uncertainty Quantification for Optimizing Internal Cooling Channel Rib Profiles}},
  year = {2023},
  eprint = {2306.00810},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.00810},
}
Figure