Uncertainty quantification for Multiphase-CFD simulations of bubbly flows: a machine learning-based Bayesian approach supported by high-resolution experiments

Yang Liu 0201, Dewei Wang, Xiaodong Sun, Yang Liu 0202, Nam Dinh, Rui HuReliab. Eng. Syst. Saf 2021 · DOI

TL;DR

This paper introduces a novel machine learning-based Bayesian approach to quantify and reduce uncertainties in multiphase Computational Fluid Dynamics (MCFD) simulations of bubbly flows. By integrating high-resolution experimental data with advanced machine learning techniques like neural networks and Gaussian processes, the authors demonstrate a significant reduction in simulation uncertainties, leading to more reliable predictions for critical engineering applications.

What Problem Does This Paper Address?

Accurately simulating multiphase flows, particularly bubbly flows, is crucial for many engineering applications, such as nuclear power plants. However, these simulations are plagued by significant uncertainties stemming from empirical constitutive relations and the inherent stochastic nature of the flow. Traditional methods struggle to capture these uncertainties comprehensively, leading to unreliable predictions. Existing experimental datasets often lack the necessary high-resolution data or comprehensive measurements required for robust uncertainty quantification (UQ). This paper aims to bridge this gap by developing a data-driven UQ framework that leverages high-fidelity experimental data to constrain and improve MCFD simulations.

Methods

The core of this work is a modular Bayesian inference approach that breaks down the UQ process into manageable modules.

  • Surrogate Modeling: To overcome the computational expense of running thousands of MCFD simulations, a surrogate model is developed. This involves:
  • Dimensionality Reduction (PCA): Principal Component Analysis (PCA) is used to reduce the high dimensionality of MCFD simulation outputs (e.g., void fraction, velocities at many grid points) to a smaller set of principal components (PCs).
  • Feedforward Neural Network (FNN): An FNN is trained to predict these reduced PC subspace outputs based on the input constitutive relation parameters. This integrated PCA-FNN approach aims to minimize both FNN prediction error and PCA-induced error.
  • Model Form Uncertainty (Gaussian Processes - GPs): Gaussian Processes are employed to model the discrepancy between the MCFD model and experimental data, representing the "model form uncertainty." This is trained using experimental measurements and the baseline MCFD simulation.
  • Bayesian Inference (MCMC): Markov Chain Monte Carlo (MCMC) methods, specifically the No-U-Turn Sampler (NUTS), are used to infer the posterior distributions of the constitutive relation parameters and to quantify model form uncertainty.
  • Probability-Boxes (p-boxes): The quantified epistemic (parameter and model form) and aleatoric (stochastic flow fluctuations) uncertainties are combined to construct p-boxes, which provide a comprehensive representation of prediction uncertainty.

Datasets

The study utilizes a comprehensive two-phase bubbly flow experimental database obtained from a facility at Virginia Tech. This database includes high-resolution measurements of:

  • Void fraction
  • Gas phase velocity
  • Liquid phase velocity
  • Bubble size (used as an inlet condition)

This dataset was collected under two distinct inlet flow conditions (Condition 1 and Condition 2) and is noted to overcome shortcomings of legacy datasets, such as providing all necessary quantities of interest and high-resolution data.

Key Insights and Ideas

  • Integrated Dimensionality Reduction and Surrogate Modeling: Combining PCA with FNN during the training process, rather than separately, is a key novelty. This integrated approach helps to reduce both the surrogate model's inherent error and the error introduced by dimensionality reduction itself.
  • Quantifying Both Epistemic and Aleatoric Uncertainty: The framework explicitly addresses both reducible (epistemic: parameter and model form) and irreducible (aleatoric: stochastic fluctuations) uncertainties, providing a more complete picture of prediction uncertainty.
  • P-boxes for Comprehensive Risk Analysis: The use of p-boxes is highlighted as a powerful tool for representing imprecise probabilities, allowing for more robust risk assessment, reliability analysis, and sensitivity analysis.
  • Experimental Data is Crucial for Inverse UQ: The paper strongly emphasizes that robust UQ, especially inverse UQ (where experimental data is used to infer model parameters/uncertainties), is critically dependent on high-quality, comprehensive experimental measurements.

Key Results

  • Significant Uncertainty Reduction: The proposed approach successfully quantified and reduced the uncertainties in MCFD predictions for bubbly flows.
  • Improved Predictions:
    • Considering only model parameter uncertainty significantly reduced prediction uncertainty and eliminated unphysical predictions (e.g., void fraction concentration in specific regions).
  • Further improvement was achieved by incorporating model form uncertainty, leading to better agreement with experimental measurements, particularly for void fraction.
  • Surrogate Model Performance: The developed FNN baseline model achieved low RMSE (0.0325) and high R² (0.996) on the testing dataset, indicating good predictive capability.
  • Model Form Uncertainty (δ(x)) Behavior: The trained GPs for model form uncertainty showed smooth patterns without unphysical oscillations, and their uncertainty bands were consistent with measurement uncertainties.
  • Discrepancies Remain: Despite the UQ efforts, some discrepancies between MCFD predictions and experimental measurements persist, particularly in void fraction at near-wall regions and gas velocity at certain locations.

Limitations and Future Work

  • Model Form Uncertainty Dependency: The current model form uncertainty term, δ(x), is assumed to be solely dependent on spatial location and independent of inlet flow conditions. The authors acknowledge that in reality, δ(x) should also depend on inlet conditions (void fraction, velocities, bubble size).
  • Need for Broader Experimental Data: To address the limitation of the model form uncertainty, the authors suggest that additional experiments covering a wider range of inlet conditions are needed. This would allow for training a more accurate model form uncertainty that incorporates inlet conditions.
  • Further Investigation of Discrepancies: The remaining discrepancies, especially in void fraction and gas velocity, warrant further investigation, potentially through more refined model form uncertainty models.
  • Computational Efficiency: The use of PyTorch and GPU acceleration ensures efficient computation, but the initial setup and training of surrogate models and GPs can still be computationally intensive.

Summary updated

BibTeX
@article{liu2021uncertainty,
  author = {Yang Liu and Dewei Wang and Xiaodong Sun and Yang Liu and Nam Dinh and Rui Hu},
  title = {{Uncertainty quantification for Multiphase-CFD simulations of bubbly flows: a machine learning-based Bayesian approach supported by high-resolution experiments}},
  journal = {Reliab. Eng. Syst. Saf.},
  year = {2021},
  doi = {10.1016/j.ress.2021.107636},
}
Figure