A bi-fidelity method for the multiscale Boltzmann equation with random parameters

Liu Liu, Xueyu ZhuJournal of Computational Physics 2020 · DOI

TL;DR

This paper introduces a novel "bi-fidelity" computational method to efficiently tackle the multiscale Boltzmann equation, a complex model for gas dynamics, when parameters are uncertain. By cleverly combining a fast, approximate "low-fidelity" model (Euler equations) with a few accurate, but slow, "high-fidelity" simulations (Boltzmann equation), the method achieves significant computational speedups while maintaining high accuracy in predicting macroscopic gas properties across a range of uncertainties.

What Problem Does This Paper Address?

Simulating the behavior of gases at a fundamental level often involves the Boltzmann equation. This equation is incredibly powerful but also computationally demanding, especially when dealing with multiple scales (e.g., very fast particle interactions and slower overall fluid motion) and uncertainties in parameters like initial conditions or collision properties. These uncertainties are crucial in real-world applications, but traditional methods that explore these uncertainties require an enormous number of computationally expensive Boltzmann simulations, often making them infeasible. The challenge is to find a way to accurately capture the impact of these uncertainties without incurring prohibitive computational costs.

Methods

The core of this work is a bi-fidelity stochastic collocation (SC) method. Here's how it breaks down:

  • High-Fidelity Model: The accurate, but slow, multiscale Boltzmann equation is solved using an asymptotic-preserving (AP) solver. AP schemes are designed to remain stable and accurate even when the physical scales become very different (e.g., when the gas behaves more like a fluid).
  • Low-Fidelity Model: A significantly faster, approximate model is used. In this paper, the authors choose the compressible Euler equations, which are a first-order approximation of the Boltzmann equation in the fluid limit. These equations capture macroscopic quantities (density, velocity, temperature) directly, bypassing the need to resolve individual particle distributions.
  • Bi-Fidelity Algorithm:
    • Offline Stage: The algorithm first uses the cheap low-fidelity model to explore the parameter space and identify the most "important" parameter points that significantly influence the solution. This is done using a greedy algorithm to select a small set of representative parameter values.
  • Online Stage: For a given parameter value, the low-fidelity model is run to obtain a solution. This solution is then used to construct a bi-fidelity approximation of the high-fidelity solution. Essentially, the low-fidelity solution's behavior at the selected important points is used to "learn" how the high-fidelity solution behaves across the entire parameter space. This involves projecting the low-fidelity solution onto a subspace spanned by the high-fidelity solutions at the important points.

Datasets

The paper does not introduce new datasets. Instead, it uses synthetic initial data and collision kernel models to represent various scenarios, including:

  • A "double-peak" initial data test to mimic non-equilibrium states.
  • A "Sod shock tube" test with discontinuous initial data.
  • A "mixed regime" test where the Knudsen number (a measure of gas rarefaction) varies spatially.

These scenarios are designed to test the method's performance across different physical regimes and complexities.

Key Insights and Ideas

  • Leveraging Macroscopic Approximations: The key insight is that even if a low-fidelity model (like Euler equations) is inaccurate in predicting the detailed particle distribution in the physical space, its macroscopic quantities can still accurately capture the variations of the high-fidelity solution across the random parameter space. This is crucial for uncertainty quantification.
  • Computational Efficiency through "Learning": The bi-fidelity approach effectively "learns" the expensive high-fidelity solution's behavior from a small number of high-fidelity runs, guided by the much cheaper low-fidelity model. This dramatically reduces the overall computational cost.
  • Robustness Across Regimes: The method shows promise in handling both fluid-like (small Knudsen number) and kinetic (larger Knudsen number) regimes, demonstrating its versatility.

Key Results

  • Significant Speedup: The bi-fidelity method achieves substantial computational speedups. For example, in one test case, the high-fidelity Boltzmann solver was about 43 times slower than the low-fidelity Euler solver for a single run, yet the bi-fidelity approach using only 10 high-fidelity runs achieved an accuracy level of O(10⁻³) in a 15-dimensional random space, far surpassing the low-fidelity model's accuracy.
  • High Accuracy: The bi-fidelity approximations closely matched the high-fidelity Boltzmann solutions, especially in characterizing the macroscopic quantities (density, velocity, temperature) across the random parameter space.
  • Error Reduction: The mean L² errors between the bi-fidelity and high-fidelity solutions decreased rapidly with an increasing number of high-fidelity runs, quickly saturating at a low level. The errors from the low-fidelity model alone were significantly larger.

Limitations and Future Work

  • Sharpness of Error Estimates: The authors acknowledge that their error estimates might not be sharp and that a deeper understanding of the Knudsen number's role is needed for more precise accuracy analysis.
  • Rigorous Stability Analysis: Proving the rigorous stability of the AP scheme used for the high-fidelity solver, particularly due to the penalization technique, remains a challenging open problem.
  • Alternative Low-Fidelity Models: The paper suggests that other low-fidelity models, such as the compressible Navier-Stokes equations, could potentially lead to even more accurate bi-fidelity approximations.
  • Future Research: The authors plan to report more results on stability and error analysis in future work.

Summary updated

BibTeX
@article{liu2020bi,
  author = {Liu Liu and Xueyu Zhu},
  title = {{A bi-fidelity method for the multiscale Boltzmann equation with random parameters}},
  journal = {Journal of Computational Physics},
  year = {2020},
  doi = {10.1016/j.jcp.2019.108914},
}
Figure