Efficient numerical methods for the uncertain Boltzmann equation based on a hybrid solver

Yiwen Lin, Liu LiuarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a powerful framework for solving the Boltzmann equation—a fundamental model for gas dynamics—when parameters like initial conditions or collision behaviors are uncertain. By combining a "hybrid" solver (which intelligently switches between expensive kinetic models and cheap fluid models) with advanced sampling techniques like Multi-Level Monte Carlo (MLMC) and Multi-Fidelity methods, the authors achieve significant computational speedups while maintaining high accuracy across different physical regimes.

What Problem Does This Paper Address?

The Boltzmann equation is the gold standard for describing how gas particles interact, bridging the gap between microscopic collisions and macroscopic fluid flow. However, it suffers from two major "curses":

  1. The Curse of Dimensionality: It operates in a high-dimensional phase space (position and velocity), making it incredibly expensive to solve numerically.
  2. The Multiscale Challenge: In regions where gas is dense (low Knudsen number, $Kn$), it behaves like a fluid; in sparse regions (high $Kn$), it behaves like a collection of particles. Standard solvers often struggle to handle both efficiently.

When you add Uncertainty Quantification (UQ)—accounting for the fact that we rarely know initial temperatures or collision cross-sections perfectly—the problem becomes nearly intractable. Standard Monte Carlo methods require thousands of expensive simulations to produce a reliable statistical average. This paper seeks a way to perform UQ without breaking the bank.

Methods

The authors propose a "Hybrid Solver" approach integrated with two non-intrusive UQ strategies:

  • The Asymptotic-Preserving-Hybrid (APH) Scheme: Instead of solving the full Boltzmann equation everywhere, the algorithm dynamically partitions the domain. It uses a cheap Euler (fluid) solver in equilibrium regions and an AP kinetic solver only where necessary (e.g., near shocks or in rarefied zones). It uses specific mathematical criteria (based on the Navier-Stokes dissipative terms and distance to equilibrium) to decide when to switch models.
  • Multi-Level Monte Carlo (MLMC): This method reduces variance by using a hierarchy of spatial/velocity meshes. Most samples are run on very coarse, "cheap" meshes, while only a few are run on the finest, most expensive meshes to correct the bias.
  • Multi-Fidelity Methods (Bi-fidelity and Tri-fidelity): These leverage the relationship between different models. A low-fidelity model (the Euler equations) is used to explore the parameter space cheaply, and a small number of high-fidelity runs (the APH or full Boltzmann solver) are used to "anchor" the results and ensure accuracy.

Datasets

The paper evaluates the methods using classic benchmarks in computational fluid dynamics, adapted for kinetic theory and uncertainty:

  • Sod Shock Tube Test: A standard benchmark for capturing discontinuities (shocks and rarefaction waves).
  • Blast Wave Test: A more complex interaction of multiple waves.
  • Mixed-Regime Cases: Specially designed problems where the Knudsen number varies smoothly across the domain, forcing the hybrid solver to adapt its strategy in real-time.

Key Insights and Ideas

  • Dynamic Domain Decomposition is Key: The APH solver doesn't just save time; it preserves the "Asymptotic Preserving" property, meaning it automatically converges to the correct fluid limit as the gas becomes denser, without needing to resolve tiny mean-free-path scales.
  • Non-Intrusive UQ: Because the MLMC and Multi-fidelity approaches are "non-intrusive," they treat the underlying solver as a black box. This makes the framework easily extensible to other complex kinetic equations.
  • Localized Uncertainty: A fascinating insight from the "uncertain collision kernel" test is that uncertainty in particle collisions mainly affects the kinetic regions. The APH-based MLMC method is smart enough to keep this uncertainty localized, preventing it from unnecessarily complicating the deterministic fluid zones.

Key Results

  • Efficiency Gains: In deterministic tests, the APH hybrid solver was 1.97x faster than the full kinetic model for the Sod shock tube and 1.51x faster for the blast wave, with negligible loss in accuracy.
  • Superior UQ Performance: The three-level MLMC scheme outperformed standard Monte Carlo methods, achieving an error level of $10^{-3}$ with significantly fewer high-resolution samples.
  • Multi-Fidelity Synergy: The bi-fidelity approach proved exceptionally efficient when "offline" computation is permissible, as it allows for near-instantaneous "online" predictions for new uncertain parameters once the initial model relationship is established.

Limitations and Future Work

  • Parameter Sensitivity: The hybrid solver relies on two threshold parameters ($\eta_0$ and $\delta_0$) to decide when to switch between fluid and kinetic models. While the authors provide values that worked well, a more rigorous way to choose these automatically for any given problem remains an open question.
  • Theoretical Analysis: While the numerical results are robust, a formal "rigorous error analysis" for the combination of APH and MLMC is still needed to provide mathematical guarantees on convergence rates.
  • Higher Dimensions: The current work focuses on 1D spatial domains. Scaling this to 2D or 3D spatial problems will introduce further complexity in managing the moving boundaries between fluid and kinetic zones.

Summary updated

BibTeX
@misc{lin2025efficient,
  author = {Yiwen Lin and Liu Liu},
  title = {{Efficient numerical methods for the uncertain Boltzmann equation based on a hybrid solver}},
  year = {2025},
  eprint = {2507.20316},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.20316},
}
Figure