Novel ensemble algorithms for random two-domain parabolic problems

Yizhong Sun, Jiangshan Wang, Haibiao ZhengarXiv 2023 · DOI · arXiv

Abstract

Modeling complex physical systems, such as atmosphere-ocean interactions, often involves significant uncertainty in parameters like friction and diffusion coefficients. Traditional Monte Carlo methods for these coupled parabolic problems are computationally expensive due to the need for repeated matrix factorizations. This paper proposes three novel ensemble algorithms that utilize a shared coefficient matrix for multiple realizations at each time step. These methods achieve over 90% reduction in computational costs while maintaining unconditional stability and optimal convergence rates.

Introduction

In mathematical physics, fluid-fluid interactions are frequently modeled using coupled parabolic equations. However, real-world parameters—such as the friction parameter ($\kappa$) and diffusion coefficients ($\nu_1, \nu_2$)—are often random variables estimated from noisy data. Standard Monte Carlo (MC) approaches require solving a unique linear system for every sample at every time step, leading to prohibitive computational costs as the number of samples ($J$) increases.

Problem Formulation

The study focuses on two-domain parabolic problems where two distinct regions are coupled across an interface. This setup is typical for modeling interactions between different physical media, such as air and water.

Geometric configuration of the two-domain problem with interface $\Gamma$.

As shown in the figure above, the domain is partitioned into $\Omega_1$ and $\Omega_2$ by an interface $\Gamma$. The normal vectors $\hat{n}_1$ and $\hat{n}_2$ define the orientation at the boundary where the two systems interact. The randomness is introduced through the physical coefficients governing the heat or fluid flow within these domains and across the interface.

Proposed Ensemble Algorithms

The core innovation lies in the Ensemble Idea, which decouples randomness from the system's "stiffness" matrix. The authors introduce three specific algorithms:

  1. Algorithm A1: Optimized for cases where only the friction parameter $\kappa$ is random. It remains stable even with large variations in $\kappa$.
  2. Algorithm A2: Handles randomness in both friction and space-time dependent diffusion coefficients ($\nu$). It uses a common coefficient matrix based on the average of random parameters for all samples.
  3. Algorithm A3: An extension of A2 that averages diffusion coefficients over time, allowing the same matrix factorization to be reused across multiple time steps, further enhancing efficiency.

By using a shared matrix for $J$ different realizations, the computational overhead is reduced to a single matrix factorization per time step (or across multiple steps in A3), followed by $J$ back-substitutions.

Numerical Results and Discussion

The algorithms were validated using synthetic benchmarks and a practical 3D fuel cell model. Key findings include:

  • Computational Efficiency: In tests with 400 realizations, Algorithm A2 and A3 reduced computation time by 91.8% and 92.1%, respectively, compared to standard partitioned methods.
  • Stability and Convergence: The methods are proved to be unconditionally stable, allowing for larger time steps without numerical divergence. They also achieve optimal convergence rates in $L^2$ and $H^1$ norms.
  • Physical Fidelity: The 3D fuel cell simulation successfully captured the "temperature jump" at the interface of steel and titanium components, demonstrating the algorithms' applicability to complex multi-material geometries.

Conclusion

The proposed ensemble algorithms provide a powerful "shortcut" for uncertainty quantification in coupled parabolic systems. By significantly reducing the cost of matrix operations, they enable more extensive simulations of random fluid-fluid interactions. Future work will focus on extending these techniques to non-linear models, such as the Navier-Stokes equations, to better represent chaotic real-world dynamics.

Summary updated

BibTeX
@misc{sun2023novel,
  author = {Yizhong Sun and Jiangshan Wang and Haibiao Zheng},
  title = {{Novel ensemble algorithms for random two-domain parabolic problems}},
  year = {2023},
  eprint = {2301.09271},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.09271},
}
Figure