Sensitivity analysis and stochastic simulations of non‐equilibrium plasma flow

Guang Lin, George Em KarniadakisInternational Journal for Numerical Methods in Engineering 2009 · DOI

TL;DR

This paper introduces a robust methodology for analyzing uncertainties in complex plasma flow simulations. By employing advanced sensitivity analysis techniques, they efficiently identify the most influential input parameters, enabling targeted stochastic simulations that significantly reduce computational cost while maintaining accuracy.

What Problem Does This Paper Address?

Simulating complex physical phenomena like plasma flows often involves numerous uncertain input parameters (e.g., material properties, boundary conditions, geometric details). Treating all these uncertainties simultaneously in stochastic simulations is computationally prohibitive. Traditional "One-At-a-Time" (OAT) sensitivity analyses can be inefficient for non-linear systems and may miss crucial interaction effects between parameters. This paper addresses the need for a more efficient and comprehensive approach to identify and rank these uncertain parameters, thereby guiding computationally expensive stochastic simulations.

Methods

The authors propose a two-stage approach:

  1. Stochastic Sensitivity Analysis: They compare several gradient-based sensitivity analysis methods to rank input parameters. These include:
  2. Morris Method: An OAT method that explores the parameter space more efficiently than basic OAT by using randomized trajectories.
  3. Multi-Element Probabilistic Collocation Method (ME-PCM) on Sparse Grids: A novel approach that combines domain decomposition (ME) with sparse grids for efficient collocation. This method is designed to handle discontinuities in the random space and offers fast convergence.
  4. Quasi-Monte Carlo (QMC) and Monte Carlo (MC) Methods: Standard sampling-based methods for uncertainty quantification.
  5. Targeted Stochastic Simulations: Once the most influential parameters are identified through sensitivity analysis, only these parameters are treated as stochastic variables (often using Karhunen-Loève expansion for random processes) in subsequent full-scale simulations. This significantly reduces the dimensionality of the stochastic problem.

The paper also details the implementation of the Probabilistic Collocation Method (PCM) and its extension, ME-PCM, including strategies for selecting collocation points using tensor products and sparse grids.

Datasets

The paper does not use traditional "datasets" in the machine learning sense. Instead, it presents two benchmark problems for evaluating their methodology:

  • One-dimensional two-fluid plasma flow: Modeled as a shock tube problem.
  • Two-dimensional single-fluid/two-temperature plasma flow: Modeled as flow past a cylinder.

These are standard test cases used in computational fluid dynamics and plasma physics to validate numerical methods.

Key Insights and Ideas

  • ME-PCM as a powerful tool: The Multi-Element Probabilistic Collocation Method on sparse grids emerges as a highly efficient technique for sensitivity analysis, demonstrating faster convergence than Morris, MC, and QMC methods, especially for problems with moderate numbers of random parameters.
  • Handling discontinuities: ME-PCM is particularly well-suited for problems with discontinuities in the random parameter space, a common issue in complex physical models.
  • Parameter ranking is crucial: The paper strongly advocates for using sensitivity analysis as a pre-screening step to identify dominant uncertain parameters. This dramatically reduces the computational burden of full stochastic simulations.
  • Non-linear interactions matter: The chosen sensitivity methods go beyond simple OAT by capturing non-linear interaction effects between input parameters, providing a more nuanced understanding of uncertainty propagation.
  • Two-fluid vs. MHD: The plasma flow examples highlight how two-fluid effects (like finite Larmor radius and Debye length) lead to deviations from standard Magnetohydrodynamics (MHD) behavior, underscoring the need for accurate modeling of these phenomena.

Key Results

  • Efficiency of ME-PCM: Figure 5 shows that ME-PCM achieves significantly faster convergence for both mean and standard deviation errors compared to Morris and MC methods when analyzing a test function.
  • Identification of key parameters: In the two-fluid plasma flow example, the Debye length ($\lambda_D$) was identified as the most important random factor influencing the flow. In the single-fluid/two-temperature flow, the ion Reynolds number ($Re_i$) and ion Prandtl number ($Pr_i$) were found to have the largest mean elementary effects on density and temperature, respectively.
  • Demonstration on plasma flows: The paper successfully applies the sensitivity analysis and subsequent stochastic simulations to two distinct plasma flow scenarios, showcasing the practical utility of the proposed methodology. The results illustrate how varying parameters like Larmor radius can transition the flow behavior from gas dynamics-like to MHD-like.

Limitations and Future Work

  • Computational cost for very large parameter spaces: While ME-PCM is efficient for moderate numbers of parameters (20-100), the authors suggest that for problems with very large numbers of random parameters (beyond 100), Morris and MC methods might become more efficient due to their sampling-based nature.
  • Generalization of ME-PCM: The paper focuses on one- and two-dimensional random spaces. Extending ME-PCM to higher-dimensional random spaces could be a direction for future work.
  • Experimental validation: The paper presents numerical simulations. Future work could involve validating these findings against experimental data.
  • New experimental work: The authors suggest that sensitivity analysis can guide future experimental designs.

Summary updated

Abstract

Abstract We study the parametric uncertainties involved in plasma flows and apply stochastic sensitivity analysis to rank the importance of all inputs to guide large‐scale stochastic simulations. Specifically, we employ different gradient‐based sensitivity methods, namely Morris, multi‐element probabilistic collocation method on sparse grids, Quasi‐Monte Carlo and Monte Carlo methods. These approaches go beyond the standard ‘One‐At‐a‐Time’ sensitivity analysis and provide a measure of the non‐linear interaction effects for the uncertain inputs. The objective is to perform systematic stochastic simulations of plasma flows treating only as stochastic processes the inputs with the highest sensitivity index, hence reducing substantially the computational cost. Two plasma flow examples are presented to demonstrate the capability and efficiency of the stochastic sensitivity analysis. The first one is a two‐fluid model in a shock tube whereas the second one is a one‐fluid/two‐temperature model in flow past a cylinder. Copyright © 2009 John Wiley & Sons, Ltd.
BibTeX
@article{lin2009sensitivity,
  author = {Guang Lin and George Em Karniadakis},
  title = {{Sensitivity analysis and stochastic simulations of non‐equilibrium plasma flow}},
  journal = {International Journal for Numerical Methods in Engineering},
  year = {2009},
  doi = {10.1002/nme.2582},
}
Figure