TL;DR
This paper introduces a novel hybrid method combining Polynomial Chaos (PC) and Dynamically-Orthogonal (DO) schemes to efficiently solve stochastic partial differential equations (SPDEs). This approach addresses a singularity issue in the DO method when dealing with deterministic initial conditions, offering comparable accuracy to existing methods but with significantly improved computational efficiency, especially in high-dimensional problems.
What Problem Does This Paper Address?
Many real-world physical systems, such as those in fluid mechanics, solid mechanics, and wave propagation, are subject to uncertainties in their parameters, initial conditions, or external forces. These uncertainties are often modeled using stochastic partial differential equations (SPDEs). Accurately quantifying these uncertainties and understanding their impact on system behavior is crucial. While methods like Monte Carlo (MC) and Polynomial Chaos (PC) are widely used, they can become computationally expensive, particularly for high-dimensional problems or when dealing with specific challenges like deterministic initial conditions in the Dynamically-Orthogonal (DO) method. The DO method, while powerful for capturing complex dynamics, suffers from a singularity when initial conditions are deterministic, making it difficult to initialize and leading to numerical issues. This paper aims to bridge this gap by proposing a hybrid approach that leverages the strengths of both PC and DO methods.
Methods
The core of this paper is the development and evaluation of a hybrid Polynomial Chaos (PC) and Dynamically-Orthogonal (DO) method.
- Dynamically-Orthogonal (DO) Method: This method represents the solution of an SPDE as a time-dependent expansion involving both stochastic coefficients and basis functions. A key feature is the "dynamically orthogonal" condition, which helps to overcome redundancy in the representation and leads to a set of deterministic evolution equations. However, when initial conditions are deterministic, the DO method faces a singularity.
- Polynomial Chaos (PC) Method: This is a well-established spectral method that represents stochastic processes using a series of orthogonal polynomials. The paper specifically mentions the use of Probabilistic Collocation Method (PCM) and Multi-Element PCM (ME-PCM) as computationally efficient variants of PC.
- Hybrid DO-PC Method: This novel approach combines the strengths of both methods. Initially, for deterministic initial conditions, the SPDE is solved using PC (PCM or ME-PCM) for a short period until the stochasticity of the solution develops sufficiently. At a chosen "switching time" ($t_s$), the solution is then projected onto a subspace determined by a Karhunen-Loeve (KL) decomposition. This KL decomposition provides the necessary initial conditions for the DO method, effectively circumventing the singularity issue. The system then switches to the DO method for the remainder of the simulation.
Datasets
The paper does not introduce new datasets. Instead, it uses well-defined analytical test cases for SPDEs, including:
- A stochastic advection equation with both time-independent and time-dependent random advection velocities.
- A Burgers equation with both exact DO components and random forcing.
- A nonlinear diffusion equation in one spatial dimension but multi-dimensional in the parametric space.
These test cases allow for direct comparison with exact solutions or established benchmarks, facilitating a rigorous evaluation of the proposed method's accuracy and efficiency.
Key Insights and Ideas
- Bridging the Deterministic-Stochastic Divide: The hybrid DO-PC method effectively resolves the singularity issue in the DO method when faced with deterministic initial conditions, a common scenario in many physical problems.
- "On-the-Fly" Basis Adaptation: The DO method, even when initiated with arbitrary orthogonal polynomials, demonstrates the ability to adapt its basis functions to accurately track the system's dynamics, eventually converging to an optimal subspace.
- Efficiency Gains in High Dimensions: The hybrid DO-PC method shows significant computational speedups compared to PCM, especially as the dimensionality of the parametric space increases. This is a critical advantage for complex, real-world problems with multiple sources of uncertainty.
- Capturing Complex Stochastic Behavior: The DO method, and by extension the hybrid approach, is shown to effectively capture both Gaussian and non-Gaussian probability density functions (PDFs) of the solution, including bimodal distributions.
Key Results
- Accuracy: The hybrid DO-PC method achieves accuracy comparable to established methods like PCM for the tested advection, Burgers, and nonlinear diffusion equations. For instance, in the advection equation, the errors for the mean and variance using DO were found to be identical to those from PCM.
- Efficiency: The DO method is significantly faster than PCM, particularly in high-dimensional parametric spaces. Figure 8 illustrates that for the advection problem, DO is orders of magnitude faster than PCM as the dimensionality increases. Figure 13 shows similar speedups for the Burgers equation.
- Convergence: The paper demonstrates exponential convergence of the hybrid method with respect to the number of DO modes used, indicating its efficiency in capturing the stochastic behavior.
- Robustness: The method successfully handles problems with deterministic initial conditions, linear and nonlinear SPDEs, and various types of random excitations (e.g., Gaussian, bimodal).
Limitations and Future Work
- Sensitivity to Parameters: The hybrid DO-PC method's performance is sensitive to the choice of the switching time ($t_s$) and the number of DO modes ($N$). The paper suggests that choosing the number of modes based on the eigenvalues of the KL decomposition is a reasonable approach, but more systematic studies on adaptive strategies for selecting these parameters are needed.
- Stiff Systems: For certain problems, the evolution equations for the stochastic coefficients can become stiff, potentially requiring stiffly stable numerical schemes for overcoming saturation.
- Adaptive Mode Selection: The paper highlights the need for adaptive strategies to dynamically add or remove modes during the simulation to efficiently capture the energy of high modes that might be neglected.
- Bi-orthogonality: Future work could explore reformulating the DO equations by imposing bi-orthogonality on the basis and stochastic coefficients, which might lead to a modified system of evolution equations.
- Multiscale Approaches: Incorporating "sub-grid-like" terms to account for the action of truncated modes is suggested as a future direction.
Summary updated