Abstract
This paper introduces SCAMR (Stochastic Collocation with Adaptive Mesh Refinement), a novel framework designed to quantify uncertainty in high-dimensional systems, particularly those exhibiting non-smooth behavior or discontinuities. By integrating adaptive mesh splitting with a hierarchical dimensionality reduction technique based on interaction checks, SCAMR effectively mitigates the "curse of dimensionality." Experimental results demonstrate that SCAMR achieves significantly higher accuracy with orders of magnitude fewer model evaluations than existing state-of-the-art methods, such as Adaptive Sparse Grid Collocation (ASGC), especially in problems involving hundreds of uncertain parameters.
Introduction
In modern engineering and physics, computer simulations are indispensable for predicting system behavior. However, the reliability of these simulations is often compromised by uncertainties in input parameters, such as material properties, geometric tolerances, or boundary conditions. Uncertainty Quantification (UQ) is the discipline dedicated to characterizing how these input uncertainties propagate to the final output.
Traditional UQ methods face two primary challenges:
- The Curse of Dimensionality: As the number of uncertain parameters increases, the computational cost required to map the parameter space grows exponentially.
- Discontinuities and Non-smoothness: Standard global polynomial expansion methods, while efficient for smooth functions, fail when the system exhibits "jumps" or sharp gradients, often resulting in Gibbs-like oscillations and poor convergence.
While methods like Adaptive Sparse Grid Collocation (ASGC) and Multi-Element Probabilistic Collocation (MEPCM) have attempted to address these issues, they often struggle to scale to high-dimensional spaces or become computationally prohibitive when variable interactions are complex.
The SCAMR Framework
SCAMR addresses these challenges through a three-pronged strategy involving flexible approximation, local refinement, and structural decomposition.
1. Generalized Polynomial Chaos (gPC) via Least Squares
Unlike traditional methods that rely on rigid interpolation or projection on fixed grids, SCAMR utilizes a least-squares approach to determine gPC coefficients. This allows for the use of unstructured point sets and provides the flexibility to reuse samples across different iterations or neighboring elements, significantly enhancing sampling efficiency.
2. Adaptive Mesh Refinement (AMR)
To handle discontinuities, SCAMR employs an adaptive refinement strategy. The algorithm monitors "abrupt variations" along the centerlines of the parameter domain. If the output deviates significantly from a second-order polynomial approximation, the domain is split into smaller sub-meshes. This "zoom-in" capability allows the method to resolve non-smooth regions with high precision without over-sampling smooth areas.
3. Two-Level Dimensionality Reduction
The core innovation of SCAMR is its use of High-Dimensional Model Representation (HDMR) to identify the underlying structure of the problem before refinement:
- Level 1: Identifies independent variables that do not interact with others, reducing the problem into several 1D sub-problems.
- Level 2: Detects pairwise interactions to decompose a massive $n$-dimensional problem into a series of smaller, manageable sub-problems.
Experimental Evaluation
The authors evaluated SCAMR against several synthetic benchmarks and a complex physics problem, ranging from 2D to 300D:
- Smooth and Non-smooth Functions: Tested on quadratic, sine-based, and discontinuous "step" functions across various dimensions.
- Stochastic Elliptic Problem: A classic UQ benchmark involving a partial differential equation (PDE) with a random diffusion coefficient, tested up to 75 dimensions.
- High-Dimensional Genz Function: A discontinuous function scaled to 300 dimensions to test the limits of the dimensionality reduction strategy.
Key Results
SCAMR consistently outperformed standard ASGC and MEPCM-A methods in both convergence speed and accuracy. A standout result occurred in the 300-dimensional discontinuous Genz function test: while the competing MEPCM-A method required approximately 123 million points to reach a relative error of 0.09, SCAMR achieved a superior error of 0.02 using only 22,000 points. This represents a reduction in computational cost of several orders of magnitude.
Discussion and Conclusion
The success of SCAMR highlights a critical insight in high-dimensional modeling: most complex systems possess "low effective dimensionality." Even when hundreds of variables are present, they rarely interact simultaneously. By identifying these interaction "islands," SCAMR avoids the exponential costs associated with full-space exploration. Furthermore, the use of centerline checking provides an efficient proxy for detecting non-smoothness without the need for exhaustive grid searches.
One identified limitation of the framework is Strong Coupling. If a system is fully coupled—where every input variable interacts non-linearly with every other variable—the dimensionality reduction step cannot decompose the problem, potentially leading to a mesh explosion. Future research may explore higher-order interactions or alternative basis functions to further extend the method's robustness in these worst-case scenarios.
Summary updated