Abstract
Evaluating complex engineering models governed by partial differential equations (PDEs) is often computationally prohibitive, particularly in high-dimensional design spaces. This paper introduces "embedded ridge approximations," a method that exploits the localized nature of scalar fields—such as pressure or strain—to perform efficient dimension reduction. By approximating global quantities of interest (QoIs) through the aggregation of localized nodal approximations, the proposed approach significantly reduces the number of required simulations. Furthermore, the authors present algorithms for the compression and recovery of these localized approximations, ensuring scalability for large-scale computational meshes.
Introduction
In fields such as aerospace engineering and uncertainty quantification, practitioners rely on high-fidelity simulations like Computational Fluid Dynamics (CFD) and Finite Element Analysis (FEA). However, the "curse of dimensionality" often renders these models impractical for optimization or sensitivity analysis, as the number of required evaluations grows exponentially with the number of input parameters.
Dimension reduction techniques, specifically active subspaces and ridge approximations, seek to identify a low-dimensional manifold within the input space that captures the majority of the variance in a global QoI (e.g., total lift). Traditionally, finding these ridge approximations requires a large number of global model evaluations, a challenge exacerbated when gradient information is unavailable. This paper addresses this bottleneck by shifting the focus from global outputs to the underlying local scalar fields that constitute them.
Methodology
Embedded Ridge Approximations
The core innovation of this work is a "divide and conquer" strategy for dimension reduction. Rather than attempting to find a single, complex ridge approximation for a global QoI directly, the authors propose computing localized approximations at individual nodes within the computational mesh.
- Nodal Approximation: For each node in the mesh, a low-dimensional (often 1D) ridge approximation is constructed using gradient-free methods such as Variable Projection (VP) or Minimum Average Variance Estimation (MAVE). This exploits the fact that local physical quantities are often sensitive to only a small subset of local geometric or environmental changes.
- Global Integration: These nodal approximations are mathematically integrated using a vector gradient covariance matrix. This "stitching" process allows the derivation of a subspace for the global QoI based on the collective behavior of the local fields.
Ridge Compression and Recovery
Physical fields in engineering simulations are typically continuous and spatially smooth, leading to significant redundancy in the ridge directions of neighboring nodes. To manage the storage requirements of nodal approximations in large meshes, the authors developed a compression framework:
- Greedy Compression: A greedy algorithm identifies and removes redundant nodes whose ridge directions can be accurately reconstructed as a linear combination of their neighbors.
- Subspace Distance Metric: The reconstruction quality is measured using a specific subspace distance metric, ensuring that the essential directional information is preserved even after significant data reduction.
Experimental Setup
The effectiveness of the embedded ridge approximation was evaluated on two distinct benchmarks:
- Analytical Benchmark: A synthetic 10-dimensional function composed of three distinct sub-functions. This setup allowed for a rigorous assessment of the probability of successful subspace recovery compared to direct methods.
- NACA0012 Airfoil: A standard aerodynamic benchmark where the airfoil geometry is parameterized by 50 Hicks-Henne bump functions ($d=50$). The surface pressure field was evaluated at 5,233 nodes using the SU2 compressible Euler flow solver. The primary QoIs were the coefficients of lift ($C_l$) and drag ($C_d$).
Key Results
Sample Efficiency and Accuracy
The embedded approach demonstrated superior sample efficiency compared to direct ridge approximations. In the analytical test, the embedded method achieved nearly 100% subspace recovery with approximately 400 samples, whereas the direct approach often failed to converge even with 800 samples. For the NACA0012 airfoil, surrogate models built using the embedded method reached Mean Squared Error (MSE) convergence with only 300–400 CFD simulations, significantly outperforming the baseline.
Data Compression
The compression algorithm proved highly effective for the airfoil pressure field. The authors successfully compressed the data by nearly 90%, removing approximately 4,700 of the 5,233 nodes while maintaining high reconstruction accuracy. This greedy approach outperformed standard $k$-medoids clustering, demonstrating the value of leveraging spatial smoothness in dimension reduction data.
Discussion and Conclusion
The success of embedded ridge approximations stems from two primary insights: the localization of physical sensitivities and the spatial smoothness of physical fields. By breaking a global problem into simpler local optimization tasks, the method avoids the difficult optimization landscapes associated with high-dimensional global ridge recovery.
While the current greedy compression algorithms do not guarantee global optimality, they provide a practical and scalable solution for high-fidelity engineering workflows. Future research may explore the application of these techniques to chaotic or globally coupled systems where localization is less pronounced. Ultimately, this work provides a robust framework for performing sensitivity analysis and optimization on complex PDE-based systems with a fraction of the traditional computational cost.
Summary updated