Sampling-efficient surrogate modeling for sensitivity analysis of brake squeal using polynomial chaos expansion

Abstract

This study explores the feasibility of using sampling-efficient surrogate modeling methods to emulate the behavior of complex Finite Element (FE) models of brake squeal in Global Sensitivity Analysis (GSA). FE-based GSA workflows are computationally expensive due to multiple solver runs under uncertain input parameters. To address this bottleneck, we investigate three Polynomial Chaos Expansion (PCE) approaches: (1) projection-based PCE with sparse grids, (2) regression-based PCE with different oversampling rates and polynomial orders, and (3) regression-based PCE with sequential sampling. These methods are applied to a nine-parameter problem, including material and operational parameters. The models are validated against 300 unseen test samples. The accuracy of GSA estimates is validated both qualitatively, based on the understanding of problem mechanics, and quantitatively, through direct estimation of GSA indices using 1,000 reference sample points and comparison with the direct and Kriging models of the commercial software LS-OPT. Results demonstrate that the proposed methods significantly accelerate uncertainty propagation and GSA estimation for time-intensive brake squeal simulations while minimizing computational cost.