Work

A Non‐Intrusive Efficient Two‐Step Model Reduction Method for Large‐Scale Thin‐Walled Acoustic Indirect BEM Systems

Year: 2025

Type: article

Abstract: ABSTRACT Performing frequency sweep analysis for acoustic performance prediction using the boundary element method (BEM) is associated with high computational costs as the associated system matrices are dense and frequency‐dependent. While model order reduction (MOR) techniques, such as Krylov subspace recycling, have been proposed for direct BEM, their application to indirect Galerkin BEM (iBEM) remains limited, as only moderate reductions are offered. In that context, this work presents a two‐step, non‐intrusive MOR method for large‐scale, thin‐walled acoustic problems modeled via iBEM, reducing both respective solution times and required memory while facilitating their employment on top of any existing commercial software. Within the first step, the proposed method employs the automatic Krylov subspace recycling (AKR) algorithm in combination with a Chebyshev approximation of the iBEM system to construct an initial reduction basis by minimizing the number of systems sampled and permitting the storage of the affinely approximated system. In the second step, the Reduced Basis Method (RBM) is employed to further reduce the dimension of the system, achieving significant savings in memory and online computational time. To improve scalability under tight memory constraints, a memory‐aware version is introduced, using frequency interval splitting with a Petrov–Galerkin projection in the second step. This version bounds memory consumption without compromising accuracy. The proposed method is implemented in MATLAB and interfaces with an external commercial BEM solver in a black‐box fashion, demonstrating compatibility with existing industrial workflows. The performance of the methodology is assessed on two industrially challenging benchmark problems. Results show that the method achieves substantial online speed‐ups compared to a single‐step reduction, while maintaining solution accuracy.

Source: International Journal for Numerical Methods in Engineering

DOI: 10.1002/nme.70171

Authors: Philip Le, Dionysios Panagiotopoulos, Elke Deckers

Institution: KU Leuven

Language: English

Topic: Model Reduction and Neural Networks

Subfield: Statistical and Nonlinear Physics

Field: Physics and Astronomy

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $4,250

Funders: European Commission, KU Leuven, HORIZON EUROPE Framework Programme