Speaker
Description
The design of vibroacoustic systems commonly requires repetitive high-fidelity finite element evaluations across wide frequency ranges and for numerous excitation scenarios, at significant computational cost. Conventional model order reduction (MOR) techniques, which are often employed to alleviate this cost, apart from inherently trading off solution accuracy for computational gain, typically face notable limitations in multiple-input multiple-output (MIMO) settings: moment-matching approaches scale poorly with the number of inputs, while modal methods struggle with large scale systems. To overcome these challenges, this work proposes an accelerated deflation-based solution strategy for large-scale vibroacoustic finite element MIMO systems. Rather than using MOR to approximate the system response, the proposed approach leverages it to construct a deflation preconditioner that significantly accelerates GMRES convergence when a frequency sweep needs to be solved for multiple inputs. The reduction basis is generated using an Automatic Krylov Subspace Recycling (AKR) algorithm, which collects relevant subspaces based on a maximum iteration criterion. To handle the high condition numbers typical of vibroacoustic finite element models, two preconditioning strategies are incorporated: scaling of the vibration, acoustic, and interface sub-matrices by their Frobenius norms, and LU-based preconditioning at the mid-frequency of a given input. Together, these measures enable practical deflation with moderate subspace sizes. Both theoretical cost analysis and numerical results on a coupled plate–cavity benchmark demonstrate substantial reductions in iteration counts compared to conventional iterative strategies. As with traditional MOR approaches, the dominant cost is transferred to an offline phase; however, the resulting online MIMO frequency sweep is not only highly efficient but also free of any compromise in solution accuracy.