Article
Deep Learning based Surrogate Modeling for Frequency Response Prediction of Structural Dynamic Models (en)
* Presenting author
Abstract:
Understanding the response of a dynamical system in the frequency domain is crucial for vibroacoustic design. Such frequency response functions are usually computed using discretization techniques, e.g. the finite element method (FEM). The FEM requires solving a potentially large linear system for each frequency step. In particular, for many-query tasks, like design space exploration, optimization or uncertainty quantification the associated computational burden gets infeasible. In this work, we investigate data-driven deep learning surrogate models to replace these computational expensive FEM computations. Data-driven surrogate models rely on the offline – online paradigm, shifting the computational cost to the offline training phase, while providing a fast-to-evaluate surrogate model in the online phase. We consider a structural dynamic plate model as a benchmark for surrogate modeling. This model allows defining complex input design spaces, such as beading pattern variations or material and geometry property variations. As target quantities, we consider high dimensional quantities, like the velocity distributions over the plate. We present deep learning surrogate models, which are inspired by neural operators, enabling to predict the field response for any given query frequency and a point in the design space. The proposed network architectures are investigated in terms of different vibroacoustic error metrics.