Noise is a major threat to physical and mental health as well as human well-being. Thus, reducing noise emissions is an important goal in many engineering disciplines, e.g. in the automotive industry or aviation. To enable quiet product designs, engineers need to simulate the acoustic properties of such systems in the frequency domain to consider acoustic measures already in early design phases. A classical simulation approach is the Finite Element Method (FEM), which is a wave resolving discretization technique to solve the governing differential equations. The main challenge to use the FEM for vibroacoustic design is its computational burden. To overcome this burden we propose deep learning surrogate models and investigate their application in a design optimization context. We consider a common benchmark in vibroacoustics, a rectangular plate subject to harmonic excitation, as design benchmark. The design space consists of beading patterns, which are local geometry deformations and thus mass-free measures to change the local stiffness of the dynamical system. The design task is to find an optimal beading pattern to reduce the velocity sum-level in a target frequency range. This design task has high relevance in engineering applications, since it is desired to design structures with low frequency response in critical frequency ranges, defined e.g. by the excitation frequency of an engine. We approach this design task using guided diffusion [1], which makes use of the gradients from a deep learning (DL) forward model and a diffusion model to create new design proposals. Thus, in the first part of this talk, we will present a novel DL surrogate model to emulate the frequency response of the rectangular plate model [2]. This data-driven surrogate is able to process beading patterns as well as different boundary conditions, material properties and load cases. We benchmark our surrogate model on several problem specific metrics and show that it outperforms baseline DL surrogates, like DeepONets or Fourier Neural Operators. The surrogate model is then used within a guided diffusion approach for inverse design. We show that the approach is able to find optimal designs, which are superior to any design present in the data set used to train the surrogate model. Furthermore, the approach can handle different target frequency ranges, as well as explore new beading patterns, which have not been part of the training distribution.