Design Optimization for Acoustics
On a daily basis, we are confronted with noise. Whether during a train ride or in a bustling room, noise can be a source of stress. The field of acoustics in engineering holds the potential to mitigate this issue by optimizing object, room or vehicle design to minimize noise levels. The prevalent method for predicting noise levels, for example in a train, is the finite-element method. While this method can give accurate predictions for specific designs, it is computationally expensive and requires manual pre-processing. The inverse application - designing to fit specific vibroacoustic needs – is not directly possible and requires iterative adjustments of the design.
We aim to leverage deep learning methods to address these limitations and make design for acoustic needs possible. In a first step, we constructed a dataset of plates with different beading patterns and computed their vibroacoustic response to harmonic forces. The beading patterns stiffen the plates and alter these responses. We now intend to build deep learning methods to predict a plate’s vibroacoustic and subsequently optimize the beadings patterns to achieve a desired vibroacoustic response.
Within the context of this project, several research questions could be explored: - What strategies can we use to represent the vibroacoustic response compactly, easing the deep learning task and improving interpretability? - How can we construct a neural network capable of optimizing beading patterns? - Can we incorporate physics-based deep learning methods to reduce our dependency on large-scale data generation? If we address the design task for these comparatively simple examples, it could establish a foundation for scaling up deep learning approaches to tackle more complex designs in the future.
To apply please send an email to Jan van Delden and Timo Lüddecke stating your interest in this research and detailing your relevant skills.