Contribution
Optimizing acoustic system simulations with Bayesian Neural Networks (BNN): Reducing computational effort through intelligent sampling
* Presenting author
Abstract:
In acoustic applications, sound pressure levels or transmission losses are typically calculated over a specified frequency range. To capture the system's behavior across the entire frequency spectrum, numerous numerical simulations are often required, performed incrementally across the bandwidth. Data-driven methods help reduce computational costs by using interpolation to estimate the system's behavior from only a few samples. Without prior knowledge of the system, sample selection is often random, equidistant, or based on user experience, leading to highly variable success and reduced reliability. While increasing the number of training points can improve the accuracy of these approximations, it also raises computational demands. Bayesian neural networks aim to address this issue by minimizing the number of training samples, selecting only those that provide the most valuable information for predicting system behavior, thereby reducing computational effort.