Abstract
Deep neural operators have seen much attention in the scientific machine learning community over the last couple of years due to their capability of efficiently learning the nonlinear operators mapping from input function spaces to output function spaces showing good generalization properties. This work will show how to set up a performant DeepONet architecture in acoustics for predicting 2-D sound fields with parameterized moving sources for real-time applications. A sensitivity analysis is carried out with a focus on the choice of network architectures, activation functions, Fourier feature expansions, and data fidelity to gain insight into how to tune these models. Specifically, a default feed-forward neural network (FNN), a modified FNN, and a convolutional neural network will be compared. This work will de-mystify the DeepONet and provide helpful knowledge from an acoustical point of view.
Original language | English |
---|---|
Title of host publication | Proceedings of 10th Convention of the European Acoustics Association |
Number of pages | 8 |
Publication date | 2023 |
Publication status | Published - 2023 |
Event | 10th Convention of the European Acoustics Association - Politecnico di Torino, Torino, Italy Duration: 11 Sept 2023 → 15 Sept 2023 https://www.fa2023.org/ |
Conference
Conference | 10th Convention of the European Acoustics Association |
---|---|
Location | Politecnico di Torino |
Country/Territory | Italy |
City | Torino |
Period | 11/09/2023 → 15/09/2023 |
Internet address |
Keywords
- Neural operators
- Sensitivity analysis
- Virtual acoustics
- DeepONet