PhD Defence by Nikolas Borrel-Jensen
Accelerated methods for computing the acoustic sound field in dynamic virtual environments with moving sources
Supervisors
- Principal supervisor: Associate Professor Cheol-Ho Jeong, DTU Electro, Denmark.
- Co-supervisor: Associate Professor Allan Peter Engsig-Karup, DTU Compute, Denmark.
Assessment committee
- Associate Professor Rasmus Ellebæk Christiansen, DTU Construct, Denmark (chair).
- Professor Jan Hesthaven, EPFL Switzerland.
- Professor Damian Murphy, University of York, UK.
Master of the ceremony
- Associate Professor Finn T. Agerkvist, DTU Electro, Denmark.
Abstract:
Realistic sound is essential in virtual environments, such as computer games, virtual and augmented reality, metaverses, and spatial computing. The wave equation describes wave phenomena such as diffraction and interference, and the solution can be obtained using accurate and efficient numerical methods. Due to the often demanding computation time, the solutions are calculated offline in a pre-processing step. However, pre-calculating acoustics in dynamic scenes with hundreds of source and receiver positions are challenging, requiring intractable memory storage.
To address this challenge, we propose using scientific machine learning techniques to predict the pressure field for any combination of sound source (think loudspeaker) and receiver positions (think microphone). The model output for a given source and receiver would be an impulse response encoding the acoustic properties of the room. Physics-informed neural networks and deep neural operators have been investigated. For the latter, we have demonstrated the method's effectiveness in a dome measuring 36m3, which includes intricate geometries. For the first time, a machine learning model can accurately predict the full-wave propagation in 3D, paving the way for future immersive experiences.
Training machine learning models often require a large amount of data that can be computationally expensive to obtain; hence this PhD thesis also investigates efficient numerical methods for generating accurate training data. Specifically, a rectangular domain decomposition method is proposed, enabling error-free sound propagation in the bulk of the domain consisting of air. The efficient but less flexible method approximating the air domain is coupled with a versatile and flexible, but less efficient method operating near the boundary, capable of accurately modeling complex geometries and boundary materials.
Contact
Cheol-Ho Jeong Associate Professor Department of Electrical and Photonics Engineering chje@dtu.dk