Speaker
Matteo Calafà
(DTU Electro)
Description
Holomorphic neural networks have recently been proposed as two-dimensional universal function approximators constrained to produce holomorphic outputs. In this work, we combine holomorphic neural networks with Vekua operators toconstruct a machine learning architecture that inherently generates solutions tothe Helmholtz equation. This approach is particularly advantageous for two-dimensional boundary value problems in room acoustics at low frequencies, astraining can be restricted to enforcing only boundary conditions, resulting in substantial speed-ups and improved accuracy compared to traditional physics-informed neural networks. Possible applications include real-time sound field prediction, inverse design and surrogate modelling.
Authors
Matteo Calafà
(DTU Electro)
Cheol-Ho Jeong
(DTU Electro)