Contribution
Data-Driven and Physics-Informed Machine Learning for Outdoor Acoustic Wave Modeling using the Linearized Euler Equations
* Presenting author
Abstract:
The Linearized Euler Equations (LEE) describe acoustic wave propagation in the atmosphere, capturing atmospheric and topographic effects on sound waves. The high computational cost of solving LEE with traditional numerical methods motivates the exploration of machine learning methods to approximate its solutions. Besides attempting to improve computational efficiency, machine learning methods could also benefit from available simulation data and measurements.This study applies Physics-Informed Neural Networks (PINNs) and Fourier Neural Operators (FNOs) to approximate solutions to the LEE, each leveraging data differently. PINNs enforce physical laws while incorporating data, using initial and boundary conditions as well as sparse numerical or measurement data to maintain physical consistency. In contrast, FNOs act as data-driven surrogate models, learning from high-fidelity numerical simulations that inherently obey physical laws, despite not explicitly enforcing physical constraints.2D results show that PINNs ensure physically-consistent solutions but only converge after sufficient amounts of training, whereas FNOs achieve high efficiency and accuracy depending on the quality and range of training data. While both methods offer promising alternatives to traditional solvers, limitations remain in handling complex boundary conditions, incorporating real-world atmospheric variability, and extending these approaches to 3D simulations.