Introduction
Predicting outdoor sound propagation is a complex challenge due to atmospheric variables such as wind velocity, temperature gradients, and irregular terrain. Traditional acoustic models often struggle to capture these interactions accurately. Consequently, researchers utilize the Linearized Euler Equations (LEE) to model the propagation of sound waves within a background atmosphere. However, solving the LEE using standard numerical techniques like Finite-Difference Time-Domain (FDTD) requires high-resolution spatial and temporal discretization, leading to prohibitive computational costs for large-scale urban environments or high-frequency signals. This study investigates the potential of physics-informed machine learning to mitigate these computational bottlenecks.
Methodology
The research evaluates two distinct neural network architectures designed to solve partial differential equations (PDEs):
- Physics-Informed Neural Networks (PINNs): This approach embeds the LEE directly into the network's loss function. The model is trained to minimize the residual of the governing equations across the spatiotemporal domain. To effectively model the oscillatory nature of acoustic waves, the authors implement a Sinusoidal Representation Network (SIREN) architecture, utilizing periodic activation functions instead of standard Rectified Linear Units (ReLUs).
- Extreme Learning Machine Ordinary Differential Equations (ELM-ODE): This method employs a shallow neural network with fixed, randomly sampled spatial weights. By decoupling spatial and temporal components, the spatial domain is represented by fixed basis functions, allowing the temporal evolution to be solved as a system of Ordinary Differential Equations (ODEs).
The primary objective is to analyze the trade-offs between precision and computational efficiency when applying these methods to acoustic benchmarks.
Experimental Setup
The performance of both models was assessed using synthetic benchmarks rather than experimental datasets. The test cases included:
- 1D Simulation: A Gaussian pressure pulse propagating through a 200m domain.
- 2D Simulation: A planar Gaussian pulse subjected to horizontal wind advection within a 200m x 200m domain.
Mean Absolute Error (MAE) was calculated by comparing neural network outputs against ground truth data generated by traditional Finite-Difference (FD) solvers.
Results and Discussion
The comparison revealed a significant disparity in computational performance and accuracy:
- Computational Efficiency: ELM-ODE demonstrated a massive speed advantage. In 1D simulations, ELM-ODE reached a solution in approximately 2.98 seconds, whereas PINNs required over an hour (1 hour 2 minutes). In 2D scenarios, the gap remained substantial, with ELM-ODE solving in 23 minutes compared to 7.6 hours for PINNs.
- Accuracy: PINNs achieved higher precision, yielding an MAE of $1.85 \times 10^{-3}$ in 1D, while ELM-ODE produced an MAE of $7.40 \times 10^{-3}$.
- Qualitative Performance: Both methods successfully replicated complex wave phenomena, including boundary reflections and wind-induced advection.
The choice of activation functions proved critical; the SIREN-based PINN was essential for capturing the high-frequency oscillations inherent in sound propagation. However, ELM-ODE's scalability may be limited in 3D applications due to the increasing size of the coupling matrix required for multi-variable systems.
Conclusion
While PINNs offer superior accuracy, the extreme computational speed of ELM-ODE makes it a highly attractive candidate for practical acoustic modeling where rapid iteration is required. Future research should focus on implementing complex, frequency-dependent boundary conditions and optimizing these models for 3D environments and real-world topographic data.
Summary updated