Multiple scattering simulation via physics-informed neural networks

Siddharth Nair,Timothy F. Walsh,Greg Pickrell,Fabio Semperlotti

2024-03-07

Abstract:This work presents a physics-driven machine learning framework for the simulation of acoustic scattering problems. The proposed framework relies on a physics-informed neural network (PINN) architecture that leverages prior knowledge based on the physics of the scattering problem as well as a tailored network structure that embodies the concept of the superposition principle of linear wave interaction. The framework can also simulate the scattered field due to rigid scatterers having arbitrary shape as well as high-frequency problems. Unlike conventional data-driven neural networks, the PINN is trained by directly enforcing the governing equations describing the underlying physics, hence without relying on any labeled training dataset. Remarkably, the network model has significantly lower discretization dependence and offers simulation capabilities akin to parallel computation. This feature is particularly beneficial to address computational challenges typically associated with conventional mesh-dependent simulation methods. The performance of the network is investigated via a comprehensive numerical study that explores different application scenarios based on acoustic scattering.

Computational Physics

What problem does this paper attempt to address?

This paper proposes a framework based on Physics-Informed Neural Networks (PINN) for simulating acoustic scattering problems. Traditional numerical methods such as the finite element method have high computational costs when dealing with high-frequency, multi-scattering, and multi-scale problems due to their dependence on discrete parameters. PINN reduces the reliance on labeled training data and the dependence on discretization by enforcing the governing equations that describe the underlying physics during the training process, providing a parallel computing-like capability.

The PINN architecture combines the superposition principle of linear wave interaction, enabling the simulation of scattering fields generated by arbitrarily shaped rigid scatterers, especially suitable for high-frequency problems. Unlike conventional data-driven neural networks, PINN does not require a large amount of training data but directly utilizes physical laws for training, ensuring the physical consistency of the predicted results.

The main contributions of this paper include two aspects: 1) basic PINN (b-PINN) for simulating acoustic scattering of single or few arbitrarily shaped scatterers; 2) super PINN (s-PINN) that solves problems in scenarios with a large number of scatterers by utilizing the superposition principle of linear acoustics and achieves parallel processing of subdomains, reducing the total simulation time.

The study evaluates the performance of the PINN framework in simulating acoustic problems with arbitrarily shaped scatterers, particularly in the mid-to-high-frequency range, through a series of numerical experiments. This approach has the potential to reduce computational costs associated with traditional methods based on solving systems of differential equations.