An adaptive method of fundamental solutions driven by physics-informed neural networks for acoustic problems

Selecting appropriate source points remains a significant challenge in the method of fundamental solutions (MFS), particularly for acoustic problems involving geometrically complex domains and multiple types of boundary conditions. Building upon the existing adaptive MFS framework developed using physics-informed neural networks, this study extends the approach to Helmholtz acoustic problems in two- and three-dimensional domains, including both interior and exterior fields. Unlike the Laplace and biharmonic problems considered in the previous framework, Helmholtz problems involve complex-valued acoustic fields. To address this issue, the real and imaginary components are separated in the loss formulation and jointly used for neural network training. Source-point locations are represented in polar coordinates with uniformly distributed angular coordinates. For a prescribed number of source points, the angular coordinates are used as neural network inputs to predict the corresponding radial coordinates, while the loss is constructed from the MFS boundary residuals and prescribed boundary conditions to adaptively optimize source-point locations within a prescribed radial range. Numerical experiments demonstrate that conventional MFS accuracy is sensitive to the manually selected source-point distance, whereas the proposed adaptive MFS reduces the dependence on empirical source-point placement. Under the test conditions considered, it achieves comparable accuracy to conventional MFS with fewer source points. In the two-dimensional cases, it requires approximately half the time of physics-informed neural networks, while in the three-dimensional cases, it shows relatively small variations in accuracy and time across the tested network configurations and requires substantially less time than physics-informed neural networks.

Authors

Institutions

Publication Details

Journal
Journal of Applied Physics
Published
2026-11-19
DOI
https://doi.org/10.1063/5.0349841
Primary Topic
Model Reduction and Neural Networks
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An adaptive method of fundamental solutions driven by physics-informed neural networks for acoustic problems

Lin Qiu, Fajie Wang, Xingxing Yue, Zhanpeng Yan et al.
Journal of Applied Physics
Model Reduction and Neural Networks
article

An adaptive method of fundamental solutions driven by physics-informed neural networks for acoustic problems

Lin Qiu, Fajie Wang, Xingxing Yue, Zhanpeng Yan, Xin Li
article en

Abstract

Selecting appropriate source points remains a significant challenge in the method of fundamental solutions (MFS), particularly for acoustic problems involving geometrically complex domains and multiple types of boundary conditions. Building upon the existing adaptive MFS framework developed using physics-informed neural networks, this study extends the approach to Helmholtz acoustic problems in two- and three-dimensional domains, including both interior and exterior fields. Unlike the Laplace and biharmonic problems considered in the previous framework, Helmholtz problems involve complex-valued acoustic fields. To address this issue, the real and imaginary components are separated in the loss formulation and jointly used for neural network training. Source-point locations are represented in polar coordinates with uniformly distributed angular coordinates. For a prescribed number of source points, the angular coordinates are used as neural network inputs to predict the corresponding radial coordinates, while the loss is constructed from the MFS boundary residuals and prescribed boundary conditions to adaptively optimize source-point locations within a prescribed radial range. Numerical experiments demonstrate that conventional MFS accuracy is sensitive to the manually selected source-point distance, whereas the proposed adaptive MFS reduces the dependence on empirical source-point placement. Under the test conditions considered, it achieves comparable accuracy to conventional MFS with fewer source points. In the two-dimensional cases, it requires approximately half the time of physics-informed neural networks, while in the three-dimensional cases, it shows relatively small variations in accuracy and time across the tested network configurations and requires substantially less time than physics-informed neural networks.

Journal of Applied PhysicsVol. 140(19)
Qingdao University (CN)
Model Reduction and Neural Networks
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon