Physics-Informed Neural Network for Elastic Wave-Mode Separation

E. A. B. Alves, P. D. S. de Lima, D. H. G. Duarte, M. S. Ferreira, J. M. de Araújo, C. G. Bezerra · 2025

Read Physics-Informed Neural Network for Elastic Wave-Mode Separation by E. A. B. Alves, P. D. S. de Lima, D. H. G. Duarte, M. S. Ferreira, J. M. de Araújo, C. G. Bezerra (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-04600.txt.

Introduction

Alves and co-authors teach a physics-informed neural network to split elastic waves into their compression and shear parts inside uneven materials. Their 2025 paper starts from the reading problem: where rock or tissue varies, the two wave modes convert into each other and blur the picture. The classic split leans on the Helmholtz decomposition.

Their network solves a single scalar equation in place of the traditional vector one, a swap that cuts cost and scales better with dimension. Trials ran on a uniform model and on realistic uneven ones. The split modes matched those from standard numerical methods closely. The shear part also leaked less across the boundary than usual, which sharpens the read.

Catalog record

CREATORE. A. B. Alves, P. D. S. de Lima, D. H. G. Duarte, M. S. Ferreira, J. M. de Araújo, C. G. Bezerra
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-04600

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-Informed Neural Network for Elastic Wave-Mode Separation E. A. B. Alves, P. D. S. de Lima, D. H. G. Duarte, M. S. Ferreira, J. M. de Araújo, C. G. Bezerra Abstract Mode conversion in non-homogeneous elastic media makes it challenging to interpret physical properties accurately. Decomposing these modes correctly is crucial across various scientific areas. Recent machine learning approaches have been proposed to address this problem, utilizing the Helmholtz decomposition technique. In this paper, we investigate the capabilities of a physics-informed neural network (PINN) in separating P and S modes by solving a scalar Poisson equation. This scalar formulation offers a dimensionally scalable reduction in computational cost compared to the traditional vector formulation. We verify the proposed method in both homogeneous and realistic non-homogeneous elastic models as showcases. The obtained separated modes closely match those from conventional numerical techniques, while exhibiting reduced transverse wave leakage. I Introduction Understanding the dynamics of elastic waves in solids is crucial for a wide range of applications, including material characterization Kweun et al. 2017 ; Yang et al. 2018 ; Kim et al. 2020 ; Lee et al. 2024 and seismology Pujol 2003 ; Tromp 2020 . These waves can be described by two modes, corresponding to longitudinal P-waves and transverse S-waves. However, when either P or S waves encounter a material discontinuity, the scattering process generates both P and S modes, a phenomenon known as mode conversion. While metamaterials can be engineered to control or suppress mode conversion fully Liu and Liu 2015 ; De Ponti et al. 2021 ; Mei et al. 2022 ; Yu et al. 2023 ; Liu and Peng 2024 , it remains unavoidable in many other cases. Examples abound in the manufacturing, motoring, aerospace, and petroleum industries, to name but a few. In the latter case, this is particularly interesting in the field of seismology, where it presents significant challenges in seismic imaging. Thus, mode separation is crucial for correctly interpreting physical properties in realistic complex media. The Helmholtz decomposition can separate P and S modes in isotropic and anisotropic Fang et al. 2024 media by imposing a divergence-free (for S wave) and a curl-free (for P wave) condition Morse and Feshbach 1953 . In the context of multiparameter inverse problems, this decomposition can mitigate ambiguities caused by the fact that different observational signatures are indistinguishable, a phenomenon known as crosstalk Tromp et al. 2005 . Therefore, suppressing this effect can enhance specific seismic imaging methods Wang and Cheng 2017 ; Pan et al. 2018a ; Pan et al. 2018b ; Wang et al. 2018 ; Zhong et al. 2021 . Wavefield separation can be efficiently computed by solving an auxiliary vector Poisson equation, where the source term corresponds to the elastic displacement Zhu 2017 . The computational cost can be substantially reduced by transforming it into a scalar equation Zheng and Yao 2023 . In both vector and scalar approaches, P and S waves are subsequently constructed by employing adequate vector operations on the solution of the Poisson equation. Therefore, regardless of the separation method used, it typically involves solving a partial differential equation (PDE). In this regard, physics-informed neural networks (PINNs) have emerged as a practical machine learning approach for solving PDEs Raissi et al. 2019 ; Karniadakis et al. 2021 . PINN belongs to a class of neural networks that incorporates prior knowledge of physical laws into the training process. Such information enables PINN to obtain a superior solution compared to standard data-driven techniques without requiring large amounts of data. Besides its application in different areas Xu and Zhang 2021 ; Zhong et al. 2022 ; Lin and Yu 2022 ; Cornell et al. 2022 ; Liu et al. 2023 ; Norambuena et al. 2024 , PINN has been successfully employed in wave dynamics description Alkhalifah et al. 2021 ; Song et al. 2021a ; Duarte et al. 2025 and inverse-problem solutions Han and Lü 2025 ; Negahdari et al. 2025 . In particular, some authors have addressed the mode decomposition problem using various deep learning techniques Kaur et al. 2021 ; Wei et al. 2022 ; Huang et al. 2023 ; Sun et al. 2024 . Recently, a physical-constrained neural network Sun et al. 2025 and a particular PINN architecture Mu et al. 2025 have been proposed to separate the elastic wavefield following the Helmholtz decomposition. However, they solve a vector Poisson equation, which leads to increased neural network complexity as the number of spatial dimensions increases. Conversely, the scalar framework can reduce this cost by a factor of 1/2 in 2D and 1/3 in 3D problems Zheng and Yao 2023 . Therefore, this formulation offers a scalable computational-cost reduction compared to the traditional vector approach, enabling the use of a less complex neural network architecture. With that motivation, we investigate PINN capabilities in separate P/S elastic modes by solving a scalar Poisson equation. Furthermore, although the focus of the present communication is in the seismic context, our findings are transferable to other systems containing elastic waves in solid media. This paper is organized as follows. We revisit the elastic wavefield separation in sec. II , highlighting the differences between the vector and scalar formulation. sec. III describes the Fourier feature-embedded PINN architecture used for solving the Poisson equation. The decomposed modes for a simple homogeneous and a realistic non-homogeneous elastic model are presented in Section IV , followed by conclusions in Section V . Figure 1: The architecture of the Fourier features physics-informed neural network used to separate the elastic wavefield by solving Eq. ( 5 ), which only requires one output w(\mathbf{x}) . Figure 2: The loss history for the homogeneous model. The learning rate was halved every 2000 epochs. Figure 3: The first two panels show the horizontal and vertical components of the elastic wavefield in a homogeneous medium for a specific instant of time. The separated elastic modes obtained with the conventional numerical solver (second row) and the proposed PINN approach (third row). The difference between them (fourth row) illustrates that both approaches yield very similar results. Figure 4: A typical Brazilian pre-salt P velocity model used as a realistic example in the separation mode. Figure 5: The same as Fig. 2 , but for the non-homogeneous velocity model Fig. 4 . Figure 6: The same as Fig. 3 , but for a non-homogeneous elastic medium shown in Fig. 4 . II Elastic Wavefield Separation The elastic wave propagation in a homogeneous and isotropic medium is described by Main 1980 \rho\frac{\partial^{2}\mathbf{u}}{\partial t^{2}}=\mathbf{f}+(\lambda+2\mu)\nabla(\nabla\cdot\mathbf{u})-\mu\nabla\times\nabla\times\mathbf{u}\,, (1) where \mathbf{u} is the elastic wavefield, \mathbf{f} is the source wave term, \rho is the density, and \lambda and \mu are the Lamé coefficients. The longitudinal and transverse wave velocities are defined as v_{p}=\sqrt{(\lambda+2\mu)/\rho} and v_{s}=\sqrt{\mu/\rho} , respectively. The elastic wavefield can be decomposed into P and S waves using the Helmholtz decomposition Zhu 2017 , that is, \mathbf{u}=\mathbf{u}^{p}+\mathbf{u}^{s}\,, (2) where these fields satisfy \nabla\times\mathbf{u}^{p}=0 and \nabla\cdot\mathbf{u}^{s}=0 . Accordingly, these modes can be obtained through the relations \mathbf{u}^{p}=\nabla(\nabla\cdot\mathbf{w})\,,\quad\mathbf{u}^{s}=-\nabla\times\nabla\times\mathbf{w}\,, (3) under the assumption that the auxiliary vector field \mathbf{w} satisfies the following vector Poisson equation: \nabla^{2}\mathbf{w}=\mathbf{u}\,. (4) The computational cost involved in solving Eq. ( 4 ) grows with the number of spatial dimensions, which makes it cumbersome

Read the complete text