Physics-informed neural networks for viscoacoustic wave propagation: forward modelling, inversion and discretization sensitivity Volume: Preprint submitted to arXiv

Chaohua Liang 1, Xingliang Peng 2, Jun Matsushima 1 1 The University of Tokyo Graduate School of Frontier Sciences Kashiwa, Japan 2 College of Water Conservancy, Jiangxi University of Water Resources and Electric Power Nanchang 330099, China · 2026

Read Physics-informed neural networks for viscoacoustic wave propagation: forward modelling, inversion and discretization sensitivity Volume: Preprint submitted to arXiv by Chaohua Liang 1, Xingliang Peng 2, Jun Matsushima 1 1 The University of Tokyo Graduate School of Frontier Sciences Kashiwa, Japan 2 College of Water Conservancy, Jiangxi University of Water Resources and Electric Power Nanchang 330099, China (2026) 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-2601-16068.txt.

Introduction

Chaohua Liang, Xingliang Peng and Jun Matsushima apply neural nets to sound waves in rock that lose strength. Their 2026 preprint sits in earth wave maths. Wave runs help teams read deep rock and find stores or risks. Grid based ways to run waves can be slow. They need step by step runs for each new guess of the ground.

The team uses nets that learn from the laws of wave move. Such nets can run waves forth and also work back to the ground form. The authors test how the grid step size in training shapes the result. Care with steps should keep the net true to real waves. The work aims for fast, sound earth wave tools.

Catalog record

CREATORChaohua Liang 1, Xingliang Peng 2, Jun Matsushima 1 1 The University of Tokyo Graduate School of Frontier Sciences Kashiwa, Japan 2 College of Water Conservancy, Jiangxi University of Water Resources and Electric Power Nanchang 330099, China
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-16068

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 networks for viscoacoustic wave propagation: forward modelling, inversion and discretization sensitivity Volume: Preprint submitted to arXiv Chaohua Liang 1, Xingliang Peng 2, Jun Matsushima 1 1 The University of Tokyo Graduate School of Frontier Sciences Kashiwa, Japan 2 College of Water Conservancy, Jiangxi University of Water Resources and Electric Power Nanchang 330099, China Abstract Seismic wave forward and inverse modeling not only provide the theoretical foundation for understanding the Earth’s complex deep structures, but also constitute core techniques for subsurface resource exploration and geological hazard assessment. Traditional numerical approaches, such as the finite-difference method (FDM) and finite-element method (FEM), typically rely on grid-based discretization and iterative forward simulations when solving inverse problems. Although deep learning has shown considerable potential in seismic applications, its effectiveness is often limited by the requirement for large volumes of labelled data and the insufficient incorporation of physical constraints. In this study, we propose a unified physics-informed neural network (PINN) framework for both forward modeling and parameter inversion of viscoacoustic wave propagation. The proposed framework is validated through comparative numerical experiments. For forward modeling, the network accurately reproduces wavefields, amplitude attenuation, and phase characteristics across different velocity models. For inversion, we evaluate the reconstruction of velocity and attenuation models from temporally sparse observations and random initial guesses. The results indicate that the proposed PINN approach achieves stable and acceptable accuracy when benchmarked against reference solutions, while exhibiting reduced sensitivity to spatial discretization compared with grid-based numerical methods. More importantly, by embedding a viscoacoustic wave equation into the learning process, the framework provides a physically consistent representation of intrinsic attenuation effects beyond purely acoustic approximations. In addition, the capability of simultaneously recovering velocity and attenuation parameters under sparse sampling conditions highlights the robustness and flexibility of the proposed method. These results demonstrate the feasibility of the method and highlight its potential for high-resolution seismic exploration. \pagerange keywords Physics-Informed Neural Networks (PINNs), Viscoacoustic wave, Forward and inverse modeling, Attenuation and dispersion, Seismic exploration. 1 Introduction Accurate seismic imaging is essential for resolving subsurface geological structures, particularly in the exploration and assessment of energy resources such as methane hydrates, commonly referred to as “combustible ice” owing to their high energy density and widespread occurrence. Over the past few decades, numerical methods such as the finite difference method (FDM) and finite element method (FEM) have been firmly established as the standard tools for seismic forward modeling and inverse analysis [ 6 , 9 , 10 ] . These grid-based approaches typically discretize the domain into meshes to approximate the partial differential equations (PDEs). Although conventional approaches, such as FDM and full waveform inversion (FWI), can achieve high accuracy in solving partial differential equations, they rely on repeated high-fidelity forward simulations during inversion, which leads to substantial computational cost. What’s more, the misfit function of FWI is prone to converging toward local minima rather than the global optimum [ 15 ] . With the rapid evolution of deep learning (DL), together with the exponential increase in GPU-accelerated computing power, data-driven paradigms have witnessed explosive growth in the field of seismology. The enhanced hardware capabilities have allowed researchers to train sophisticated neural networks on massive seismic datasets, establishing complex non-linear mappings between measurements and subsurface properties. For example, Wu and Lin [ 17 ] proposed InversionNet, a framework that employs a deep convolutional encoder-decoder architecture to directly learn the non-linear mapping from seismic data to subsurface velocity structures. Similarly, Recurrent Neural Networks (RNNs) have been successfully applied to seismic impedance inversion [ 1 ] , leveraging the temporal correlations within seismic traces. Furthermore, Generative Adversarial Networks (GANs) have shown remarkable performance in seismic data reconstruction [ 5 ] . However, these supervised methods require lots of labeled data, which is difficult to obtain in real-world seismic exploration. Besides, these models are essentially black-box mappers, therefore may generate physically unreasonable results. Recently, Physics-informed neural networks (PINNs) by Raissi [ 11 ] have received widespread attention as a new scientific computation algorithm. PINNs explicitly embed the governing Partial Differential Equations (PDEs) into the loss function as a regularization mechanism, which provides new ideas for addressing problems faced in conventional methods. Specifically, in inverse problems, PINNs integrate the forward and inverse processes into a unified framework. By treating the unknown parameters in PDEs as trainable variables, PINNs leverage automatic differentiation to update these parameters simultaneously with the outputs of the forward process. PINNs is now applied to miscellaneous scientific domains including fluid mechanics [ 11 , 12 ] , geophysics [ 13 , 14 ] , solid mechanics [ 4 ] , heat transfer [ 18 ] and biomedical imaging [ 16 ] . However, vanilla PINNs face certain challenges in seismic exploration. Studies indicate that vanilla PINNs suffer from point-source singularity in solving wave equations. To address this problem, a hybrid initialization strategy that utilizes the solutions generated by the FDM method at early time steps as initial constraints is proposed [ 2 ] . What’s more, while Physics-Informed Neural Networks (PINNs) have recently emerged as a powerful tool for solving seismic wave equations, the majority of existing studies restrict their scope to ideal acoustic or elastic media [ 13 , 14 , 19 ] , explicitly neglecting the intrinsic attenuation properties of the subsurface. However, in realistic geological formations, the viscosity of rocks and the presence of fluids inevitably cause the conversion of mechanical energy into heat. This phenomenon, known as seismic absorption, induces velocity dispersion and amplitude dissipation, which significantly alter the waveform characteristics [ 8 ] . Furthermore, their application to inverse problems—particularly in complex viscoacoustic media—remains noticeably underexplored. To address these limitations, we establish a physics-informed architecture specifically designed to capture these complex viscoacoustic behaviors and is able to run forward and inverse process simultaneously. In this paper, we propose a unified physics-informed neural network framework for viscoacoustic wave propagation, capable of both forward modeling and physical parameter inversion. To avoid point-source singularities, early-time wavefields are incorporated as initial constraints, and open boundary conditions are employed to suppress artificial boundary reflections. The main contributions of this work are threefold. First, we demonstrate that the proposed framework accurately models viscoacoustic wavefields, capturing both attenuation and phase characteristics in homogeneous and layered velocity models. Second, we show that the PINN formulation remains stable under coarse discretization and sparse sampling conditions, exhibiting a fundamentally different degradation behavior compared to finite-difference solvers. Third, we introduce a simultaneous forward–inverse strategy that enables robust recovery of both velocity and absorption parameters from random initial models, highlighting the framework’s ability to solve the inverse problem by treating physical parameters as learnable variables. 2 Problem Formulation In this paper, our goal is to investigate the ability of PINNs for viscoacoustic propagation in 2D homogeneous media. The governing equations based on the Maxwell rheological model for propagating pressure in the time-space domain are given by [ 3 ] : \frac{1}{v^{2}}\frac{\partial^{2}P}{\partial t^{2}}-\nabla^{2}P+\frac{g}{v}\frac{\partial P}{\partial t}=S(x_{s},t) (1) Where P , v , and t denote the pressure, velocity, and time, respectively. S represents the source term. Consistent with standard seismic modeling practices, a Ricker wavelet is employed as the source function in this study. g is the absorption coefficient, given by: g=\frac{2\pi f_{0}}{vQ} (2) Where f_{0} and Q denote the

Read the complete text