Abstract:Physics-informed neural networks (PINNs) have great potential for flexibility and effectiveness in forward modeling and inversion of seismic waves. However, coordinate-based neural networks (NNs) commonly suffer from the "spectral bias" pathology, which greatly limits their ability to model high-frequency wave propagation in sharp and complex media. We propose a unified framework of Fourier feature physics-informed neural networks (FF-PINNs) for solving the time-domain wave equations. The proposed framework combines the stochastic gradient descent (SGD) strategy with a pre-trained wave velocity surrogate model to mitigate the singularity at the point source. The performance of the activation functions and gradient descent strategies are discussed through ablation experiments. In addition, we evaluate the accuracy comparison of Fourier feature mappings sampled from different families of distributions (Gaussian, Laplace, and uniform). The second-order paraxial approximation-based boundary conditions are incorporated into the loss function as a soft regularizer to eliminate spurious boundary reflections. Through the non-smooth Marmousi and Overthrust model cases, we emphasized the necessity of the absorbing boundary conditions (ABCs) constraints. The results of a series of numerical experiments demonstrate the accuracy and effectiveness of the proposed method for modeling high-frequency wave propagation in sharp and complex media.

What problem does this paper attempt to address?

### Problems Addressed by the Paper The paper primarily addresses a series of challenges encountered when simulating seismic wavefields in complex media, particularly the accurate modeling of high-frequency wave propagation. Specifically: 1. **Point Source Singularity**: Traditional Physics-Informed Neural Networks (PINNs) face the issue of point source singularity when solving wave equations in both time and frequency domains. This problem is mitigated by approximating the Dirac delta function using smooth spatial mapping functions, such as Gaussian functions. 2. **Spectral Bias**: Neural networks tend to learn low-frequency functions more easily than high-frequency functions, a phenomenon known as "spectral bias." The paper introduces Fourier feature mapping, which transforms input coordinates into a series of high-frequency sine waves, enhancing the neural network's ability to learn high-frequency functions. 3. **Absorbing Boundary Conditions (ABCs)**: To better simulate wave propagation in infinite or semi-infinite media, the paper explores the importance of integrating absorbing boundary conditions into the loss function. By introducing boundary conditions based on parabolic approximation as soft regularization terms, the issue of wave propagation in infinite domains is addressed. 4. **Hard Embedding of Initial Conditions**: To ensure the accurate implementation of zero initial conditions, a hard embedding scheme is introduced, guaranteeing that the system is at rest at the initial moment. 5. **Time Domain Decomposition Training Strategy**: To improve the modeling capability of PINNs for complex wave propagation behavior, a time domain decomposition strategy is introduced. This strategy decomposes the optimization problem over the entire time domain into several relatively simpler problems, thereby reducing the optimization difficulty. The experimental results using the non-smooth Marmousi model and the overthrust model demonstrate the accuracy and effectiveness of the proposed methods in modeling high-frequency wave propagation in complex media.

Physics-informed Neural Networks with Fourier Features for Seismic Wavefield Simulation in Time-Domain Nonsmooth Complex Media

Physics-enhanced neural networks for half-space seismic wave modeling

Wave simulation in non-smooth media by PINN with quadratic neural network and PML condition

Physics-constrained Neural Networks for Half-Space Seismic Wave Modeling

PINN-Based Seismic Wavefield Simulation With Learnable Multiscale Fourier Feature Mapping and Adaptive Activation Function

Modeling multisource multifrequency acoustic wavefields by a multiscale Fourier feature physics-informed neural network with adaptive activation functions

SeismicNet: Physics-informed neural networks for seismic wave modeling in semi-infinite domain

Physics-informed Neural Networks (PINNs) for Wave Propagation and Full Waveform Inversions

Helmholtz-equation Solution in Nonsmooth Media by a Physics-Informed Neural Network Incorporating Quadratic Terms and a Perfectly Matching Layer Condition

Accelerating the Convergence of Physics Informed Neural Networks for Seismic Wave Simulation

Diminishing Spectral Bias in Physics-Informed Neural Networks Using Spatially-Adaptive Fourier Feature Encoding

Physics-informed neural network for seismic wave inversion in layered semi-infinite domain

Adaptive-Sampling Physics-Informed Neural Network for Viscoacoustic Wavefield Simulation

Research progress of physics-informed neural network in seismic wave modeling

Self-adaptive physics-driven deep learning for seismic wave modeling in complex topography

Practical Aspects of Physics-Informed Neural Networks Applied to Solve Frequency-Domain Acoustic Wave Forward Problem

Fourier Domain Physics Informed Neural Network

Multi-frequency wavefield modeling of acoustic VTI wave equation using physics informed neural networks

Numerical Solver-Independent Seismic Wave Simulation Using Task-Decomposed Physics-Informed Neural Networks.

Physics-informed Neural Networks for Fully Non-Linear Free Surface Wave Propagation

Seismic Inversion Based on Acoustic Wave Equations Using Physics-Informed Neural Network