Physics-informed neural networks (PINNs) have been used by researchers to solve partial differential equation (PDE)-constrained problems. We evaluate PINNs to solve for frequency-domain acoustic wavefields. PINNs can solely use PDEs to define the loss function for optimization without the need for labels. Partial derivatives of PDEs are calculated by mesh-free automatic differentiations. Thus, PINNs are free of numerical dispersion artifacts. It has been applied to the scattered acoustic wave equation, which relied on boundary conditions (BCs) provided by the background analytical wavefield. For a more direct implementation, we solve the nonscattered acoustic wave equation, avoiding limitations related to relying on the background homogeneous medium for BCs. Experiments support our following insights. Although solving time-domain wave equations using PINNs does not require absorbing boundary conditions (ABCs), ABCs are required to ensure a unique solution for PINNs that solve frequency-domain wave equations, because the single-frequency wavefield is not localized and contains wavefield information over the full domain. However, it is not trivial to include the ABC in the PINN implementation, so we develop an adaptive amplitude-scaled and phase-shifted sine activation function, which performs better than the previous implementations. Because there are only two outputs for the fully connected neural network (FCNN), we validate a linearly shrinking FCNN that can achieve a comparable and even better accuracy with a cheaper computational cost. However, there is a spectral bias problem, that is, PINNs learn low-frequency wavefields far more easily than higher frequencies, and the accuracy of higher frequency wavefields is often poor. Because the shapes of multifrequency wavefields are similar, we initialize the FCNN for higher frequency wavefields by that of the lower frequencies, partly mitigating the spectral bias problem. We further incorporate multiscale positional encoding to alleviate the spectral bias problem. We share our codes, data, and results via a public repository.
4 September 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Surface-related multiples (SRMs) can significantly interfere with primaries if not properly addressed. The robust estimation of primaries by sparse inversion (REPSI) eliminates EPSI's free parameters but relies on a highly complex spectral-projected gradient (SPG$\ell _{1}$) algorithm. Compared to SRM elimination (SRME) via prediction and adaptive subtraction, EPSI-type methods require many iterations and incur high computational costs due to extensive matrix multiplications and forward/inverse fast Fourier transforms (FFTs) on prestack shot gathers. Moreover, the common assumption of a constant wavelet across all shots may fail for field data. Based on the feedback primary-multiple model, we propose a simple and efficient estimation of multiples and sources (EMS) method to estimate primaries, SRMs, and both shot-constant and shot-variant sources without subsurface model assumptions or source priors. We first employ a simple and accelerated linearized Bregman (ALB) algorithm to solve the large-scale sparsity-promoting problem of retrieving the surface-free Green's function. Tests demonstrate ALB converges faster than SPG$\ell _{1}$. Shot-constant and shot-variant sources are then estimated via a straightforward, exact, and efficient line-search scaling, replacing the iterative and expensive LSQR algorithm in REPSI. Finally, EMS is significantly accelerated using GPUs, which effectively handle the computational demands of EPSI-type methods. Validation on synthetic and field open datasets confirms EMS's effectiveness, robustness, and benefits. EMS outperforms REPSI in convergence speed, reconstruction quality, simplicity, and runtime, and surpasses SRME with more accurate SRM estimates. Our EMS code is released as open source.