Improved Training Convergence of Seismic Multifrequency Wavefield Simulation Based on GaborPINN With Halton Sequence

In solving for frequency-domain seismic wavefields, traditional numerical methods, such as frequency-domain finite difference (FDFD), require multiple computations for each fixed frequency. To overcome this issue, we choose to incorporate frequency distribution into the input of the physics-informed neural network (PINN) to predict multifrequency wavefields with a single training session. However, for VanillaPINN, simply increasing the input dimensions without embedding the sinusoidal feature of seismic waves into the network will lead to convergence difficulty. To address this problem, we develop a seismic wavefield simulation method based on GaborPINN to simulate multifrequency wavefields with improved training convergence. Furthermore, we propose to use the Halton sequence to collect the input coordinates, and the wavefield results demonstrate that the Halton sequence, with its more uniform distribution, achieves higher accuracy than the conventional random sampling. We demonstrate the superiority of the proposed method over the VanillaPINN and the Fourier feature PINN on a section of the Sigsbee2A model.