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

Seismic wave forward simulation is the basis of seismic inversion and imaging. In recent years, deep learning methods represented by physics-informed neural networks (PINNs) have been widely applied in the field of scientific computing, and have obtained significant attention as mesh-free methods in seismic inversion. Currently, the training of PINN is usually constructed by fixed points in seismic wavefield simulation, and the large number of training points can result in high-computational costs. This letter proposes an adaptive strategy for PINN training point selection based on the velocity distribution and residuals of viscoacoustic wave equations. This strategy initializes the positions and quantities of training points based on the velocity distribution of the model and dynamically adjusts the training points based on the values of the loss function during the calculation process. The proposed approach reduces the number of training points, thereby improving the training efficiency of PINN. Numerical examples including the homogeneous model, the layered model with irregular topography, and the Marmousi2 model show that the proposed sampling strategy effectively reduces the calculation time by 54.37% while ensuring accuracy.