Physics-Informed Attention-Based Neural Network for Full-Waveform Inversion

Summary Seismic full-waveform inversion (FWI) is a nonlinear highly ill-posed inverse problem that is commonly used to estimate a high-resolution velocity model of the subsurface. Performing a successful FWI study requires an accurate initial model of the subsurface that fits the observed data with an error of less than half the period. In this study, we introduce physics-informed neural networks (PINNs) with an attention mechanism for performing FWI without requiring an initial model. We compare the results of the traditional FWI with the ones of PINN with and without attention mechanism. The three methods are assessed using the Marmousi model. Our study indicates that using PINNs with an attention mechanism can lead to higher accuracy compared to the traditional FWI and PINNs without the attention mechanism.