Slope assisted Physics-informed neural networks for seismic signal separation with applications on ground roll removal and interpolation

Abstract

The knowledge of the local slope field of prestack seismic data is essential in several seismic signal processing tasks. Building on our previous slope-assisted, physics-informed seismic interpolation framework, dubbed PINNslope, we introduce a series of enhancements that elevate the framework's versatility. This ultimately enables its application to different signal separation problems, with a specific focus on ground roll removal. To begin with, the local slope estimated using our physics-informed neural networks framework is compared against the analytical local slope and those obtained from several conventional slope estimation algorithms. This comparison showcases that our prediction better approximates the analytical one. Second, we use the derived relation between the analytical slope and the physics-informed neural networks estimated slope to constrain the slope estimation network in the ground roll removal problem, predicting only the clean reflections and avoiding the prediction of the ground roll. To address the large difference in the frequency content of the field data, we utilize a time derivative term in the loss function to emphasize the amplitude of the comparatively higher frequency reflection arrivals. Furthermore, we modify the framework loss function and architecture to demonstrate the possibility of predicting two separate components of the seismic data according to the estimate of two local slopes that can span opposite or different ranges of values between each other. The effectiveness of the double slope framework is demonstrated on a proof of concept of the deblending problem and for the interpolation of complex aliased data characterized by conflicting dips, two tasks that were not achievable using our single slope prediction network implementation.