Contribution of Frequency and Training Model on AI-Based Velocity Prediction

Velocity model building is important in providing subsurface velocity models for workflows such as seismic imaging and interpretation. Velocity model building techniques, such as ray-based tomographic approaches are not very effective in complex geological settings. Full waveform inversion (FWI) approaches are computationally intensive and sensitive to an initial model. The physics-guided deep learning-based velocity model building, that involves deterministic, physics-based modelling and data-driven deep learning components, is designed to capture the subsurface salt body shapes and locations, with a small amount of training models. In this paper, we discuss the influence of dominant frequency and training models on the velocity prediction by using a hybrid physics-guided neural network method. Our results show that, the higher the dominant frequency, the more accurate the prediction accuracy of the salt body shapes and background information. For more complicated velocity models and real datasets, simple synthetic training models are not capable of capturing the salt body shapes, nor the background information. A more practical synthetic training set with much more smoothed background layered structures is more suitable for predicting complicated models.