Developing deep learning methods for pre-stack seismic data inversion

Abstract

Inversion of pre-stack seismic data is important for building accurate models of hydrocarbon reservoirs used to estimate reserves and set up efficient production strategies. Conventional methods often involve challenging procedures. The limited bandwidth, noisy nature, and fact that the pre-stack seismic data has more than one component all affect how accurate and stable the inversion solution is. Deep learning (DL) is a cost-effective and accurate method for predicting spatially distributed properties. We utilized specific deep-learning algorithms to create a methodology for estimating porosity based on angle gathers. The workflow involved trace editing for amplitude variation compensation, noise reduction, and organizing data into appropriate common depth point (CDP) gathers. Well-logging data and approved seismic horizons were utilized to obtain the corresponding CDP gather’ porosity labels. From initial CDP gathers, we computed a series of seismic attributes and developed an efficient technique to identify the most relevant ones that we utilized to train base models. We explored different hyper-parameters to determine the optimal characteristics for the method's objective function. We subsequently developed an integration approach that assigns appropriate weights to aggregate individual base models to create a robust model. We developed a statistical analysis method to determine the confidence level in the final prediction. In real-world scenarios, the new methodology outperformed conventional and popular DL inversion approaches, with an R2 of 0.989 compared to 0.954 and 0.967, a mean porosity of 0.174868 compared to 0.174985, and a mean uncertainty of ±0.000714 at the well location.