A practical ML workflow to quantify dynamic reservoir property changes from 4D seismic

Abstract

This study demonstrates how machine learning is used to quantify reservoir property changes from 4D seismic. We apply an end-to-end machine learning workflow to deliver saturation and pressure changes in reservoir with time. We screened and categorized the sources impacting seismic attributes to understand their influences on the final prediction of pressure and saturation changes. In the Gulf of Mexico (GOM) case study, we utilized a synthetic seismic as the training dataset, generated from a history-matched simulation model and a calibrated rock physics model. The synthetic seismic provided direct estimations of the input features’ quality and uncertainty ranges. Finally, we applied an iterative approach to estimate changes in the water saturation, gas saturation and pore pressure of the reservoir. The iterative approach compares the predicted outputs against production, injection and pressure data gathered at the producers and the injectors creating a closed feedback loop. The proposed ML workflow can be applied to other fields with 4D seismic data after calibrating 4D synthetic seismic with a known rock physics model.