Unraveling the hidden relationships between seismic multiattributes, well-dynamic data, and Brazilian pre-salt carbonate reservoirs productivity: A shallow versus deep machine-learning approach

Abstract

Listen

Translate article

During the initial phases of an E&P project, the most reliable data on the reservoir’s deliverability are acquired via drill stem tests (DST), which provide the productivity per flow unit whenever production logging tool data are available. However, DSTs are restricted to a few kilometers, whereas seismic data cover large areas. The integration of these data has been challenging, particularly due to the difference in scale between them. So, a new workflow to determine the relationship between poststack seismic attributes and reservoir productivity using classic supervised (shallow) and deep-learning regression algorithms was developed. The DST parameters were predicted over the entire seismic cube, which can be extremely valuable for the decision-making process. The data set is from the Brazilian deep-water presalt carbonate reservoirs of the Mero Field, which is a well-explored area with a plethora of test and production data. It is adjacent to an underexplored Central Libra appraisal plan area, which is covered by the same seismic survey. Thus, any relationships between seismic attributes and well productivity data observed at the Mero Field are extrapolated to the adjacent underexplored area. Ten seismic attributes and DST data from 10 wells of the Mero Field were used to train shallow and deep-learning supervised regression algorithms for the prediction of flow capacity and productivity index seismic cubes. Twenty development wells (blind tests) were used for the assessment of our predictive models. The highest percentage of correct predictions at the blind test wells (85%) was obtained with random forest regression using six attributes derived from a spectrally balanced full-stack volume, and neither amplitude-variation-with-offset nor inversion data were needed. Deep learning provided lower performance (75%) at a higher computational cost. It demonstrated a new reservoir derisking tool that can be used for project optimization in areas covered by the same seismic survey.