Machine-Learning Assisted Reservoir Property Prediction: A Case Study from the Triassic Snadd and Kobbe Formations, Norwegian Barents Sea

Dimitrios Oikonomou1*, Behzad Alaei1, Eirik Larsen1, Christopher A-L. Jackson1,2, Idar A. Kjørlaug3, Kristian Helle3, Ryo Sakamoto3

1Earth Science Analytics As, Prof. Olav Hanssensvei 7A, 4021, Stavanger, Norway, 2Basins Research Group (BRG), Department of Earth Science and Engineering, Imperial College, LONDON, SW7 2BP, UK, 3Moeco Oil & Gas Norge AS, Haakon VIIs gate 1, 0161 Oslo

AAPG ACE, Salt Lake City, USA, May 2018

Prediction of reservoir quality and architecture are key steps of the E&P workflow. We have developed a machine-learning workflow for prediction of sedimentary facies associations, porosity, and permeability based on well- and 3D seismic data. The workflow is applied on a large dataset from the Norwegian Barents Sea.

The traditional workflow has two stages: i) petrophysical and sedimentological evaluation to derive rock- and fluid properties in wells, and ii) integration of rock physics and seismic quantitative interpretation to derive, and apply elastic to reservoir property transform functions to seismic data. The traditional approach is largely based on basin calibrated empirical relations between well logs and reservoir properties, adjusted using core data.

To improve workflow efficiency and accuracy, we replace the traditional approach with a machine-learning approach that: i) predicts reservoir properties directly from wireline logs and core data, and ii) predicts reservoir properties directly from elastic properties derived from 3D seismic. The machine-learning workflow is purely data driven, does not need manual calibration, and is thus more efficient and accurate than the traditional approach for large data sets.

First, we apply a series of machine learning regression and classification techniques to build models for porosity, permeability, and sedimentary facies based on elastic properties derived from well logs. We measure model accuracy and ability to generalize using various metrics, including blind tests.

Secondly, we infer reservoir architecture and quality in 3D based on inverted seismic data. The workflow is flexible, as it enables conversion of geological constraints to quantitative features that are used in combination with the elastic features.

In this talk we show how this workflow is applied to a case study of the Triassic of the Norwegian Barents Sea, using measured data from all available cores, wireline logs, and from two selected seismic datasets. We illustrate the workflow, by starting with regional raw data used for regional rock-property estimation, and ending up with prospect reservoir models for volumetric and fluid-flow simulation. We document, using blind tests, improved precision compared to the traditional approach. This, combined with improved workflow efficiency, makes our machine-learning approach attractive for all stages of the E&P workflow from regional screening to detailed reservoir characterization studies.