Modeling Petrophysical Property Variations in Reservoir Sand Bodies Using Artificial Neural Network and Object Based Techniques

Abstract

This study aims at reducing uncertainty in the prediction of petrophysical properties (porosity, water saturation and net to gross) of a field at locations that do not have well data by employing geostatistical simulations, artificial neural network and object facies modeling techniques in modeling the petrophysical property variations across a field. The project has an objective of establishing standard workflows that can be adopted in modeling petrophysical property variation in a field. Four classes of model were developed which are: the pure artificial neural network model, the pure variogram model, the collocated cokriging model and the bivariate cross plot model. The Gullfaks 3D seismic data, Checkshot data and Well log was used in implementing the project using the Petrel 2013 Software. Stratigraphic modeling, seismic interpretation, Depth conversion and 3D Grid Construction were done to provide an appropriate 3D Grid Structure for Petrophysical modeling. Facies and Petrophysical properties were interpreted at well locations inorder to provide data for modeling facies and petrophysical properties in 3D. The Truncated Gaussian Simulation was used in modeling the Gross Depositional Environment in order to provide an appropriate framework for 3D Facies modeling using Object based techniques. The Pure artificial neural network petrophysical models were generated using 3D Seismic data and petrophysical interpretations at well locations, the pure variogram models were developed using the Object facies model and petrophysical interpretations at well locations, the collocated cokriging models were generated by combining the variogram approach and the pure artificial neural network models using collocated cokriging and the bivariate cross plot models were generated by combining the variogram approach and the artificial neural network models using back propagation from their bivariate cross plot. The pure artificial neural network models gave the best prediction for porosity, net to gross and water saturation closely followed by the bivariate crossplot model for porosity and net to gross, the collocated cokriging model came second for water saturation prediction accuracy however, the pure artificial neural network model lacked geological patterns of facies which should naturally be present for porosity and net to gross, The pure variogram and collocated cokriging showed the region of oil water seperation more clearly than the artificial neural network method. Appropriate combination of these methods is therefore required for optimal Petrophysical property modeling.

SAMA: Spatially-Aware Model-Agnostic Machine Learning Framework for Geophysical Data

“…In addition to well-log data and core data, seismic data contain useful information that can be used to predict porosity and water saturation. Studies adapting machine learning approaches for earning from seismic data used models such as support vector Regression (SVR) [21], ANNs [22] and adaptive neuro-fuzzy inference system optimized by a genetic algorithm [18].…”