Machine Learning Application to Permeability Prediction Using Log Core Measurements: A Realistic Workflow Application for ReservoirCharacterization

摘要

The use of Artificial Intelligence continues to grow in popularity within the geosciences in view of ever-growing complexity and magnitude of available subsurface data. This is equally evident by the need forfaster and accurate interpretations required to find hydrocarbons in ever more challenging and increasinglycomplex basins. This drive is made necessary in a continuously evolving and cost conscious petroleumindustry business environment. Advances in computing architecture now easily allows for more common application of machine learningtechniques in day to day geoscience workflows. The use of machine learning in permeability prediction isbecoming ever more common place as more specialists adopt this technique for modelling and predictionpurposes. Typical machine learning techniques include Fuzzy Logic, Artificial Neural Networks (ANN) andSelf Organizing Maps (SOM) amongst others which are run both in supervised and unsupervised modes.The described workflow in this paper was carried out using an available commercial standard petrophysicalpackage with ANN built in modules. This paper describes a typical workflow for predicting reservoirpermeability based on an integrated workflow utilizing core measurements integrated with available logdata. Permeability is a key rock parameter for understanding fluid flow dynamics and flow rates and itsmodelling usually poses some unique challenges. Traditionally and statistically, this can be done at afairly coarse level in cored wells by utilizing Poro-Perm correlations that usually do not capture fine scalevariability observed at core scale measurement. These Poro-Perm transforms are subsequently applied onuncored wells to predict permeability. This paper analyses a workflow that aims to utilize a depth-normalizedlog and core data set trained using an Artificial Neural Network (ANN) module, blind tested on few keycored wells and subsequently used to predict permeability in uncored wells. In conclusion, the recommendedworkflow will ensure much more realistic and better matching permeability predictions.