Machine Learning Assisted Reservoir Properties Prediction in a Brownfield Offshore Malaysia

Abstract

Abstract Accurate prediction of porosity and permeability are crucial to the understanding of fluid distribution and hydrocarbon potential within targeted reservoir. However, increasing reservoir heterogeneity always possess a challenge to conventional method that simplify these complexities, while the high cost of coring acquisition makes it more difficult to validate the result. This paper describes an innovative technique for reservoir properties prediction that combines well logs, core analysis and machine learning tested at 5 wells located in a brownfield offshore Malaysia. Standard well logs of gamma ray, bulk density and neutron porosity were used as main inputs with routine core analysis as targeted outputs. The inputs data were checked and corrected for any light hydrocarbon or borehole washout effect that may affect the learning process. Due to nonlinear relationship between the variables, classification machine learning method of Random Forest was chosen for the study. The ratio of training and blind test was set at 70:30. The algorithm then tested at 100% of the data and benchmarked with core porosity, core permeability and human evaluation judgment based on well logs response at uncored interval. Both porosity and permeability prediction showed R2 of 85% and 80% during blind test which indicated high reliability on training model algorithm. In general, prediction result has a very good correlation with core data at good and thick sand. While at shaly sand, the correlation quality was reduced. At uncored interval, the prediction quality showed a realistic reading when comparing with well logs data response. This further proved the confident in the machine learning algorithm away from core interval. The R2 for porosity prediction was observed higher compare to permeability with average of possibly due to smaller scale ranges. The study showed another perspective to predict reservoir properties using a data driven approach and indicated a very promising result. No constant was used as compared to empirical calculation which introduced less uncertainty into the model. With reduce dependability in core data, the saved cost can be utilized in another multidiscipline task to increase the hydrocarbon recovery.