Reinventing Reservoir Characterisation and Properties Prediction through Core Based Machine Learning Application

Abstract

Advances in machine learning algorithm and the discoveries of various open source codes in recent years have led to a new insight in utilizing legacy data to delve into the unknown generating something out of poor data availability and presence. Subsurface evaluation of reservoir parameters in many ways are affected by uncertainties in the tool measurement, calibration accuracy and subjective narrative of geological processes involved. Often, difference in measurement scale among log, core, seismic acquisition and reservoir model construction will aggravate the uncertainties in petrophysical evaluation. This paper will highlight the novel approach of utilizing legacy core data from operator’s assets to predict and capture the uncertainties range of reservoir parameters used through the application of machine learning model. Several case studies demonstrating the accuracy of the prediction in multiple geological settings such as clastic, carbonate, CO2 storage site and deep water thin bed deposits will be also presented. The overall methodology can be divided into three phases where phase 1 involves data digitization into a structured format followed by data quality check based on industry standards and various rules applied to identify the outliers. Phase 2 involves features extraction from core images data utilizing deep learning network to evaluate core quality condition while phase 3 demonstrates the predictive capability of various machine learning and regression models to predict reservoir properties away from the wellbore and establish an acceptable uncertainty range of input parameters to be used in the reservoir modeling job. Predicted properties are validated and blind-tested with the newly acquired core data to further evaluate the accuracy and applicability of the machine learning approach. In addition, benchmarking against the producing reservoir analogs is also conducted to gauge the impact to hydrocarbon volume calculation and economic viability of the project. Results show that accuracy of properties prediction are acceptable for various reservoir depositional environment and comparison with producing reservoir analogs indicate 10% to 20% difference in the properties range which is acceptable considering the scarce data availability and overlap in the input parameters used in the machine learning model. Based on the established parameters range hydrocarbon volume in place, reserves and economics of the field development project could be further sensitized and evaluated prior to FID decision. In conclusion, machine learning algorithm proves its viability in predicting and quantifying the reservoir parameters uncertainties for a more holistic subsurface evaluation work. Application of legacy core data to generate valuable subsurface insights particularly in those limited data areas has assisted to delineate certain reservoir properties for certain basin, certain field and certain geological settings which serve the purpose of finding new exploration plays and monetizing the development opportunities.