Enhancing Rock Property Prediction from Conventional Well Logs Using Machine Learning Technique - Case Studies of Conventional and Unconventional Reservoirs

Abstract

Abstract It is timely for our industry to introspect on ways for step improvement in the utility of the wireline and logging-while-drilling logs which remains central to any asset development. Interpretation is limited by our current understanding of rock-fluid physics in source rocks, which is still developing. The gap is clearly evident in unconventional source rock interpretation where approximations such as pseudo-Archie approach are used for saturation estimation. This paper presents the use of emerging knowledge in machine learning to demonstrate its applicability for improving the total organic carbon (TOC) estimation in an unconventional well and permeability prediction in a conventional well. We have used the support-vector regression (SVR) technique, which is a new machine learning technique. Vast amount of logging data can be quickly processed using this technique. Limited core data is used to train the SVR algorithm. In this work, we first use the SVR technique to establish a correlation between conventional well logs (e.g., gamma ray, formation resistivity, neutron porosity, bulk density) and core measurements, thereafter building a rock property-prediction model as a function of well logs selected. Two field datasets from a South American well and a Mississippi Canyon well were selected to validate the method. Both wells contain a suite of logs and few core TOC and permeability data. Various combinations of conventional well logs were studied to check if the prediction accuracy can be improved. The results of the two case studies reveal that the SVR technique provides accurate and reliable TOC and permeability predictions. We observed in the case study of TOC prediction that incorporating the carbon weight fraction log as an input, in addition to the conventional well logs, improves the TOC prediction because the carbon weight fraction log provides information about the amount of carbon, which eventually helps the SVR algorithm to learn better from the data. Meanwhile, for the case study of permeability prediction, we observed that the conventional well logs are sufficient to generate a good permeability prediction model. Additional logging data including the nuclear-magnetic-resonance (NMR) logging data do not improve significantly the prediction accuracy. In conclusion, SVR technique could be used to improve our log interpretation. This technique can be easily adapted to predict rock mechanical properties and especially useful for unconventional reservoirs where traditional models may not be applicable and new methods are still evolving. Such new data analysis technologies could optimize our logging service and core analysis planning.