Machine Learning Based Prediction of Porosity and Water Saturation from Varg Field Reservoir Well Logs

Abstract

Accurate estimation of reservoir parameters such as fluid saturations and porosity is important for assessing petroleum volumes, economics and decisionmaking. Such parameters are derived from interpretation of petrophysical logs or time-consuming, expensive core analyses. Not all wells are cored in a field, and the number of fully cored wells is limited. In this study, a time-efficient and economical method to estimate porosity, water saturation and hydrocarbon saturation is employed. Two Least Squares Support Vector Machine (LSSVM) machine learning models, optimized with Particle Swarm Optimization (PSO), were developed to predict these reservoir parameters, respectively. The models were developed based on data from five wells in the Varg field, Central North Sea, Norway where the data were randomized and split into an unseen fraction (10%) and a fraction used to train the models (90%). In addition to the unseen fraction, a sixth well from the Varg field was used to assess the models. The samples are mainly sandstone with different contents of shale, while fluids water, oil and gas were present. The ‘seen’ data were randomized into calibration, validation and testing sets during the model development. The petrophysical logs in the study were Gamma-ray, Self-potential, Acoustic, Neutron porosity, bulk density, caliper, deep resistivity, and medium resistivity. The log based inputs were made more linear (via log operations) when relevant and normalized to be more comparable in the algorithms. Feature selection was conducted to identify the most relevant petrophysical logs and remove those that are considered less relevant. Three and four of the eight logs were sufficient, to reach optimum performance of porosity and saturation prediction, respectively. Porosity was predicted with R2 = 0.79 and 0.70 on the model development set and unseen set, for saturation it was 0.71 and 0.61, a similar performance as on the training and testing sets at the development stage. The R2 was close to zero on the new well, although the predicted values were physical and within the observed data scatter range as the model development set. Possible improvements were identified in dataset preparation and feature selection to get more robust models.