When using wireline log to characterize formation properties for an area we often run into incomplete datasets. One way to address this lack of data is to create synthetic curves to use in the analysis. This paper will cover workflows to generate synthetic photoelectric (PE) and unconfined compressive strength (UCS) logs. Modern well log data sets usually include PE logs which provides important information about the lithology of the formations the wellbore intersects. However, many legacy wells do not have PE logs that we need in order to understand the lithology of these formations. Similarly, it is important to obtain UCS data from mechanical failure tests done to core samples to determine the strength of the rock. Obtaining and testing core samples for the entire zones of interest is both expensive and time consuming. Synthetic well logs can be a reliable and cheaper alternative to predicting PE and UCS values rather than running a new set of logs or coring and testing the whole zone of interest.
In the first workflow, Synthetic PE logs were generated using wireline logs from over a hundred wells that included gamma ray, density, neutron, resistivity logs and volume of clay. The data was randomly partitioned into a 70:30 split for training and validation data set respectively. Model competition among a suite of machine learning algorithms such as Linear Regression, Artificial Neural Networks (ANNs), Decision Trees, Gradient Boosting and Random Forest was used to select the best algorithm based on the least average squared error (ASE) of the validation dataset.
In the second workflow, UCS data was generated using wireline logs and core rebound hammer data from fourteen wells including gamma ray, density, porosity, neutron, clay volume, kerogen volume, compressional slowness, shear slowness, Young's modulus (static and dynamic) and Poison's ratio. Variable clustering was used to remove collinearity, decrease variable redundancy, and choose the best variables for analysis. Cluster analysis was performed on the chosen variables to identify factors that differentiate data segments from the population. The data was randomly portioned into a 70:30 training and validation split and model competition amongst the suite of machine learning algorithms mentioned above was used to select the champion model based on the least ASE of the validation dataset.
Results show that neural networks and random forests generated the best prediction of UCS and synthetic PE logs compared to other machine learning algorithms used.