Abstract

The objective of this paper is to highlight a new machine learning application to predict a synthetic caliper log using other available Logging While Drilling (LWD) well logs. LWD logs and wireline (WL) mechanical caliper logs for 17 tight gas wells, with more than 120,000 depth data points, were used to train a machine learning model. LWD logs included gamma ray, sonic, density, azimuthal density, porosity, photoelectric log, and mineralogy. The LWD logs were used as main input parameters in the model together with the maximum hole dimeter (Cmax) from wireline data. The model output being the synthetic maximum borehole diameter (Cmax_P) was compared with the measured caliper log. The model was trained with various machine learning algorithms including ensemble algorithms such as XGBoost and recurrent neural network (RNN). In the training process, Leave-One-Out method was implemented where 16 wells were used for training to predict the maximum borehole diameter in the remaining blind well. The results of current dataset showed good prediction accuracy with an average absolute percentage error of less than 6%. In addition, we demonstrated incorporating a physical feature related to geological layering by training the RNN with a window of multiple depth points, which clearly resulted in a predicted caliper curve that was significantly less noisy and had closer characteristics to the measured caliper curve.

Introduction

In multi-stage fracturing well completions, running a wireline mechanical caliper log is critical to obtain reliable measurements of borehole geometry along the lateral to identity packers’ placement depth for achieving adequate zonal isolation between successive fracturing stages. However, in many instances, wireline caliper logs cannot be run due to borehole conditions and drilling challenges. In such cases, various logging-while-drilling (LWD) well logs could be utilized to derive a synthetic caliper log.

This content is only available via PDF.
You can access this article if you purchase or spend a download.