Access Sproule's Special Market Report for Insights into Global Crude Markets Download

expertise iconContact An Expert

About Us

Sproule Presents a New Machine Learning Procedure for Generating Synthetic Logs in the Montney

Monday, March 2, 2020

Ilia Chaikine, Petroleum Engineer at Sproule and PhD 2020 candidate presented a study on the Montney at the Artificial Intelligence and Big Data Technical Meeting on March 2nd, 2020 in Houston, Texas. Ilia discussed a new machine learning procedure for generating synthetic logs in unconventional Reservoirs.

Presentation Abstract

During hydraulic fracture treatment, the rock mechanical properties surrounding a horizontal wellbore dictate how fractures propagate through the formation which heavily influences post-fracture production performance. Yet, despite the huge number of horizontal wells drilled and completed to date, there is still a poor understanding of the relationship between rock mechanics and production performance. The objective is to determine this relationship from data analytics, more specifically, convolutional-recurrent neural networks, by modelling the DTS log in tight rock formations.

Mechanical rock properties are estimated with three sets of log measurements: shear sonic travel time (DTS), compressional sonic travel time (DTP) and bulk density (RHOB). Due to cost and time to process. DTS logs are often missing. In this study, a hybrid convolutional-recurrent neural network (c-RNN) was developed to predict synthetic DTS curves. C-RNN have the advantage that they learn sequential data unlike traditional neural network (ANN) which do not have this capability. The synthetic DTS curve was generated by using five inputs: x, y and z coordinates, RHOB and DTP for every point along the wellbore.

This study focuses on the Montney formation in Alberta, Canada. Out of the 180 vertical and deviated wells in the study area, only 14 wells had DTS measurement available, thus the only suitable method for determining experiment accuracy was the “leave-one-out” cross-validation method. In this method, 13 wells were used as training data with one well as a test, the experiment was run a total of 14 times (one for each test well) and the results of all 14 experiments were examined and compared. Using the c-RNN the average mean absolute percent error (MAPE) along the entire Montney formation for the 14 wells came out to 1.2%. This result was superior when compared to that of a basic ANN, simple baselines, and empirical correlations. The results demonstrated that the new c-RNN method is a cost-effective and fast alternative to running DTS logs for new wells and can be applied to any formation as along as there is a sufficient number of existing wells, typically < 20, with DTS logs for verification.

The novelty of the research reported is on the use of hybrid convolutional-recurrent neural networks for modelling the DTS log in a tight rock formation. The approach is widely applicable to other tight rock resources and is relatively easy to implement.

To read more about the Artificial Intelligence and Big Data Technical Meeting, click here.