Single and Multi-Well Synthetic Well Log Generation using Multivariate Analysis

Abstract

Abstract With the rapid development in machine learning (ML) and deep learning (DL) algorithms, computations power and availability of the massive amount of data, ML and DL approaches have gained a lot of interest in oil and gas industry and brought the data science and analytics into the forefront of this industry. Using traditional neural networks (NNs) to normalize log and generate synthetic well logs is not a new idea. However, the recent advancement in ML and DL methods encourages to further research and revisit the prediction power, discuss the methodological limits and further improve the approach and prediction algorithms in single and multi-well synthetic well log generation for reservoir characterization and formation evaluation. In this study a data-driven procedure was implemented based on deep neural networks for density generation using multivariate inputs of well log data. The density prediction was formulated as a depth series regression problem where multiple inputs including different open hole logs are used as the input of a reverse model that estimates density. Different recurrent DL structures including Long Short Term Memory, Gated Recurrent Units, and Bidirectional Recurrent Neural Networks were tested in this study to select the most time-efficient and high performance model for density log generation. The models were fed with different number of available curves to explore a relationship and potential predictive power of several variables on the target curve (density) to eliminate unnecessary and irrelevant curves and end up with a model that uses a few curves to generate the target curve. Training datasets consist of real well log data of 25 wells logged offshore Middle East. Different metrics (RMSE, MAE, and R2) were used to compare the performance of the models and stacked LSTM of three layers showed the highest performance. Using this approach companies could cut costs by generating high quality synthetic data with a faster turnaround and without the need of relogging.