Synthetic well log modeling with light gradient boosting machine for Assam-Arakan Basin, India

Abstract

Synthetic logs can help in the identification of petrophysical and geological properties of reservoirs, where the actual well logs are incomplete or absent. This study presents new and novel machine learning (ML) algorithms to generate pseudo well logs using the readily available conventional wireline logs. The machine learning algorithm works well with conventional wireline logs because they can deduce implicit relationships and efficiently detect trends in the logs. Moreover, identifying ups and downs (or peaks and valleys) of the subsurface characteristics is essential for any log. In this work, we have elucidated the application of Light Gradient Boosting Machine (LightGBM) and compared its performance with Category Boosting (CatBoost) and Random Forest (RForest). Grid search cross-validation (GridSearchCV) was done to optimize the hyperparameters of the algorithms towards a better balancing between accuracy and speed. To check the robustness of the algorithms, a number of input and output logs, considering variation in trends of logs with depth, were chosen to train the network. The well logs from Assam-Arakan Basin (India) include neutron porosity (NPHI) log, photoelectric (PE) log, caliper (CALI) log, density correction (DRHO) log, gamma-ray (GR) log, and bulk density (RHOB) log along with depth. We have evaluated the well log models and machine learning algorithms using determination coefficient (R2), mean square error (MSE), and root mean square error (RMSE) for training, testing, and validation. It has been shown that the use of machine-learning algorithms to generate synthetic logs improves reservoir characterization since they are computationally efficient, accurate, and cost-effective. The results showed that the LightGBM outperformed all the other algorithms in terms of speed and accuracy, in addition to its simplified hyperparameter tuning.