Identification of lithofacies from well log data in the upper Assam basin using machine learning techniques

Well logging can be classified under the general category of big data, as the datasets are intricate for conventional data processing application software to handle. This study aims at working on three entire wells drilled up to 1751–2690.4 m in upper Assam basin for lithofacies identification with each well comprising of huge datasets. Conventional methods for lithofacies classification are a challenging task for this enormous amount of data. It can be subject to biases, substantially laborious and time consuming. To deal with this problem, machine learning (ML) classification algorithms have come into picture, which systematically speed up accurate prediction of lithofacies. Five such algorithms (Support vector, Random forest, Multi-layer Perceptron, K Nearest Neighbors and Decision tree) are chosen and then evaluate each of the model’s performance exhibiting varying hyperparameters and input features. We established a thorough approach that entails all classifiers operating concurrently within the same ML framework. The best algorithm was selected following the comparison of each model’s Jaccard index and F1-score based on optimized hyperparameters. In addition, introduction of water saturation, a derived petrophysical property affects the model’s performance. This feature selection process brags the importance of inclusion of useful features to improve the accuracy of any ML model’s prediction capability. We observed that after introduction of water saturation profile, Jaccard index and F1-score of the best classification model, Random Forest has increased by on average 10.74% and 7.82%, respectively. Finally, we correlated identified lithofacies, including the reservoir facies from well to well in the study area.