Decoding the Stratigraphic Heterogeneity of Bengal Basin, India Using Supervised Machine Learning-A Case Study

Abstract

In oil and natural gas exploration, Machine Learning (ML) has gained noteworthy prominence for its ability to decode complex subsurface geology. ML commonly applies advanced statistical algorithms to build robust predictive regression and classification models. On-land Bengal Basin tasted exploratory success recently but displays great ordeal of stratigraphic heterogeneity. This paper discusses application of support vector machine (SVM), random forest (RF) and self-organizing map (SOM) ML algorithms with supervisions towards comprehensive modeling of the complex Miocene facies from Bengal on-land area for the maiden time, difficult otherwise by subjective conventional approach. Advanced geochemical logs (such as Elemental Capture Spectroscopy, ECS) and core data usually are very useful to quantify the facies downhole. However their availability also demands increased operation time and cost. Whenever available such data can be treated seamlessly with ML to test and build quantitative facies classification model from limited resources to over a region. Facies classification by ML not only makes the most of the available data but also eliminates the undesired subjectivity in addressing the subsurface heterogeneity with higher confidence. Gamma Ray, Neutron-Density, Resistivity, Sonic derived P-wave & S-wave velocities and suitably engineered log derivatives are mathematically modeled from the study area to classify facies using SVM, RF and SOM ML algorithms with ECS and core data calibration. Miocene sediments of the study area shows presence of six distinct facies viz. Claystone, Silty Claystone, Clayey Sand, Sandy Clay, Sand and Clean Sand. Facies data are trained by ML algorithms with multifold cross validation and returns credible accuracy for SVM, RF and SOM. The statistics driven facies model has been extrapolated for area where ECS or core data are not available but common logs are and yields geologically acceptable outputs. During exploration and field development stages such ML driven quantitative facies model improves the understanding of the subsurface from reservoir and non-reservoir point of view. ML Facies modeling captures the transition from shelf to fluvial depositional environment in the study area. Association frequency of different facies helps to visualize the changes from low/transitional to higher energy regime on a fine scale within Miocene. This paper discusses appropriate workflow, SVM kernel selection and hyper-parameter optimizations for SVM, RF and SOM that dictate the quality of facies model for Bengal basin. Heterogeneous stratigraphic play of Bengal on-land area demands accurate and quantitative subsurface lithological understanding for deploying fine exploration and development strategies, which can be addressed by this study. Nonetheless RF/SVM appear to be better facies classifier than SOM for Miocene sediments of Bengal from overall classification accuracy especially for less populous facies and calculation time.