Explainable AI-Based Geophysical Logging Parameters Selection for Enhancing Rock Brittleness Prediction
TL;DR
Based on XGB-based SHAP analysis, it is found that acoustic shear and compressional travel times are most contributing variables for BI forecasting other than deep resistivity, gamma ray, bulk density and neutron porosity.
Abstract
A reliable rock brittleness index (BI) profile plays a key role in maintaining safe drilling operation and hydraulic fracturing during hydrocarbon exploration and field development. Also, BI is vital for mining projects such as tunneling and rock mass excavation for mineral extraction. Instead of destructive testing measurements for BI, the machine learning (ML) algorithms are widely used to obtain continuous BI profile using real field logging data. The major objectives of this study are a) to evaluate the efficacy of different ML algorithms and b) to find the most contributing geophysical logging parameters for predicting BI of clastic sedimentary formation. The data-driven computational models are developed and optimized using ML techniques of extreme gradient boosting (XGB) and random forest (RF) algorithms along with support vector regression (SVR) using a list of processed 1550 datasets from logging parameters. The explainable AI (XAI) of Shapley Additive Explanations (SHAP) is applied to identify features of importance to ensure transparency and interpretability of the BI model's predictions. Results revealed that XGB outperforms with high determination of coefficient (99%) and least statistical error compared to RF and SVR algorithms to predict BI using logging data of offshore Volve oil field. Based on XGB-based SHAP analysis, it is found that acoustic shear and compressional travel times are most contributing variables for BI forecasting other than deep resistivity, gamma ray, bulk density and neutron porosity. These features importance also verified and confirmed by SHAP value with RF. The applied methodological strategies and research findings from the study could be applied to refine the data-driven model architectures for better performance and reliability in forecasting geomechanical properties and formation evaluation with precise subsurface datasets in petroleum and mining operations.