Machine Learning-Driven Analysis of Geophysical Wellbore Data for Underground Storage Applications

Abstract This study explores the application of data-driven approaches to enhance the interpretation of geophysical wellbore data for carbon capture and storage (CCS) in the Illinois Basin. Extensive data from exploration and pilot-scale projects are analyzed to assess their potential for generating synthetic datasets that can effectively replicate some missing information for ongoing projects. Specifically, various machine learning (ML) models, including random forest, gradient boosting, feed-forward neural networks, and others, are employed to predict sonic velocities based on the mineralogical composition of participating rock and to identify fracture locations using petrophysical logs. ML-based approaches significantly outperform the traditional effective media interpretation, exhibiting improved accuracy in sonic velocity prediction and capturing small-scale heterogeneity of the formations. However, the prediction of fracture presence remains challenging due to data imbalance and the complex interplay between fractures and inherent heterogeneity of geologic formations. Application of t-distributed stochastic neighbor embedding (t-SNE)—an advanced technique for multidimensional data visualization—reveals that spatial heterogeneity strongly influences geophysical properties, limiting ML models’ performance in detecting fractures. Despite these challenges, ML approaches show promise in complementing traditional methodologies, enabling faster and more informed decision-making processes during early project stages. Future research is needed to address data limitations and enhance the reliability of ML models in diverse geological settings.