Accurate estimation of shear and Stoneley wave transit times is fundamental for seismic analysis and reservoir characterization, particularly within the oil and gas sector. Traditionally, these parameters have been derived from Dipole Shear Sonic Imager (DSI) logs. Yet, challenges such as inconsistent data resulting from various geological conditions can hinder effective analysis. A new study addresses these issues by leveraging machine learning techniques to predict missing and unreliable values within DSI logs.
Conducted on data from two wells located in southern Iran, the research implemented eight distinct machine learning models: Random Forest (RF), Gradient Boosting (GB), Support Vector Regression (SVR), Multiple Linear Regression (MLR), Multivariate Polynomial Regression (MPR), CatBoost, LightGBM, and Artificial Neural Networks (ANN). By employing common log measurements such as computed gamma ray (CGR), bulk density (RHOB), and compressional wave transit time (DTC), along with depth-based lithologies, the study marks significant progress toward improving subsurface interpretation.
The research began with extensive preprocessing to remove outliers and noise and normalize the data. Invalid DTC values were corrected and utilized to predict the shear wave transit time (DTS) and Stoneley wave transit time (DTST). After this, results demonstrated Random Forest as the best performer compared to its counterparts.
Machine learning has emerged as a powerful tool for enhancing reservoir evaluation and characterizing elastic properties of formations. The methodologies adopted in this study exemplify how advanced computational techniques can facilitate the management of complex datasets, especially when traditional well-logging methods face limitations.
Throughout the study, Poisson's ratio—a parameter indicative of materials' elastic properties—was utilized to validate the DTC and DTS log values. Specifically, values were designated as missing at depths where Poisson's ratio fell below 0.15, adhering to established guidelines. Through this metric, the researchers ensured the accuracy and relevance of the DSI log assessments, lending greater credibility to their predictive models.
For the dataset comprising 11,218 data points, 80% were allocated for training the models and 20% for testing their performance. The final model validation was conducted using 10% of the dataset to confirm reliability through blind testing—evaluations conducted without prior exposure to the data set. Such rigorous methodologies are pivotal to producing reproducible and relevant outcomes across the oil and gas industry.
All algorithms entered this comparative framework, evaluated based on various metrics, including the coefficient of determination (R²), mean squared error (MSE), and root mean squared error (RMSE). Results revealed significant discrepancies among algorithmic performances. Random Forest consistently emerged superior across stages, showcasing its robustness and efficiency, particularly when managing invalid DSI data typical of complex geological formations.
For illustrative comparison, models such as Gradient Boosting and other machine learning forms also showed promise yet did not achieve the same reliability offered by Random Forest. This precise performance echoes a growing consensus within the scientific community on the efficacy of machine learning applications for subsurface management.
Importantly, the findings facilitate not only technical advances within the oil and gas sector but also reinforce the broader applicability of artificial intelligence across similarly challenging domains. By enhancing the predictability of sonic log parameters, this novel approach has the potential to reshape decision-making processes concerning reservoir management.
Looking forward, the researchers highlight the necessity for continued investigation, particularly with regards to integrating new features and exploring alternative machine learning techniques. Future efforts may pivot toward enhancing reservoir characterization to reinforce the methodological frameworks around DSI logs and their respective roles within oil and gas exploration.
Overall, the study strongly advocates for the utility of machine learning methodologies, illustrating their ability to solve complex analytical problems and enhancing the decision-making capabilities of resource management strategies.