Development of machine learning-based TOC estimation model using well logging and sidewall core data for source rock evaluation in the Jeju Basin, South Sea of Korea

초록

In this study, a machine learning-based model was developed to estimate total organic carbon (TOC) from well logging data for enhanced source rock evaluation. The model enables estimation of a continuous TOC curve along well logs, even in core-limited intervals. The dataset consisted of well logging data and TOC analyses of sidewall core samples obtained from the O, J3, and J5 wells in the Jeju Basin. After pairing well logging data with corresponding TOC values, input features were selected considering the number of available data pairs. Subsequently, a well-to-well standard normalization was performed to account for inter-well variability of the gamma-ray log responses. A total of 118 datasets were divided into 113 for train and 5 for test data. A performance comparison between random forest (RF) and extreme gradient boosting (XGBoost) models revealed that XGBoost demonstrated superior performance. Specifically, on the test dataset, XGBoost achieved a coefficient of determination (R-2) of 0.84 and a mean absolute error (MAE) of 0.09 wt.%, significantly outperforming RF (R-2 = 0.11, MAE = 0.19 wt.%). Application of the developed model to intervals within the O well where core data were unavailable revealed an underestimation in sections with TOC exceeding 1 wt.%. Future improvements within high-TOC intervals can be achieved through data augmentation or training with TOC experimental data using cutting samples.

키워드