Using Machine Learning for Geosteering During In-Seam Drilling

To maximize the coal seam gas production, it is critical to use geosteering to maintain the drill bit within coal seams. The gamma ray log is usually used as the coal/noncoal indicator to maintain the drill bit; however, the gamma ray log is a lagging indicator because its sensors are behind the drill bit. This can impair the drilling efficiency and subsequently increase non-productive time (NPT). In this paper, a machine learning approach is implemented to generate the gamma ray log (regression task) and identify coals (classification task) during drilling. The data is first filtered with positive rate of penetration (ROP) and depth increment. Then outliers are removed and samples are classified as coal/noncoal using the gamma ray log. The machine learning algorithm (i.e., XGBoost) is implemented to train and test the samples. To evaluate the results, the R², mean absolute error (MAE), and root mean square error (RMSE) are used for the regression task. The precision, recall, and F1 score are used for the classification task.

A case study is performed with data from one well in the Surat Basin, Australia. It is observed that ROP is usually higher in coals and lower in noncoal formations. The R² of the regression task from XGBoost is 0.6175. The MAE and RMSE are 1.293 counts per second (CPS) and 1.996 CPS, respectively. The general trend of generated gamma ray log is close to the original gamma ray log from logging-while-drilling (LWD) tools. For the coal/noncoal classification task, the precision, recall, and F1 score are 0.85, 0.88, and 0.86, respectively. Thus, XGBoost can effectively distinguish coals from noncoal formations during in-seam drilling. The developed machine learning model has the potential to identify coals and improve drilling efficiency during real-time in-seam drilling.