ABSTRACT

This study investigates the feasibility of using machine learning algorithms to identify the presence of coal (at a scale comparable to wireline logging), utilising drilling parameters from coal seam gas (CSG) wells located in the Surat Basin, Queensland, Australia. Generally, during the drilling operation and before wireline logging, the presence of coal lithologies is inferred from elevated gas levels liberated into the drilling mud. However, the reliability of gas sensors can be compromised, necessitating operations geologists to rely on fluctuating drilling parameters to make decisions. To address this, a supervised classification model using the XGBoost algorithm has been developed to predict coal in the absence of reliable gas sensor data, improving data available to the operations geologists for decision-making purposes. Trained on data from over 150 wells, the classification model identifies coal lithologies by analysing typically available, high-resolution drilling parameters. While these parameters vary due to physical changes in the drilled lithology, they are also significantly overprinted by operational factors. Underpinned by iterative exploratory data analysis, the machine learning workflow involved processing a large amount of raw data, defining the predictive target, feature engineering and model development. Traditional machine learning performance metrics, such as F1 score, recall and precision, have been used in conjunction with business-based metrics to compare model iterations. Even though the model faces challenges related to class imbalances, overfitting and variable operational environments, results demonstrate the utility in predicting the location of coal and assisting operational geology workflows in wells where gas readings are unreliable or unavailable.