Unlocking the Potential of Mud Logs and LWD in the Gulf of Thailand Using Machine Learning

Abstract The gas field in the Gulf of Thailand stands as one of the largest natural gas reserves in Thailand, with over 30 years of development history and more than 1,000 penetrated wells. With the power of machine learning and availability of historical data, there are opportunities to revolutionize traditional processes. Use of machine learning for log synthesis can help reduce expenses and operational risks associated with traditional well logging methods, including service fees, rig time, and potential retrieval challenges. This study aims to utilize machine learning techniques to predict well logs by analyzing Mud Log and Logging While Drilling (LWD) data. The proposed machine learning workflow utilizes raw mud log data (gas show content) and LWD measurements (gamma-ray and resistivity) to predict essential well logs—bulk density (RHOB), neutron porosity (NPHI), and compressional sonic travel time (DTC). This study employs various machine learning techniques, including Random Forest Regressor, Gradient Boosting Regressor Neural Networks and Support Vector Regression (SVR). The data sources come from the wells in the Gulf of Thailand, ranging from depths of 1,000 to 2,800 meters TVDSS. With a resolution of 0.1 meter, the training dataset will consist of 10,000 to 28,000 points for each well. These wells exhibit porosity levels between 13% to 30% and permeability below 100mD. The formations mainly composed of channel and point bar sands within a fluvio-deltaic system. Challenges arise from variations in reservoir properties across different depositional facies, clay content, and fluid presence, affecting prediction accuracy between original training data and predicted wells. The model effectively predicts well logs utilizing gamma-ray, resistivity, gas content data, and applies feature creation through polynomial transformation as input. Among the models tested, the Random Forest Regressor and Gradient Boost Regressor models prove to be the most accurate predictor. The random forest model achieves prediction accuracies with Mean Absolute Percentage Errors (MAPE) for bulk density (RHOB), neutron porosity (NPHI), and compressional sonic travel time (DTC) of 0.965%, 9.27%, and 2.536%, respectively. Such a predictive model is suitable for existing wells with only LWD and Mud Log data, lacking open-hole logs due to operational issues encountered during drilling. The cost savings from reducing well log fees and rig time in one of the fields could amount to $1 million USD per month. Beyond log synthesis, the methodology opens avenues for identifying reservoir and fluid types, enhancing decision-making in oil and gas exploration and production activities in the Gulf of Thailand.