Real-Time Compressional Sonic Log Prediction from Drilling and Mud Gas Data Using Machine Learning

Sonic logs are important for deriving elastic moduli of rocks, which can be useful in calculating in-situ stresses, estimating safe drilling mud weight, controlling wellbore stability, and constructing velocity models for seismic processing. Practically, determining the geomechanical information of the subsurface in real-time can alleviate operational risks and improve formation evaluation. Since sonic logs are not acquired in real-time, machine learning can be utilized to estimate them in real-time using drilling parameters and mud gas data. This study uses Random Forest machine learning technique to predict compressional wave slowness in real-time by utilizing surface drilling parameters and mud gas data. Out of a total of five wells, the regression model is trained with data from four wells. The input parameters for each depth point include conventional drilling parameters (rate of penetration, torque, weight on bit, etc.) and mud gas data. Various preprocessing techniques were applied on the input parameters prior to model training to ensure good quality. Validation was performed on wells with existing sonic logs that were not included in the training. Model performance is measured by the correlation coefficient and the mean absolute percentage error. Results show that predicting compressional sonic logs in real-time is feasible using machine learning. First, a model was tested to observe the effect of excluding certain depth intervals with high uncertainty in data values on model performance. The model's performance was enhanced and gave better correlation coefficient and lower mean absolute error. Therefore, cleaning data of uncertain intervals before running the model can improve sonic log prediction. Second, we investigated the effect of adding mug gas data as input features on model performance. Improvement in sonic log prediction was observed in some cases, and not in others. A sensitivity analysis was conducted on the developed models to determine the relative importance of the various input parameters on compressional sonic log prediction. Valuable information can be extracted from sonic logs in real-time to reduce operational risks associated with drilling. This study demonstrates the effectiveness of utilizing real-time data for compressional wave slowness prediction, saving significant time and cost. Through the use of machine learning, data cleaning, and model fitting, prediction can be automated. This allows for scaling up the analysis on all the data available. We plan to apply additional preprocessing techniques, include more wells, and perform feature selection on the data to improve the prediction accuracy.