Improving Velocity Modeling with Surface Drilling Data Using Machine Learning

Abstract

Abstract Subsurface velocity models play a vital role in the conversion of seismic data from travel time to horizon depth. However, in practice, the calibration of velocity models can be challenging due to the limited availability of sonic logs. As a result, depth uncertainties may be introduced to the seismic images, potentially impacting the accuracy of geological interpretation. In this paper, we leveraged the known correlations between drilling parameters (e.g., rate of penetration, weight on bit and torque) and mechanical properties of formation rocks to construct a machine-learning model that estimates sonic velocity. Trained with well data from an unconventional field, the model successfully generated predictions that closely tracked actual sonic logs in blind test wells. The workflow consists of data preparation, exploratory data analysis (EDA), modeling and testing, and interpretation. The steps of data preparation and EDA were both conducted with the guidance of domain knowledge to ensure any preprocessing and interpretation are contextually reasonable. In the modeling phase, a two-stage Random Forest regression model was trained and fine-tuned, in which the second-stage model learns from the residuals from the first-stage predictions. Overall, the model demonstrated effectiveness by capturing the trend of sonic velocity with respect to depth across the majority of intervals. We observed that the prediction errors are overall symmetric around zero but varied with depth, which was likely caused by data quality issues in shallower intervals. Among the input features, true vertical depth made the dominant contributions to the prediction, likely due to the low dips and low lateral variability of the local geology. Other key predictors included revolutions per minute, hook load and weight on bit. By incorporating estimated sonic velocity profiles, the resolution of existing velocity models can be enhanced both vertically and spatially, while avoiding additional costs for data acquisition. In addition, the capability of estimating velocity during drilling offers new opportunities to inform real-time decisions in operations, as updates could potentially be made to the horizons ahead of the bit as new drilling data stream in.