Optimization of Drilling Parameters While Drilling Surface Holes Using Machine Learning and Differential Evolution

Summary Downhole vibrations while drilling surface hole sections can cause inefficient drilling. Downhole sensors can be used to provide real-time data on vibration levels encountered during drilling operations. This information helps the drilling crew to identify and address the factors causing excessive vibrations by adjusting drilling parameters based on real-time feedback to maintain or enhance the rate of penetration (ROP). The high cost, however, hinders the operator from using such sensors in each well. This research presents a workflow that coupled machine learning (ML) with an optimization algorithm to improve the drilling operation by enhancing the ROP while reducing the severity of downhole vibrations (i.e., lateral and torsional) without using downhole sensors. The ML modeling included multiclass-multioutput classification (MMC) to predict the severity of downhole vibration and regression analysis to predict the ROP. Different ML models, including K-nearest neighbors (K-NN), decision trees (DTs), random forest (RF), gradient boosting (GB), and extreme gradient boosting (XGBoost), were trained using data from eight historical wells drilled in a field of interest. The most accurate model was then combined with an optimization algorithm, differential evolution (DE), to optimize the drilling operation in Well No. 9. Four different optimization scenarios were explored to determine the optimal drilling parameters, surface rotary speed (RS) and weight on bit (WOB), to enhance the drilling efficiency. The values of RS and WOB parameters were varied within the traditional formation’s operational window, and a range of ±30%, 50%, and 70% of the original values applied during actual drilling in Well No. 9. The analysis showed that the RF was the most accurate model during the testing phase. The MMC achieved a Jaccard score of 0.83, while the regression achieved R2 and root mean square error (RMSE) values of 0.86 and 0.37, respectively. The results also revealed that all optimization scenarios were able to minimize downhole lateral and torsional vibrations almost across all drilled formations in Well No. 9. Moreover, none of the optimization scenarios resulted in a significant increase in the ROP in the uppermost drilled formation, except for a minor improvement observed in the top section. Scenarios 1 and 2 did not enhance the ROP in the lowermost drilled formations, while Scenarios 3 and 4 exhibited a higher improvement. The optimization workflow described in this paper demonstrates the potential for ROP enhancement while continuously monitoring downhole vibrations during drilling subsequent offset wells without the need to install downhole sensors, hence, reducing the overall cost of the well.