Pore Pressure Prediction by Empirical and Machine Learning Methods Using Conventional and Drilling Logs in Carbonate Rocks

Precise pore pressure estimation has high significance in terms of drilling and development operations. Regarding its necessity, empirical and intelligence methods have been introduced for prediction purposes. Efforts have been devoted to enhancing the accuracy of formation pressure interpretation and capable hybrid machine learning (ML) approaches. The present study addressed four drilled wells in the carbonate reservoir of Asmari in the Middle East. Due to the complexity of the pore pressure analysis in mentioned formations, some hybrid ML methods were proposed. The observer data for two of the wells were conventional logs and two other ones included conventional logs in addition to drilling logs. In the process, first, the datasets were organized for the case wells and the best features were indicated through combined artificial neural networks (ANN) and a multi-objective optimization technique. To apply the most effective parameters, the pore pressure of the Asmari reservoir was predicted using the least-square support vector machine (LSSVM), ANN, and random forest (RF) approaches. The particle swarm optimization (PSO) and Bayesian method were applied to increase the accuracy of the ML procedures. LSSVM-PSO and RF-Bayesian approaches showed the highest coefficient of determination, (on average, 0.97 and 0.96, respectively), as well as the least average absolute percentage error (AAPE) (1.3 and 1.4, respectively). Then, the ML methods were integrated with empirical formulas called Bowers, Zhang, and compressibility methods for pore pressure prediction whereas the outputs displayed superiority of the ML approaches. According to the outcomes, the proposed method outperformed hybrid approaches as a reliable and fast method for formation pressure estimation in the Asmari formation and other challenging carbonate reservoirs. Moreover, highly effective features were introduced for pore pressure prediction among drilling and well logs for such reservoirs which can prevent drilling instability and reduce non-productivity. The results can improve the reliability of pore pressure prediction and accuracies while providing a procedure through utilizing the appropriate estimation technique for carbonate reservoirs. Highlights Carbonate reservoirs require accurate pore pressure prediction due to complexity. Artificial neural networks with multi-objective optimization are effective for feature selection process. A hybrid of random forest and least-square support vecto