Real-Time Prediction of Acoustic Velocities While Drilling Vertical Complex Lithology Using AI Technique

Mechanical rock properties are often determined using sonic log data compressional velocity (V-p) and shear velocity (V-s). However, a sonic well log is not always acquired due to deteriorated hole condition (i.e., hole washout), sonic tool failures, especially in high-pressure, high-temperature (HPHT) wells, and relatively high cost. This paper introduces two data-driven models, namely artificial neural network (ANN) and random forest (RF), to estimate V and V across different formations that are characterized by deep burial depth and strong heterogeneity. Two types of actual field data were used to develop the models: (i) drilling surface parameters, which include flow rate, standpipe pressure, rotary speed, and surface torque, and (ii) acoustic velocities V and V, which were acquired by a conventional sonic log. Well-1 and Well-2 with data points of 6,846 were used to develop the models, while Well-3 with 1,016 data points was used to evaluate the capability of the developed models to generalize on an unseen data set with different statistical behavior. Furthermore, Well-3 was used to compare the accuracy of the developed models with the earliest published correlations in estimating the Vs. The results showed that the RF outperformed the optimized ANN in estimating V and V in Well-3. The RF predicted the Vp with a low average absolute percentage error (AAPE) of 0.9% and correlation of coefficient (R) of 0.87, while the AAPE and R were 6.7 % and 0.45 in the case of ANN Similarly, the RF estimated the Vs with an AAPE of 1.1% and R of 0.85, whereas the ANN predicted the V with an AAPE of 9.5% and R of 0.40. Furthermore, the RF was the most accurate in determining V in Well-3 compared to the earliest published correlations.

Automated summary

This study introduced two data-driven models, artificial neural network (ANN) and random forest (RF), to estimate sonic velocities across different formations. The results showed that the RF outperformed the optimized ANN in estimating the velocities, with lower average absolute percentage error and higher correlation coefficients.