Application of Blend and Stack Machine Learning Models to Predict Shear Wave Velocity Comparing to Greenberg-Castagna Method – A Case Study in One of South West Iranian Deep-Hard Carbonate Reservoir

Shear wave velocity is used in geophysical, petrophysical, and geomechanical investigations for a variety of purposes, namely drilling wellbore stability, sand output, and fracturing. Shear wave velocity is typically not present in all wells for a variety of reasons. In the current work, shear wave velocity in one of the challenging deep-hard carbonate oil reservoirs in southwestern Iran was estimated using the widely used correlation of Greenberg-Castagna modified by Gassmann in comparison to extensive machine learning approaches. Shear wave velocity (Vs) is proportional to the square root of Young s modulus and inversely proportional to Poisson s ratio. Factors such as rock properties and subsurface stress affect minimum horizontal stress. A 20% error in Vs log can significantly impact calculated values of Poisson s ratio and Young s modulus and minimum horizontal stress (δh,min). The impact depends on the context. Accurate measurement and interpretation of Vs and parameters are crucial in subsurface studies. Thus, with correct values of δh,min, fracturing design to break down formation could be performed. This paper presents a comprehensive study on predicting shear wave velocity (Vs) using machine learning techniques. A comparison between the traditional Greenberg-Castagna method and a machine learning approach was made on a TROF well data of deep hard carbonate reservoirs. Results showed that the R2 for Greenberg-Castagna modifying by Gassmann is 0.32, which is not scientifically and technically satisfying. On the other hand, a machine learning approach was applied to the data, resulting in the identification of six best-fit models, with Extreme Gradient Boosting as the best single learner with an R2 of 0.892. The original model outperformed the tuned model, with insufficient data and model limitations identified as the main reasons. To further improve model accuracy, stack and blend models were applied, resulting in R2 values of 0.9099, 0.9018, and 0.8826 for stack, tuned stack, and both blend and tuned blend respectively. The calculation time for 3613 data-points selected to calculate Vs using the Greenberg-Castagna method was approximately 11 minutes, while the calculation time for 4996 data-points selected to calculate Vs using the machine learning approach was approximately 20 minutes for each run. The data was split for training and testing purposes, with 3497 data-points used for modeling and 1499 data-points for predictions.