Machine Learning Based Mechanical Earth Model: A Case Study

Abstract

ABSTRACT: Determining the state of the in-situ stresses in a given field is crucial for proper completion selection, mud weight window determination, and hydraulic fracturing (HF) design and execution. Hassi Tarfa (HTF) field is tight sandstone reservoir in Algeria, where HF is the main stimulation technique being applied to increase oil recovery. However, determining rock mechanical properties such as stress tensor is challenging as most wells do not have sonic logs. This study seeks to apply different artificial intelligence (AI) techniques to predict the geomechanical properties of the HTF field using selected wells with and without sonic logs. The project is divided into three main sections. First, a complete set of logs for fifteen wells are used to train and validate several artificial intelligence algorithms that use depth, porosity, photoelectric and density logs to estimate sonic logs. The Root Mean Square Error (RMSE) is applied to rank the effectiveness of each method, and select the best model. The selected AI algorithm is then used to predict sonic measurements for the wells without acoustic data. In the second section, triaxial tests conducted on 23 core samples are used to train and test an Artificial Neural Network (ANN). The ANN predicts static Young’s Modulus and Unconfined Compressional Strength (UCS) from dynamic elastic moduli. In the third section, poro-elastic model is used to create a 1D Mechanical Earth Model (MEM) for the selected wells. The results showed that among all models, ANN is the most accurate model to synthetize sonic logs, and to predict the values of static Young’s modulus based on its corresponding dynamic value as well as UCS based on its relationship with dynamic Young’s modulus with an average absolute error of less than 2%. 1. INTRODUCTION A proper understanding of reservoir rock mechanical behavior is crucial in minimizing wellbore related problems such as wellbore collapse, reservoir subsidence and production of sand. This requires correct estimation of geomechanical parameters of the formation. The elastic properties include Young modulus, Poisson’s ratio, bulk and shear moduli and strength properties including uniaxial compressive strength (UCS), tensile strength, friction angle and cohesion (Parapuram et al., 2017). Accurate estimation of the geomechanical properties are vital in assuaging the risks linked with drilling and reservoir productivity enhancement (Tariq et al., 2017). In contrast, Abdulraheem et al. (2009) reported that wrong estimation of rock mechanical properties could result in incorrect field development planning and loss of revenue. Obtaining rock samples at many different depth and carrying out laboratory testing to estimate static geomechanical parameters using triaxial is tedious and expensive. The common approach is to use well log data, particularly sonic logs, to estimate dynamic rock mechanical parameters, yet this requires calibration against static values obtained from the core data at some depths. This is while there are instances where velocity logs are not available or unreliable (Abdulraheem et al., 2009). In such cases, the use of developed empirical correlations is a common approach to predict geomechanical parameters (Wadwa et. al., 2010; Najibi et al., 2015). The various correlations are bound to definite types of rock, a given depth, basin age as well as specific underground conditions and therefore they are not generalized for other fields.