Unlocking Reservoir Potential: Machine Learning-Driven Prediction of Reservoir Properties and Sweet Spots Identification

Abstract

Abstract Reservoir properties prediction and sweet spots identification from seismic and well data is an essential process of hydrocarbon exploration and production. This study aims to develop a robust and reliable approach to predict reservoir properties such as acoustic impedance and porosity of a fluvio-deltaic depositional system from 3D seismic and well data using Machine Learning techniques and compare the results with conventional stochastic inversion. A comprehensive machine learning methodology has been applied to predict reservoir properties in both log-to-log and log-to-seismic domains. First, 1D predictive models were created using an Ensemble modelling process which consists of 4 models each from Random Forest, XGBoost and Neural Networks. This was used to predict missing logs for eight wells. Subsequently, a 3D time model with 2ms temporal thickness was built and a seismic stack volume, seismic attributes volumes (envelope, sweetness, RMS Amplitude etc.) and low frequency model were resampled to the model resolution. The conventional post-stack stochastic inversion process is executed in the model to generate acoustic impedance, which is subsequently utilized to compute porosity through the acoustic impedance versus porosity transform. 3D predictive models are then created by incorporating seismic attributes, low frequency model and the target acoustic impedance log (AI) to establish a relationship and predict the 3D acoustic impedance property within the model. Additionally, another regression function is generated, employing the predicted acoustic impedance versus porosity, to forecast the 3D porosity property. Machine Learning 1D predictive models enabled the prediction of partial or full missing logs such as gamma ray, density, compression sonic, neutron porosity, acoustic impedance (AI), and porosity (PHIE) to complete the full logs coverage on eight wells in the reservoir zones. XGBoost 1D models produced the best results for training with R^2 score of 0.93 and validation score of 0.87. The stochastic inversion approach enabled the generation of high-resolution acoustic impedance and porosity properties in the 3D model. 3D predictive models established a relationship of seismic attributes volumes with well logs (AI) at well locations and predicted the acoustic impedance property in the whole 3D volumes away from the wells. To assess the prediction accuracy, we employed a randomly-selected blind wells approach, and the optimal model achieved an 82% validation accuracy. Notably, Neural Networks exhibited superior performance in proximity to the well locations, with a decline in quality observed as we moved away from the wells. On the other hand, Random Forest and XGBoost consistently produced continuous results. The predictive properties of AI and porosity were combined to train an unsupervised Neural Network model for facies prediction. This process aided in identifying sweet spots associated with the optimal reservoir sand saturated with hydrocarbons. Machine learning prediction produced quick and satisfactory results that are comparable with conventional seismic inversion output but with minimum intervention of an interpreter and demonstrated the ability to handle large datasets. The applied approach allows the generation of multiple models using various seismic attributes to identify the best sand reservoir sweet spots for well placement and field developments projects. This approach can be used at an early stage of exploration where few wells are available. The output reservoir properties can be directly included in a 3D static model.