Application Of Machine Learning In Reservoir Characterization Of The Tight Sandstone

Reservoir characterization is the process of accessing the viability of the reservoir with the determination of the physical reservoir parameters i.e.porosity,permeability,lithology,formation thickness,saturation,etc.Different data types i.e.seismic data,well-log data,and petrophysical core data are manipulated individually or integrated together to gain a better insight into these reservoir parameters.Similarly,approaches like empirical,and model-driven deterministic methods including soft computing methods like machine learning(ML)have been established to determine these parameters.Machine learning is a widely accepted and appreciated approach in reservoir prediction.Numerous machine learning algorithms are successfully employed and proven to be better performing than the conventional methods of reservoir parameter prediction.The subpar performance of conventional methods and their time and capital inefficienc y aspires to machine learning usage in reservoir characterization.In this study,reservoir parameters:S-wave velocity,porosity,permeability,and shale percentage are predicted using machine learning regression algorithms:Support vector regression(SVR),Random Forest(RF),and Deep neural network(DNN).S-wave velocity is predicted using well logs and porosity,permeability,and shale are predicted using both well logs and seismic data.This study makes use of the data from a tight sandstone reservoir in the Hangjinqi area of the Ordos basin in China.It establishes the applicability of ML algorithms for a heterogeneous reservoir with a comparative ML regression model analysis.S-wave velocity is predicted from well log attributes:acoustic(AC),density(DEN),and gamma-ray log(GR)with a good performance by all three models.However,SVR outperforms the remaining with an accuracy of 0.972 R~2scores.The model performance is validated with a blind well test as well.Porosity,Permeability,and Shale percentage were predicted with inputs P-impedance,S-impedance,V_p/V_s ratio,and density.All models performed well individually.But DNN performs better in the determination of all physical parameters.The correlation coefficient of the best-performing DNN model for porosity,permeability,and shale percentage is 0.828,0.808,and 0.933 respectively.Model SVR gives more correlation after DNN followed by algorithm RF.To have a wider insight and generalization of the reservoir our work is extended to the seismic data usage.P-impedance,S-impedance,and V_p/V_s are generated with simultaneous seismic pre-stack inversion.These inverted attributes are used as the input to the pre-trained ML models.Inverted porosity,permeability,and shale content visualized against the porosity,permeability,and shale of each well confirmed DNN’s better performance than SVR and ML.In this way fit for purpose ML models were trained and applied for tight sandstone reservoir parameters prediction.And its efficiency for the task is established.