Integrating Machine Learning for Accurate Pore Pressure Prediction in Fault Conglomerate Formation of Junggar Basin.

Pore pressure stands as a foundation parameter for an optimal mud density window evaluation; thus, its accurate prediction plays an essential role in the success and safety of the well-drilling process. However, its prediction in complex geological formations remains one of the challenges, particularly in tectonic setting areas in which various complex geoprocess mechanisms drive the formation pressure system. Fracture structure, strong burial change, and strong formation heterogeneity of the Upper Wuerhe formation in the Jinlong 2 block form unique geological characteristics that impair the accuracy of parametric conventional prediction models. Integration of multiple petrophysical data is necessary to lower this prediction uncertainty. In this study, a new multivariate PP prediction model accounts for the influence of fracture presence, and the rock mechanical properties leveraging machine learning (ML) were proposed. Four ML models; deep neural network (DNN), support vector machine, XGB, and random forest were employed to relate the effective stress of formation with six well logs; neutron porosity, density, shale content, longitudinal acoustic velocity, borehole overgage, and resistivity anisotropy logs where the final formation PP was obtained based on Terzaghi's effective stress theorem. The study used 1680 data sets extracted from well logs from JIN 216 well, of which 80% were used as a training subset and 20% as a testing subset. All ML models have very good data set generalization; however, the DNN model presents the most accurate pore pressure prediction, delivered at an R (2) value of 0.9821 and root mean square error (RMSE) of 0.0344 g/cm(3). Furthermore, the ML shows about an 83% decrease in RMSE compared to that of Sayers and Eberhart Philips in both JLHW 204 and JLHW 261 wells found in the same block; hence, it saves as a potential alternative multivariate prediction model capable of improving PP prediction accuracy in fracture conglomerate formation.