A machine learning methodology for multivariate pore-pressure prediction

Accurate pore-pressure prediction is of essential importance to hydrocarbon exploration and development. A multivariate prediction model of multiple petrophysical data is required to adequately reflect variations in pore pressure. However, the parametric multivariate models with assumptions on lithology, predominantly sand or shale, are theoretically inaccurate for mixed lithologies and require a tedious calibration process. Here, we propose a new method of utilizing machine learning (ML) techniques for pore-pressure prediction with a nonparametric multivariate model of petrophysical properties (sonic velocity, porosity, and shale volume). The training dataset for the ML models is constructed using petrophysical properties extracted from well log data and theoretical effective stress in the normally compacted interval. Bowers' unloading relation is invoked herein to account for abnormal pressure generated by unloading. Four ML algorithms, including the multilayer perceptron neural network, support vector machine, random forest, and gradient boosting machine, are applied to well log data from a set of offshore exploration wells. The results suggest that the proposed method using ML makes pore-pressure predictions in good agreement with pore-pressure measurement, and random forest outperforms the other ML algorithms in terms of goodness-of-fit, generalizability, and prediction accuracy. Compared with methods based on parametric models, the proposed method based on ML produces more accurate pore-pressure prediction and better capture the onset of overpressure.