Development of a New Empirical Formula Using Machine Learning for Pore Pressure Prediction in the South Sumatera Basin

Abstract

Accurate pore pressure prediction is crucial for maintaining wellbore stability and preventing drilling hazards. Therefore, this research aimed to present a new empirical method derived from machine learning models, applied to two wells in South Sumatra Basin (S-3 and S-4) comprising 214 depth intervals. The method integrated geomechanics principles, statistical correlation analysis, and neural network optimization to generate an interpretable and transferable equation. The internal parameters of the trained model were extracted and reformulated into a transparent empirical expression that engineers could apply directly in practice. This was distinct from the conventional black-box artificial neural network (ANN). Model performance was rigorously validated against analytical pore pressure measurements. Additionally, the method achieved strong predictive accuracy, with coefficients of determination (R²) of 0.94 and 0.91 for S-3 and for S-4, and root mean square error (RMSE) of 115 psi and 142 psi, respectively. These values represented a significant improvement compared to traditional methods. For example ANN-derived formula reduced RMSE by 28% and 22% in contrast to Eaton’s equation and the Bowers velocity–effective stress relationship. It also outperformed Normal Compaction Trendline (NCT) method in intervals with abrupt lithological changes. The clear identification of significant predictors, namelytemperature, gamma ray, porosity, and water saturation, helped bridges the gap between machine learning accuracy and engineering usability. The results showed that converting advanced computational models into interpretable tools significantly enhanced operational safety, reduced non-productive time, and improved drilling efficiency in Indonesian most prolific hydrocarbon provinces.