Pore pressure prediction assisted by machine learning models combined with interpretations: A case study of an HTHP gas field, Yinggehai Basin

Accurately estimating pore pressure in high-temperature and high-pressure (HTHP) formations is crucial to prevent drilling problems and engineering disasters. In this study, four machine learning models were developed, namely decision tree, random forest, gradient boosting decision tree, and extreme boosting decision tree models, along with the classical Eaton's method, to rapidly predict pore pressure in an HTHP gas field in the Yinggehai Basin using seven variables, including 'depth' and six common well-logging series. The results demonstrate that the machine learning models outperform the classical Eaton's method in pore pressure prediction. Among the models, the decision tree model exhibits the best performance, with the highest correlation coefficient (R2 = 0.98) and the lowest root mean squared error (0.61). Additionally, the Graphviz tool and Shapley additive explanations method were used for the visualization and interpretation of the decision tree model to gain a deeper understanding of the model's working mechanism and the reasons behind the obtained results. The interpretation revealed that the variable of 'depth' was the most significant input parameter in the decision-making process of the model. This study provides insights into the 'black-box' interpretation of machine learning models and demonstrates that tree-based machine learning models can accurately predict pore pressure in undersurface HTHP formations.