Modèles géostatistiques hybrides et Machine-learning pour la prédiction des propriétés pétrophysiques des réservoirs sous R

Abstract

This thesis aims to develop an integrated methodology to improve the modeling of petrophysical properties of oil reservoirs, particularly permeability and porosity, by combining geostatistical approaches (such as kriging) and Machine Learning techniques (such as Random Forest and XGBoost), primarily using the R environment with Python as a complementary tool. The adopted methodology includes several progressive steps : exploratory data analysis, spatial structure modeling using variogram, kriging interpolation, development of Machine Learning models, construction of a hybrid approach combining the advantages of both methods, and geostatistical simulation for uncertainty analysis and risk assessment. Results show that the variables, especially permeability, exhibit an asymmetric distribution and high spatial variability, requiring a logarithmic transformation. Variogram analysis confirmed the existence of a significant spatial structure. Machine Learning models, particularly Random Forest, outperform linear regression in predictive accuracy. The hybrid approach provided the best results, combining the predictive power of Machine Learning with the spatial consistency of kriging, producing more accurate and realistic maps. Geostatistical simulation helped identify high-uncertainty zones and assess risks through the calculation of P10 and P90 percentiles. This methodology can be applied to other fields such as hydrogeology, natural resource management, and environmental studies. It also opens perspectives for developing more advanced models such as deep neural networks and digital twins of oil reservoirs.