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

Abstract

This thesis presents an integrated methodology for predicting petrophysical properties (permeability K1 and porosity PHIE1 ) of oil reservoirs by combining geostatistics and ma-chine learning using Python. Based on real spatial data (file TPgroupe2. xls), an exploratory analysis and a logarithmic transformation (log K1 ) are performed. Spatial continuity is mo-deled using an experimental variogram (spherical model), followed by ordinary kriging to generate estimation maps with uncertainty. Machine learning models (Random Forest, XGBoost) capture nonlinear relationships but ignore spatial dependence. A hybrid approach combining the Random Forest trend with kri-ging of residuals is proposed, improving both predictive accuracy and spatial consistency. Sequential Gaussian Simulation (SGS) generates multiple equiprobable realizations, yielding P10 (pessimistic), P50 (median), P90 (optimistic) maps and a spread (P90 - P10) map, provi-ding rigorous quantification of spatial uncertainty for decision support in reservoir enginee-ring.