Hybrid AI Framework for Predicting Geomechanical Risks and Hole Instability During Well Planning

Abstract Wellbore instability is a major contributor to Non-Productive Time (NPT) in drilling, leading to stuck pipe, sidetracks, and costly well abandonment. Conventional geomechanical assessments rely on deterministic models and expert inter- pretation, limiting scalability across complex formations. This study introduces a Hybrid AI Framework that integrates structured sensor analytics with unstructured historical text to predict geomechanical risks during well planning. The framework fuses Machine Learning (ML) models trained on Logging-While-Drilling (LWD) and real-time drilling pa- rameters with Large Language Model (LLM) extraction of instability events, causes, and mitigations from Daily Drilling Reports (DDR) and End-of-Well Reports (EOWR). A weighted Geomechanical Risk Index (GRI) combines ML-based probabilities (PML) and LLM-derived contextual similarity scores (PLLM) using tunable fusion weights α and β. Valida- tion on four Volve field wells (30,470 intervals; 17.4% event rate) using Leave-One-Well-Out cross-validation shows that Random Forest with GRI fusion achieves outstanding performance (F1 = 0.993, ROC-AUC = 1.000, PR-AUC = 0.999). Optimal weights (α = 0.95, β = 0.05) highlight ML dominance enhanced by contextual knowledge. Pressure-related parameters, rolling statistics, and torque fluctuations are identified as key predictors. The framework enables proactive, data-driven geomechanical risk assessment and supports intelligent well design within digital drilling systems.