Machine Learning Approach to Support Mud Gas Evaluation and Detect DBM in Oil Wells

Summary Real-time interpretation of mud gas data remains a major challenge in exploratory oil well drilling due to operational variability and the lack of robust models correlating gas signatures with reservoir fluids. This study presents a machine learning approach to identify geochemically relevant zones and mitigate artifacts from Drill Bit Metamorphism (DBM). The solution integrates drilling parameters, advanced gas measurements, mud properties, PVT lab data, and geological interpretations from Brazilian wells. Data preprocessing—performed via an integrated dashboard— included quality control, outlier removal, and feature selection using Multidimensional Scaling (MDS) and K- Means clustering. Kernel Ridge Regression was applied to produce two key outputs: gas affinity curves (C2 and C2C) and a DBM severity curve derived from ethylene (ETHE) distribution. Results showed strong model performance (R2 ≥ 0.84), effectively identifying productive zones and reducing the impact of drilling-induced anomalies. This AI-driven workflow supports early fluid evaluation by providing a probabilistic, data-informed interpretation of mud gas data. Its integration into surface logging enhances decision-making during drilling and reduces geological uncertainty in hydrocarbon exploration.