Physics-Informed Machine Learning for Real-Time Formation Characterization Using Surface Drilling and Mud-Gas Data
Abstract
Conventional wireline logging provides the industry standard for quantitative petrophysical characterization and remains essential for reservoir definition and completion design. This study investigates whether real-time surface drilling parameters recorded during the drilling of vertical well ASTT-6-Deep (4,094 m MD) can supplement wireline evaluation by providing continuous, real-time lithology screening. The dataset comprises 4,069 depth intervals at 1-m spacing, of which 3,538 carry expert mud-logger-assigned lithology labels across six classes: Shale, Claystone, Dolomite, Limestone, Sandstone, and Silt. Fifty-five features are engineered from 42 active mud log channels across four categories: (A) seven mechanical features, including Mechanical Specific Energy (MSE) and the novel Lithology Signature Index (LSI = Torque/WOB ÷ ROP); (B) ten gas-chromatography ratio features; (C) six hybrid composite features; and (D) twelve five-point rolling-window depth-trend statistics. Random Forest and XGBoost classifiers are evaluated under 5-fold stratified cross-validation with within-fold SMOTE oversampling. Both models achieve a Macro-F1 of 0.6523 ± 0.011 (Random Forest) and 0.6514 ± 0.010 (XGBoost), with overall accuracy of 73.1% and 73.7%, respectively. Dolomite is the most reliably predicted class (F1 = 0.8951/0.8967), while Silt is the most challenging (F1 ≈ 0.33). The authors explicitly acknowledge that these accuracy levels are insufficient to replace wireline formation evaluation. The framework is positioned as a real-time complementary screening tool that provides continuous formation context and guides wireline decisions, not as a wireline substitute. The results compare favorably with recently published studies on machine-learning-based lithology prediction from drilling and mud-gas parameters (Ibrahim et al. 2023; Ziadat et al. 2023; Mensah 2023; Khalifa et al. 2023; Nguyen et al. 2026).