A Geomechanically Augmented Neural Network with Heterogeneity-Adaptive Data Splitting, Systematic Hyperparameter Optimization, and LSTM-FCNN Hybrid Architecture for Rate of Penetration (ROP) Prediction

The complex, heterogeneous nature of many subsurface environments makes accurate Rate of Penetration (ROP) prediction both critical and challenging for achieving drilling efficiency, cost control, and operational safety. Although artificial intelligence has demonstrated strong potential in extracting nonlinear patterns from drilling and well-log data, its application to heterogeneous formations remains limited by: (i) overreliance on operational parameters that lack formation-physics context, (ii) rigid train–test splits that ignore geological variability, and (iii) heuristic hyperparameter selection practices that are not reproducible. This study presents a geomechanically augmented deep learning framework applied to two vertical wells in a Middle East carbonate-clastic field (Well A: 9375 records, 1000–3370 m; Well B: 4443 records, 1945–3131 m). Five contributions are introduced: (1) a physics-informed input space integrating lithology-specific geomechanical properties (UCS, CCS, Young’s modulus, shear modulus, friction angle), validated against core measurements (R2 = 0.79–0.95); (2) a heterogeneity-adaptive train–test partitioning strategy demonstrating that formation complexity, rather than a fixed universal ratio, governs the optimal split; (3) a residual Fully Connected Neural Network (FCNN) with Swish activation and systematic hyperparameter sensitivity analysis; (4) a rigorous preprocessing pipeline comprising 99th-percentile Winsorization, interaction-term feature engineering (WOB × CCS, RPM × UCS), Lasso selection, Z-score normalization, and Gaussian noise augmentation, with all transforms fitted exclusively on training data to prevent leakage; and (5) a hybrid LSTM-FCNN that processes depth-ordered sequences via Savitzky–Golay denoising and a ten-step sliding window. The standalone FCNN achieved R2 = 0.8641 (Well A) and R2 = 0.9062 (Well B). The LSTM-FCNN improved intra-well accuracy to R2 = 0.9877 and R2 = 0.9551 and resolved a severe cross-well transfer asymmetry (B → A: R2 = 0.0388 for FCNN versus R2 = 0.8217 for LSTM-FCNN; A → B: R2 = 0.8963), confirming that depth-sequential modeling captures transferable formation patterns across contrasting lithological profiles.