Drillability-informed ROP prediction and parameter optimization in complex formations

Rate of penetration (ROP) is a key indicator of drilling efficiency in complex formations and directly affects drilling time and cost. Under heterogeneous geological conditions, variations in formation drillability and mismatches between drilling parameters and formation characteristics can reduce rock-breaking efficiency and constrain ROP. This study develops a drillability-informed framework integrating formation characterization, ROP prediction, and drilling parameter optimization. Multi-source drilling and logging data are processed using wavelet denoising and normalization, and informative variables are selected using mutual information. Formation drillability is quantified using drillability level, brittleness index, and strength gradient index, and K-means clustering is employed to generate three engineering-oriented drillability categories rather than formal geological strata. A constrained fusion neural network, termed CTFusion-Net, is then constructed by combining physics-inspired computational unit features, parallel CNN–TCN feature extraction, and local monotonicity regularization within a quantitatively defined valid drilling envelope. On the independent test well, CTFusion-Net achieved an RMSE of 0.244 m/h, an MAE of 0.197 m/h, and an R 2 of 0.943. Relative to the CNN baseline, RMSE and MAE were reduced by 45.8% and 39.6%, respectively. Ablation analysis further showed that, when CU features were retained, monotonicity regularization reduced the monotonicity violation ratio from 13.1 to 5.8%, while providing a modest additional improvement in prediction accuracy. For parameter optimization, the Grasshopper Optimization Algorithm was used to optimize weight on bit and rotary speed within field-derived operational bounds. Across 25 independent runs, it achieved a model-predicted ROP improvement of 26.94 ± 0.63% and an MSE reduction of 12.36 ± 0.38%, outperforming Differential Evolution and Dual Annealing under the investigated conditions. The proposed workflow provides an integrated approach for characterizing drillability-dependent drilling responses and identifying promising parameter combinations. The reported improvements represent offline, model-based optimization potential and require further field validation and local recalibration before application to other formations.