Accurate prediction of the rate of penetration (ROP) in deep and ultra-deep wells remains a major challenge due to complex downhole conditions and limited real-time data. To address the issues of physical inconsistency and weak generalization in conventional da-ta-driven approaches, this study proposes a mechanism-guided and attention-enhanced deep learning framework. In this framework, drilling physical principles such as energy balance are reformulated into differentiable constraint terms and directly incorporated in-to the loss function of deep neural networks, ensuring that model predictions strictly ad-here to drilling physics. Meanwhile, attention mechanisms are integrated to improve feature selection and temporal modeling: for tree-based models, we investigate their implicit attention to key parameters such as weight on bit (WOB) and torque; for sequential models, we design attention-enhanced architectures (e.g., LSTM and GRU) to capture long-term dependencies among drilling parameters. Validation on 49,284 samples from 11 deep and ultra-deep wells in China (depth range: 1226–8639 m) demonstrates that the synergy between mechanism constraints and attention mechanisms substantially improves ROP prediction accuracy. In blind-well tests, the proposed method achieves a mean absolute percentage error (MAPE) of 9.47% and an R² of 0.93, significantly outperforming traditional methods under complex deep-well conditions. This study provides reliable intelligent decision support for optimizing deep and ultra-deep well drilling operations. By improving prediction accuracy and enabling real-time anomaly detection, it enhances operational safety and efficiency while reducing drilling risks. The proposed approach offers high practical value for field applications and supports the intelligent development of the oil and gas industry.

Lithology identification in complex heterogeneous formations faces the challenges of overlapping well-log responses and long-tailed data distributions, causing traditional flat models to misclassify fine-grained lithologies. To address this, we propose a hierarchical routing method integrating geological priors and rock mechanics. A joint resampling strategy combining stratified undersampling and SMOTE first rectifies data bias. Next, conventional well logs are inverted into mechanical parameters, such as uniaxial compressive strength and abrasiveness, constructing a high-dimensional physical discriminative space. Finally, a two-stage network is deployed: an XGBoost gating model performs macroscopic rock classification to block cross-facies noise, followed by a dynamic routing mechanism distributing samples to deep neural network (DNN) sub-experts for fine-grained decoupling. Tested on actual multi-block logging data, the XGBoost-DNN architecture achieved a global accuracy of 0.85 and improved the macro F1-score from 0.55 under the baseline model to 0.84. Integrating physical mechanisms with a hierarchical topology effectively overcomes bottlenecks in fine-grained lithology identification, offering robust geological interpretability and engineering value.