Neural network pore pressure prediction model with embedded physical constraints

• Addresses the limitations of traditional and conventional machine learning methods in utilizing multi-dimensional data and aligning with geological laws, which tend to produce geologically illogical predictions, by conducting research on neural network-based pore pressure prediction embedded with physical constraints. • Applies Spearman correlation analysis and SHAP value interpretation to screen and reduce dimensions of key features for pore pressure prediction, avoiding overfitting risks and computational waste caused by feature redundancy. • Constructs a loss function coupling data constraints and physical constraints (based on the Eaton pore pressure model), and uses the Bayesian optimization algorithm for global optimization of model hyperparameters. • Designs multi-scenario controlled experiments to quantitatively analyze the weight coefficient of data constraints, finding that the minimum mean absolute error (MAE) is achieved when the weight coefficient is 0.7. • In performance comparison experiments on the F well dataset, the Transformer model outperforms SVR, LSTM, and ANN models in prediction accuracy (with superior MAE, RMSE, etc.) due to its self-attention mechanism that advantages in correlating non-local geological features. • Verifies that the Transformer model embedded with physical constraints can effectively integrate multi-dimensional data (e.g., logging, seismic data), improve prediction accuracy, and enhance the interpretability of geological laws through attention weight visualization, providing a new technical path for safe optimization and intelligent decision-making in drilling engineering Accurate prediction of formation pore pressure is a core prerequisite for the safe and efficient execution of drilling engineering. Aiming at the shortcomings of traditional methods and conventional machine learning approaches in utilizing multi-dimensional data and aligning with geological laws—issues that tend to produce predictions violating geological logic—this study conducts research on neural network-based pore pressure prediction embedded with physical constraints. Based on 35 types of data from 6 wells in the BZ block, 8 core features were selected through Spearman correlation analysis and SHAP after preprocessing. A dual-coupled loss function integrating "data constraints + physical constraints" was constructed. Transformer hyperparameters were determined via Bayesian optimization, and the weight coefficient of data constraints was confirmed through multi-scenario control experiments. Comparisons were made with models without physical constraints and the traditional Eaton model. The results show that the Transformer-Phy model coupled with Eaton constraints performs optimally: in the test on Well F, its R² reaches 0.9821, with an error reduction of 27.34% compared to the unconstrained Transformer, outperforming other models. It has a maximum error of 9.79% and exhibits good cross-well adaptability. This achievement can optimize drilling fluid density design, reduce the risks of well blowouts and lost circulation, provide a reliable technical paradigm for continental clastic rock blocks, lay a foundation for multi-physics field coupling and promotion across multiple basins, and support the safe and efficient development of deep oil and gas resources.