PIFLT: A Physics-Informed and Feature-Enhanced Hybrid Model for Formation Pore Pressure Prediction

Summary Pore pressure prediction is a critical focus in formation mechanics and petroleum engineering. Traditional methods, such as the Eaton method and the equivalent depth method, primarily rely on single-parameter empirical formulas to estimate pore pressure. However, these approaches are limited in their effectiveness, particularly when dealing with complex geological conditions and high-frequency data fluctuations. While deep learning techniques have demonstrated the ability to extract intricate feature relationships from data and generate data-driven pore pressure prediction models, they are often constrained by their “black-box” nature, resulting in limited interpretability and generalization capabilities. To address the aforementioned challenges, we present a deep learning framework that integrates physical knowledge into the existing long short-term memory (LSTM)-transformer fusion model. The proposed framework introduces a local feature enhancement mechanism within the LSTM structure, enabling it to more effectively capture localized data fluctuations. Additionally, physical constraints are strategically embedded within the transformer decoder, aligning the model training process with fundamental physical principles, thereby enhancing the interpretability and reliability of the predictive outcomes. The proposed model was thoroughly validated through extensive experimentation on multiple measured data sets. The experimental findings demonstrate that the proposed method significantly enhances prediction accuracy, offering an innovative approach to pore pressure prediction research.