Abstract: Formation pore pressure is the foundation of well plan, and it is related to the safety and efficiency of drilling operations in oil and gas development. However, the traditional method for predicting formation pore pressure involves applying post-drilling measurement data from nearby wells to the target well, which may not accurately reflect the formation pore pressure of the target well. In this paper, a novel method for predicting formation pore pressure ahead of the drill bit by embedding petrophysical theory into machine learning based on seismic and logging-while-drilling (LWD) data was proposed. Gated recurrent unit (GRU) and long short-term memory (LSTM) models were developed and validated using data from three wells in the Bohai Oilfield, and the Shapley additive explanations (SHAP) were utilized to visualize and interpret the models proposed in this study, thereby providing valuable insights into the relative importance and impact of input features. The results show that among the eight models trained in this study, almost all model prediction errors converge to 0.05 g/cm3, with the largest root mean square error (RMSE) being 0.03072 and the smallest RMSE being 0.008964. Moreover, continuously updating the model with the increasing training data during drilling operations can further improve accuracy. Compared to other approaches, this study accurately and precisely depicts formation pore pressure, while SHAP analysis guides effective model refinement and feature engineering strategies. This work underscores the potential of integrating advanced machine learning techniques with domain-specific knowledge to enhance predictive accuracy for petroleum engineering applications.