Development of a Novel Physics-Informed Neural Network Architecture for Real-Time Pore Pressure Prediction During Drilling in Complex Pressure Regimes of the Deep Offshore Nigeria

Real-time pore pressure prediction during drilling in complex overpressure regimes involves the integration of heterogeneous sensor data streams with geological prior knowledge in ways that conventional statistical models cannot efficiently represent. This study develops a novel physics-informed neural network architecture specifically designed for real-time pore pressure prediction in complex offshore Nigeria deepwater wells. The network architecture embeds the Eaton pore pressure prediction equation and the Bowers unloading model as physical constraints within the loss function of a recurrent neural network, ensuring that predictions remain physically consistent with established geomechanical relationships while learning wellbore-specific corrections from drilling data. Training data comprising 28 deep wells from Nigerian deepwater blocks drilled between 2012 and 2022, including drill-while-drilling-calibrated pore pressure profiles, were used to construct the training dataset. A transfer learning approach is implemented to enable model initialisation from pre-trained weights and rapid fine-tuning on new well data in real time as drilling progresses. The physics-informed neural network outperforms both conventional Eaton-based prediction and a purely data-driven LSTM network in held-out test wells, reducing mean absolute pore pressure prediction error from 0.31 psi per foot for conventional methods to 0.14 psi per foot. The architecture enables confident real-time prediction 300 metres ahead of the bit using look-ahead seismic velocity inputs. Keywords: physics-informed neural network, pore pressure prediction, deepwater Nigeria, machine learning, real-time drilling.