Abstract

Predicting pore pressure (PP) in oceanic slope settings, where sedimentation rates are relatively high, remains challenging and highly uncertain; however, deep learning (DL) offers a promising new approach. A key limitation of DL methods is their reliance on large volumes of training data, which is often scarce for deep reservoir wellbore logging. To overcome this challenge, we explore a transfer learning (TL) approach. Eaton's empirical model is applied to predict PP using wellbore logging data from the Northern Hikurangi Margin, New Zealand—an area influenced by gas hydrates, fluid migration, cold seeps, and creep-like deformation. The study utilizes 7,278 data records extracted from wellbore logging data of IODP Expedition 372, from well U1517A. Sonic transit time, gamma ray, bulk density, neutron porosity, and NMR porosity data were extracted from the well U1517A to construct the theoretical framework. The normal compaction trend (NCT) curve is used to evaluate the fitting accuracy for lowpermeability zone data. Statistical analysis using the Pearson correlation coefficient matrix with PP, to identify suitable input parameter combinations for developing the TL regression model. The dataset was divided into 80% for training and 20% for testing. The TL model achieved the highest accuracy in PP prediction, with an R² value of 0.759 and a root mean square error (RMSE) of 0.478. Additionally, the TL model incorporating past information reduced the RMSE by approximately 73% compared to the model without prior PP data. The model accurately predicts pore pressure and applies to other complex geological settings