Deep Learning for Simultaneous Petrophysical Inversion of Pre-Stack Seismic Data: BiLSTM and CNN–BiLSTM Approaches for Direct Estimation of Porosity, Water Saturation, and Shale Volume

Abstract

This thesis presents a deep learning framework for the direct estimation of reservoir petrophysical properties, porosity (φ), water saturation (Sw), and shale volume (Vsh), from pre-stack seismic data. Conventional inversion workflows rely on linearized approximations and multi-stage processing, leading to uncertainty accumulation and poor recovery of lithological parameters such as Vsh. To overcome these limitations, Bidirectional Long Short-Term Memory (BiLSTM) and hybrid CNN–BiLSTM networks were developed for simultaneous prediction of φ, Sw, and Vsh directly from multi-angle seismic reflections. A synthetic dataset of 10,000 geological models was generated using Gassmann fluid substitution, Zoeppritz-based angle-dependent reflectivity at five incidence angles, and Ricker wavelet convolution, calibrated to North African clastic reservoirs. The CNN–BiLSTM architecture achieved the best performance, with RMSE values of 0.0149 for φ, 0.0639 for Sw, and 0.0461 for Vsh, outperforming classical Gauss–Newton inversion, particularly for Vsh estimation where the conventional method failed. Monte Carlo Dropout provided predictive uncertainty estimates, and the models remained highly robust to seismic noise across a wide signal-to-noise ratio range without retraining.