Work

Physics-Constrained Deep Sequential Learning for Continuous Pore Pressure and Fracture Gradient Prediction in Heterogeneous Carbonate Reservoirs Using Conventional Wireline Logs

Year: 2026
Type: preprint
Abstract: Accurate assessment of predrilling and post-drill of formation pore pressure (Pp) and fracture gradient (FG) is quite important for defining a safe mud-weight window and avoiding kicks. Also, it is crucial for mitigating lost circulation and improving casing shoe depths. In Middle Eastern reservoirs, specifically heterogeneous carbonate reservoirs, traditional empirical correlations such as Eaton's sonic and resistivity equations as well as Bowers' velocity-stress relations and Matthews-Kelly fracture models often fail. Failure is attributed to carbonate rock stiffness, diagenetic cementation and secondary porosity beside chemical compaction decouple acoustic velocity from vertical effective stress (σ'v). Moreover, particular dipole sonic logs (Δts), borehole image logs and laboratory rock-mechanics core measurements are not often acquired across entire development wells because of extreme cost and operational hazards. However, using conventional wireline logs and as follow; Gamma Ray (GR); Bulk Density (ρb); Neutron Porosity (ϕN); Deep Resistivity (Rt); Caliper (CALI), and Compressional Sonic transit time (Δtc) is to overcome these limitations. This study suggests a novel work of Physics-Constrained Bidirectional Gated Recurrent Unit with Temporal Depth Attention (PC-BiGRU-Attn) framework. When well-log profiles are formulated as bidirectional stratigraphic depth sequences (L = 31, samples = 1.94 m aperture) instead of isolated depth points, the sequential network captures vertical compaction transitions beside caprock sealing effects and lithofacies boundaries across carbonate subunits. Both, Terzaghi's effective stress principle and geomechanical boundary constraints (Phyd ≤ Pp < Pf ≤ σv) are embedded directly into a hybrid physics-informed loss function to eradicate physically excluded pressure spikes in uncalibrated intervals. Field validation across three anonymized wells (WELL X1, WELL X2, and WELL X3, comprising 60,186 depth measurements spanning 1,247.12–3,392.62 m) demonstrates that the proposed PC-BiGRU-Attn architecture outperforms standalone empirical methods (Eaton, Bowers, and Matthews–Kelly), Multilayer Perceptrons (MLP), Gradient Boosted Trees (GBR), 1D Convolutional Neural Networks (1D-CNN), and standard Bidirectional Long Short-Term Memory (Bi-LSTM) networks. On the unseen blind test well (WELL X2, N = 8,946), PC-BiGRU-Attn achieves a coefficient of determination (R²) of 0.9876 (RMSE = 60.14 psi, MAPE = 1.30%) for pore pressure and R² = 0.9961 (RMSE = 32.39 psi, or FG RMSE = 0.0047 psi/ft, MAPE = 0.46%) for fracture pressure, reducing blind-well pore and fracture pressure RMSE by 37.2% and 49.6%, respectively, relative to point-wise ANN models while maintaining a 0.00% physical violation rate.
Language: English
Subfield: Fuel Technology
Field: Energy
Open Access status: green
OpenAlex ID: W7225905664
Cites: 19
Cited by:
Related to: 10