Hybrid CNN-LSTM model for predicting wettability alterations in shale reservoir based on experimental techniques

Abstract

The study proposes a hybrid model from Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for the optimization of CO₂ sequestration in low clay or organic content, brittle, quartz-rich, siliceous shale reservoirs. The model solves the problem of wettability prediction under the low permeability of shale and complex fractures, which triggers high computational complexity and low generalization of conventional and existing machine learning methods. Using quartz as a proxy for siliceous shale, we train the model with CT scan data, SEM, high-speed camera images, and geomechanical measurements under supercritical CO₂ (SC-CO₂) and water injection conditions. SC-CO₂ injection increases quartz porosity by 4.0 % and reverses wettability to hydrophobic in order to expand storage capacity, but decreases caprock sealing. Water injection increases porosity by 2.8 %, doubles permeability, and promotes hydrophilic wettability to favor CO₂ trapping. The model achieves 92 % accuracy (F1-score 0.905 for hydraulic fractures) and low error rates (MSE 0.015 for porosity, 0.018 for wettability) with real-time CO₂ injection optimization possible. Validation in organic- and clay-rich reservoirs would be required for broader applicability, but for siliceous shales, the model is remarkably adequate.