Deep learning assisted monitoring framework for geological carbon sequestration

Abstract

Geological Carbon Sequestration (GCS) is expected to play a vital role in mitigating the harmful effects of climate change. However, the environmental risks associated with ${\text{CO}}_2$ plume migration, caprock fractures, induced seismicity, and fatal reactivation motivate the development of a high-resolution monitoring framework capable of predicting reservoir pressure build-up and ${\text{CO}}_2$ plume location in real-time. This study presents a Deep Learning (DL)-based monitoring framework that predicts reservoir pressure build-up and ${\text{CO}}_2$ plume location from sparse geophysical measurements during GCS injection operations. The proposed surrogate model consists of an Encoder–Decoder Artificial Neural Network (ANN) that inverts sparse surface uplift and injection pressure measurements to predict the pressure build-up field. A Long Short-Term Memory (LSTM)-based surrogate model is developed to capture the spatio-temporal evolution of the ${\text{CO}}_2$ saturation field, using the predicted pressure build-up history and ${\text{CO}}_2$ injection schedules. The synthetic dataset used is parameterized by variations in permeability, porosity, and dynamic ${\text{CO}}_2$ injection schedules. An ensembled eXplainable Artificial Intelligence (XAI) feature attribution approach is used to identify the most effective surface uplift sensor locations. The performance of the multi-network monitoring surrogate model is studied in the form of a case study involving the injection of ${\text{CO}}_2$ into a single wellbore. The accuracy of the proposed surrogate model demonstrates its potential for real-time monitoring of GCS processes.