Leakage risk assessment in geologic carbon sequestration using a physics-aware ConvLSTM surrogate model

Abstract

The secure implementation of geological carbon sequestration (GCS) critically hinges on accurately localization of CO₂ leakage through inverse modeling of plume migration dynamics in heterogeneous reservoirs. This process is inherently challenged by subsurface uncertainties and the complexity of multiphase flow. Advances in various deep-learning-based surrogate models have been made to improve computational efficiency. Especially, Physics-Informed Neural Networks (PINNs) have gained widespread application due to their integration of the partial differential equation (PDE) into the loss function. However, conventional PINNs still face critical limitations in handling two-phase flow dynamics and high-dimensional parameter spaces due to the discretization requirements of PDE. To address these challenges, we propose a Physics-Aware Convolutional LSTM (PA-CLSTM) surrogate model that intrinsically embeds flow gradient information into the ConvLSTM architecture. Unlike PINNs which require PDE discretization as part of the loss function, PA-CLSTM encodes physical constraints through Sobel operator-derived velocity fields in latent space, thereby avoiding the need for PDE discretization, while maintaining compatibility with spatiotemporal feature extraction. Validation with a synthetic 2D saline aquifer demonstrate, PA-CLSTM achieves a five-fold acceleration over numerical simulations (TOUGH2-ECO2N) and a 67% reduction of inversion RMSE (from 1.65 to 0.59) of estimated permeability field in the focused area, compared to purely data-driven ConvLSTM. Meanwhile, PA-CLSTM inversion results accurately localize the CO₂ leakage. Compared to the ConvLSTM, the leakage location estimation RMSE decreased from 7.44 to 1.09, approaching the numerical simulation result of 0.68. In this work, we introduce the PA-CLSTM model in GCS, which significantly improves the inversion speed compared to numerical simulation and enhances accuracy compared to another surrogate model ConvLSTM.