Monitoring and prediction of internal strains upon supercritical CO2 leakage along the fault

Monitoring internal fault parameters during CO 2 leakage provides critical insights into plume migration and leakage rates in subsurface environments. To explore this, a laboratory experiment simulating supercritical CO 2 leakage along a fault was conducted. Fifteen parameters were continuously recorded, including five internal strains on the fault surface and ten external variables (inlet pressure, five axial strains, and four hoop strains). Five regression methods – multiple linear regression (MLR), multivariate adaptive regression splines (MARS), ridge, Lasso, and Elastic Net regression – were applied to develop explicit models for predicting single internal strains. Additionally, back propagation neural networks (BPNN) were used to construct multi-input, multi-output models for simultaneous prediction of multiple internal strain components. Key findings include: (1) The curves of the 15 monitoring parameters change dramatically in the experiment, especially during the supercritical CO 2 transition phase. (2) All regression models achieved high accuracy in reproducing the experimentally observed strain evolution. (3) Among the tested approaches, MARS provided a closer fit to localized strain variations in the phase-transition interval than the linear regression models. (4) For BPNN, a network with two hidden neurons achieved a balance between predictive accuracy and model complexity for single-strain prediction. This study demonstrates that combining high-resolution experimental monitoring with machine learning can effectively capture the relationship between external stress–strain signals and internal fault responses.