An intelligent method for predicting the escape characteristics of blowout leaks from offshore CO2 storage wells based on improved transformer and Bayesian-optimized SVR

Offshore CO 2 geological storage is crucial for carbon neutrality but poses wellbore leakage risks. This study provides a dataset of 10,270 offshore CO 2 blowout leakage conditions from coupled wellbore-seawater multiphase flow simulations. We propose a hybrid Transformer-BO-SVR model, integrating an improved Transformer with Support Vector Regression (SVR) optimized by Bayesian optimization (BO). Nine physical parameters (well depth, water depth, geothermal gradient, wellbore diameter, etc.) are used to predict seabed leakage mass flow rate, bottom hole pressure, and CO 2 sea-surface-reaching fraction. The model employs feature-wise independent embedding for continuous tabular variables, a gated residual module for selective feature filtering, and full-feature flattening followed by Bayesian-optimized SVR for prediction of three targets. Numerical analysis reveals that seawater plume evolution depends mainly on crossflow velocity and water depth, while CO 2 sea-surface-reaching fraction is governed by water depth and hole size. Sobol analysis identifies water depth as the primary controlling parameter. The model achieves high accuracy with R 2 of 0.981, 0.957, and 0.984 for the three targets, demonstrating its superiority in predicting the complex, multiphase flow characteristics of CO 2 blowout leakage in offshore storage systems.