Scalable Geomechanical Modeling for CO2 Sequestration Using Progressive Neural Operators with Low-Rank Adaptation

The complexity of coupled fluid-solid interactions and the high computational cost of traditional simulations contest the accurate modeling of subsurface systems for geological CO2 sequestration. Machine learning-based reduced order models (ROMs) offer a scalable alternative but often require extensive training data. This study extends the Progressive Improved Neural Operator (p-INO) framework, which leverages progressive learning to enhance generalization and reduce data requirements in large-scale PDE simulations. To address the increasing parameter burden inherent to progressive learning, we introduce a LoRA-enhanced p-INO that employs low-rank adaptation to reduce computational costs while maintaining predictive accuracy. Results from datasets derived from the Illinois Basin Decatur Project demonstrate that p-INO significantly outperforms baseline models, particularly in data-limited scenarios, and that LoRA-enhanced p-INO achieves efficient parameter reduction with minimal accuracy loss. This combined approach provides a robust, scalable solution for geomechanical modeling in CO2 sequestration applications.