Physics-informed deep learning for prediction of CO₂ storage site response

Area of Science:

• Geological Engineering
• Computational Science
• Artificial Intelligence

Background:

• Accurate prediction of CO2 plume migration and pressure is crucial for safe and economical carbon storage.
• Traditional numerical simulations are computationally expensive, while data-driven models risk physical inconsistency.

Purpose of the Study:

• To develop a physics-informed deep learning method for predicting carbon storage site response.
• To improve the accuracy and reliability of CO2 injection modeling.

Main Methods:

• Utilized deep neural networks incorporating governing flow equations (conservation of mass and Darcy's law).
• Modified the loss term in deep learning models with physical constraints.
• Compared results against baseline data-driven models (MLP, LSTM) using a 3D synthetic dataset.

Main Results:

• Physics-informed deep learning significantly improved prediction accuracy compared to traditional data-driven models.
• The model effectively approximated CO2 saturation, pressure evolution, and water production rate.
• Demonstrated effectiveness even with limited training data.

Conclusions:

• Incorporating domain knowledge into deep learning enhances predictive accuracy for CO2 storage.
• The proposed method offers a reliable and efficient approach for CO2 storage management and risk assessment.