Physics-informed gaussian process regression for reproducible and uncertainty-aware CO2 injectivity prediction

Area of Science:

• Geological Engineering
• Machine Learning
• Chemical Engineering

Background:

• Accurate prediction of CO2 injectivity decline is crucial for safe geological carbon storage.
• Current machine learning models often lack uncertainty quantification, limiting their practical application.

Purpose of the Study:

• To develop a physics-informed Gaussian process regression (PC-GPR) framework for predicting Relative Injectivity Change (RIC).
• To embed physical constraints from DLVO theory and Civan-Kozeny-Carman models into the prediction framework.
• To provide uncertainty quantification for enhanced decision-making in carbon storage operations.

Main Methods:

• Developed four Gaussian Process (GP) variants, including a PC-GPR model.
• Integrated physical constraints (DLVO, Civan-Kozeny-Carman) into the GP framework.
• Employed Leave-One-Out cross-validation, k-fold cross-validation, and bootstrap confidence intervals for validation.
• Utilized GP posterior calibration (ECE) and split-conformal prediction intervals for uncertainty quantification.

Main Results:

• The GP-Base model demonstrated strong predictive performance (LOO R2=0.9401) with well-calibrated uncertainty (ECE=0.026) and reliable coverage (97.7%).
• The PC-GPR-M variant effectively enforced physical constraints, reducing DLVO monotonicity violations to 1.5%.
• Identified high-risk injection conditions (salinity >30,000 ppm, jamming ratio >0.04).

Conclusions:

• The PC-GPR framework offers an uncertainty-aware approach for predicting CO2 injectivity decline.
• This method enhances the safety and efficiency of geological carbon storage.
• Provides a baseline for future physics-informed machine learning research in subsurface energy applications.