📄 earth science

Predictive Modeling of CO₂ Wettability at Brine‑Mineral Interfaces as a Mechanistic Framework for Sustainable Carbon Storage

This study develops a robust predictive framework for CO₂ wettability at brine-mineral interfaces by evaluating fifteen machine learning models on 630 empirical observations, identifying CatBoost as the optimal algorithm with an R² of 0.9357 to enhance the security and efficiency of geological carbon storage through the identification of key governing parameters like pressure and mineral composition.

Original authors: Mohammadreza Kazemi

Published 2026-09-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammadreza Kazemi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Earth's atmosphere is warming, and a primary driver of this change is the accumulation of carbon dioxide released by human industry. To slow this trend, scientists are exploring ways to capture this gas and lock it away deep underground in porous rock formations known as saline aquifers. These are vast, sponge-like layers of rock filled with salty water, located miles beneath the surface. The goal is to pump the carbon dioxide down there, where it becomes trapped by the rock and the water, remaining isolated from the atmosphere for thousands of years. However, for this plan to work, engineers must be certain the gas will stay put. If the gas can easily move through the rock or escape, the storage site fails.

The key to keeping the gas trapped lies in a property called wettability. In simple terms, wettability describes which liquid a rock surface prefers to stick to: the salty water or the carbon dioxide. If the rock prefers the water, the water coats the rock grains, forming a protective barrier that stops the gas from moving freely. This is the ideal scenario for safe storage. If the rock prefers the gas, the gas can slip through the pores and potentially escape. This preference is measured by an angle formed where the gas, the water, and the rock meet. A small angle means the rock is water-loving and the gas is trapped; a large angle means the rock is gas-loving and the gas is free to move. Predicting this angle under the extreme heat and pressure of the deep Earth is incredibly difficult because the interactions between the rock, the water, and the gas are complex and change constantly.

Mohammadreza Kazemi, a researcher at the Petrochemical Research and Technology Company in Iran, set out to solve this prediction problem using a modern approach. Instead of relying on traditional equations that often struggle with such complexity, he turned to machine learning, a branch of artificial intelligence where computers learn patterns from data. He gathered a massive collection of 630 real-world measurements from previous scientific studies. These measurements recorded how carbon dioxide, brine, and different types of minerals—such as quartz, calcite, feldspar, and mica—interacted under various conditions of pressure, temperature, and saltiness. The challenge was to find a mathematical way to look at these conditions and accurately predict the resulting contact angle without needing to run expensive and time-consuming physical experiments every time.

To find the best method, Kazemi tested fifteen different computer algorithms, ranging from simple, straight-line models to complex, multi-layered networks that mimic the human brain. He trained these models on most of the data and then tested them on the remaining portion to see how well they could predict new, unseen scenarios. The results were clear: the simple models failed to capture the intricate, non-linear nature of the rock-fluid interactions, often missing the mark significantly. On the other hand, the most complex models sometimes learned the training data too perfectly, memorizing the noise rather than the underlying rules, which caused them to perform poorly when faced with new data.

One specific algorithm, known as CatBoost, emerged as the clear winner. It proved to be the most reliable tool for this task, achieving a level of accuracy that allowed it to predict the contact angle with remarkable precision. In the testing phase, the model's predictions aligned with the actual experimental data so closely that it captured 93.57 percent of the variations in the results, a performance far superior to the other methods. This high level of accuracy means the model can be trusted to forecast how the gas will behave in different underground environments without needing to physically test every single site.

Beyond just making accurate predictions, the study also used a technique to understand why the model made its decisions. This analysis revealed which factors mattered most. The most influential driver was the pressure of the system; as pressure increases, the behavior of the gas changes significantly. The second most important factor was the initial state of the rock surface, specifically a baseline measurement of how wet the rock was before the gas arrived. Surprisingly, the specific type of mineral called mica played a much larger role than other common minerals like quartz or feldspar. The presence of mica, a mineral with a layered structure, strongly influenced whether the rock would hold onto the water or let the gas take over.

The study concludes that this machine learning framework provides a robust and stable way to model these complex interactions. By using this tool, engineers can better select safe sites for carbon storage and design injection strategies that maximize the amount of gas trapped securely underground. While the current model is based on existing data and does not yet account for the slow chemical changes that happen over thousands of years, it represents a significant step forward. It offers a way to navigate the uncertainty of deep geological storage, ensuring that the carbon dioxide pumped into the Earth remains safely locked away, protecting the atmosphere for the future.

Technical Summary: Predictive Modeling of CO₂ Wettability at Brine‑Mineral Interfaces

Problem Statement
Accurately predicting the interfacial wettability of carbon dioxide (CO₂) in deep saline aquifers is a critical prerequisite for ensuring the structural and residual trapping security of geological carbon sequestration. While mechanisms such as structural and residual trapping are well-established, their effectiveness is highly sensitive to the equilibrium contact angle at the CO₂-brine-mineral interface. Traditional mathematical or strictly empirical estimations of this angle are often inadequate due to the inherent complexity and highly non-linear interactions among thermodynamic states (pressure, temperature), brine salinity, and heterogeneous mineralogical compositions. Furthermore, relying on extensive laboratory replications for every site-specific scenario is prohibitively resource-intensive. This study addresses the need for a robust, data-driven framework capable of forecasting contact angles with high fidelity across diverse geochemical environments without the computational expense of repeated physical experiments.

Methodology
The research employs a comprehensive, data-driven workflow centered on the aggregation and analysis of a curated dataset comprising 630 empirical observations of multiphase interfacial behaviors. The data spans varied mineralogical compositions (quartz, calcite, feldspar, and mica) and thermodynamic conditions.

Data Preprocessing and Analysis: The dataset includes independent variables such as operational pressure (MPa), temperature (K), brine salinity (M), mineral weight fractions, and baseline wetting metrics (static, advancing, receding angles, and $\theta_0$ ). The target variable is the equilibrium contact angle (degrees). The author conducted rigorous statistical analyses, including Monte Carlo-based outlier detection (retaining extreme values as physically realistic rather than erroneous) and Pearson correlation analysis to understand linear interdependencies.
Modeling Framework: Fifteen distinct machine learning algorithms were systematically trained and evaluated. These ranged from fundamental linear regressions (Linear Regression, Ridge, Lasso, Elastic Net) to advanced ensemble methods (Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost), kernel-based methods (SVR, Gaussian Process), and deep learning architectures (ANN, CNN, Decision Trees).
Validation Strategy: A rigorous five-fold cross-validation framework was applied to the training subset (80% of data) using a grid search for hyperparameter optimization. The remaining 20% was reserved strictly for testing to ensure an unbiased assessment of generalization capabilities.
Performance Metrics: Models were evaluated using the Coefficient of Determination ( $R^2$ ), Mean Squared Error (MSE), Mean Relative Deviation (MRD%), and Relative Error (RE%).
Interpretability: To ensure transparency and identify physical drivers, the study utilized SHAP (SHapley Additive exPlanations) analysis and feature importance metrics derived from Random Forest to interpret the optimal model's decision boundaries.

Key Results
The comparative evaluation revealed distinct performance tiers among the algorithms:

Linear Models: Traditional linear architectures (LR, Ridge, Lasso, Elastic Net) demonstrated significant underfitting, failing to capture the complex non-linear thermodynamic interactions, with testing $R^2$ values hovering around 0.48.
Overfitting Issues: Basic Decision Trees exhibited severe overfitting, achieving a near-perfect training $R^2$ of 0.9983 but collapsing to a testing $R^2$ of 0.7446.
Optimal Performance: The CatBoost algorithm emerged as the superior predictive framework. It achieved the highest testing $R^2$ of 0.9357 and a minimized testing MSE of 20.1358. Crucially, the performance differential between its training ( $R^2$ = 0.9877) and testing phases was narrow, indicating robust generalization and minimal overfitting compared to other high-performing models like XGBoost and ANN.
Visual Diagnostics: Taylor diagrams, parity plots, and residual distribution analyses (violin plots and density functions) confirmed that CatBoost maintained spatial concordance with empirical data and minimized systematic bias, accurately replicating the multi-modal statistical distribution of the contact angle.

Feature Importance and Drivers
SHAP interpretability analysis identified the paramount physical parameters governing the predictive output:

Operational Pressure: Identified as the most dominant thermodynamic driver, influencing CO₂ density and interfacial tension.
Baseline Wettability ( $\theta_0$ ): The initial wetting state was found to be a critical constraint on the final equilibrium angle.
Mica Weight Fraction: Among mineralogical components, the presence of mica (a phyllosilicate) showed the strongest positive correlation with contact angle, indicating a significant shift toward CO₂-wetness, whereas quartz showed a negative correlation (maintaining water-wetness).

Significance and Claims
The paper claims that the deployed machine learning methodology provides a mathematically stable and robust predictive framework for modeling complex interfacial wettability, overcoming the limitations of traditional empirical estimations. By integrating advanced computational algorithms, the study asserts that site-selection criteria can be optimized and injection strategies dynamically calibrated. This capability fundamentally enhances the macroscopic security of deep geological carbon storage operations by enabling reliable forecasting of residual and structural trapping potentials under diverse reservoir conditions.

The author maintains a modest scope regarding future implications, noting that the current methodology is constrained by the static nature of the aggregated tabular data, which may not fully capture long-term geochemical degradation or localized mineral dissolution. Consequently, the paper suggests that future work should focus on integrating time-series experimental data and coupled reactive transport modeling to adapt the framework to evolving subsurface mineralogical alterations, rather than claiming immediate deployment for dynamic, real-time field operations.

Technical Summary: Predictive Modeling of CO₂ Wettability at Brine‑Mineral Interfaces

More like this