📄 chemistry

Thermodynamics-Informed Graph Learning for Predicting CO2 Solubility in Water and Multicomponent Brines

This paper proposes a thermodynamics-informed graph learning framework to predict CO2 solubility in saline brines by modeling chemical interactions as graphs with physical constraints, demonstrating that while graph-based descriptors improve relative-error performance compared to traditional models, high-salinity extrapolation remains a significant challenge.

Original authors: Abu Bakker Siddique

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abu Bakker Siddique

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

Imagine you are trying to predict how much carbon dioxide (CO2) a glass of salty water can "drink" before it gets full. This isn't just a chemistry problem; it's a crucial puzzle for cleaning up our atmosphere. If we want to store CO2 underground in salty water pockets (aquifers) to stop global warming, we need to know exactly how much gas those pockets can hold.

This paper is like a team of scientists trying to build the ultimate "thirst calculator" for salty water using a mix of old-school math and new-school artificial intelligence.

Here is the story of their experiment, explained simply:

The Problem: The "Flat List" vs. The "Social Network"

For a long time, computers predicted this thirst by looking at a flat list of ingredients. They would say, "Okay, we have 5 grams of salt, 2 grams of another salt, and this much heat and pressure." The computer treats these ingredients like a grocery list where the items don't really talk to each other.

But in reality, a salty solution is more like a busy social network.

The salt ions (charged particles) are like people at a party.
Some are positive (cations), some are negative (anions).
They push and pull on each other.
The CO2 gas is a guest trying to squeeze in.
The water is the room they are all in.

The authors realized that treating the ingredients as a flat list misses the "party dynamics." They wanted to teach the computer to see the relationships between the ions, not just their names.

The Solution: A "Chemical Graph"

To fix this, the team built a Thermodynamics-Informed Graph Learning model. Think of this as giving the computer a map of the party instead of a grocery list.

The Map (The Graph): They turned every sample of water into a drawing (a graph).
- Nodes (Dots): Each dot is a character: CO2, Water, Sodium, Chloride, Calcium, etc.
- Edges (Lines): The lines connect the dots, showing who is interacting with whom. This captures how the ions push and pull on each other.
The Rules (Thermodynamics): They didn't just let the AI guess wildly. They gave it a set of "common sense" rules based on physics, like:
- Rule 1: If you squeeze the gas (increase pressure), the water should drink more.
- Rule 2: If you add more salt, the water should drink less (this is called the "salting-out" effect, like how adding salt to a crowded room makes it harder for new guests to enter).
- Rule 3: You can't have negative amounts of gas.

The Experiment: The Race

They tested their new "Graph Model" against the current champions of this field, which are powerful tree-based models (like XGBoost). Think of the tree models as very smart, fast calculators that are great at memorizing patterns but don't understand the "social dynamics" of the ions.

They ran the race using 3,323 real-world experiments collected from previous studies.

The Results: Who Won?

The results were a bit of a surprise and a bit of a reality check:

The Old Champion (Tree Model): The standard "flat list" model was incredibly strong. It got a score of 98.5% accuracy. It was the fastest and most accurate at predicting the exact numbers.
The New Graph Model (Without Rules): When they just used the "party map" without the physics rules, it did slightly worse on exact numbers but was slightly better at predicting the percentage of error.
The New Graph Model (With Rules): When they added the "common sense" physics rules, the model became very good at following the laws of nature (like knowing that more salt means less gas). However, it didn't beat the old champion on raw accuracy.

The Big Takeaway:
The new "Graph" approach didn't immediately crush the old methods. In fact, the old methods were still the most accurate for the data they had.

However, the paper argues that the Graph approach is like a better student for the future. While the old model is great at memorizing the test questions it has seen, the Graph model understands the principles of how the ions interact. This means it might be better at guessing the answer for a new type of salty water it has never seen before (like extremely salty water or a mix of salts no one has tested yet).

The Conclusion

The authors admit that for now, the "old way" (Tree models) is still the king of accuracy. But their new "Graph" method is a promising prototype. It successfully taught the computer to see the chemical "social network" and follow the rules of physics.

The paper concludes that to make this truly useful, they need to prove it can handle extreme conditions (like very high salt levels) better than the old models. Until then, it's a very strong proof-of-concept that shows how mixing chemistry rules with AI maps can help us understand the complex "party" happening inside a drop of salty water.

Technical Summary: Thermodynamics-Informed Graph Learning for Predicting CO2 Solubility

Problem Statement
Accurate prediction of CO2 solubility in saline water is critical for estimating dissolution trapping in carbon capture, utilization, and storage (CCUS) projects. While recent gradient-boosting models (e.g., LightGBM, XGBoost) have achieved high accuracy on compiled databases, they typically treat brine composition as a flat vector. This approach fails to explicitly represent the underlying chemical structure, such as ion-ion, ion-water, and CO2-brine interactions. Consequently, these models may struggle with extrapolation to unseen brine chemistries or high-salinity regimes where physical interactions are complex.

Methodology
The authors propose a thermodynamics-informed graph-learning workflow designed to encode the relational structure of chemical mixtures. The methodology consists of four primary components:

Graph Representation:
Each brine sample is converted into a chemical graph $G = (V, E, u)$ .
- Nodes ( $V$ ): Represent species including CO2, H2O, and specific ions (Na+, K+, Ca2+, Mg2+, Cl−, SO4 2−).
- Features: Node features include molality, ionic charge, radius, and hydration descriptors.
- Edges ( $E$ ): Connect active species to water/CO2 and ion-ion pairs, encoding attributes like charge products and interaction types (e.g., opposite charge, CO2-ion).
- Global State ( $u$ ): Includes temperature, pressure, ionic strength, and total molality.
Architecture (Preliminary Implementation):
Rather than a full message-passing Graph Neural Network (GNN), the preliminary study utilizes a ThermoGraphMLP. This architecture:
- Converts raw tabular inputs into 41 chemically meaningful graph descriptors (e.g., ionic strength, charge-balance error, cation/anion equivalent concentrations, and edge-summary descriptors).
- Feeds these standardized descriptors into a three-layer neural network.
- Trains in log-solubility space to balance the influence of high-solubility samples and align with relative-error metrics.
- Uses a softplus activation for the final output to guarantee non-negative solubility predictions.
Thermodynamics-Informed Objective:
The training loss function combines supervised learning with soft physical constraints:
- Supervised Loss ( $L_{sup}$ ): Mean squared error in log-space.
- Pressure Consistency ( $L_P$ ): Penalizes predictions where solubility decreases as pressure increases.
- Salting-Out Consistency ( $L_{salt}$ ): Penalizes predictions where solubility increases as ionic strength increases.
- Optional Regularization: Includes terms for positivity and potential residual learning against equation-of-state models.
Validation Protocol:
The study employs a stratified split (training/validation/test) and evaluates performance using $R^2$ , RMSE, and Average Absolute Relative Deviation (AARD). Crucially, it tests extrapolation capabilities via holdout sets for high salinity and specific salt systems, comparing the proposed model against tabular XGBoost and graph-descriptor XGBoost baselines.

Key Results
The workflow was tested on 3,323 cleaned experimental records extracted from a published database (Liu et al., 2026).

Tabular Baseline: Standard XGBoost on raw tabular inputs achieved $R^2 = 0.9856$ , RMSE = 0.00098, and AARD = 6.23%.
Graph-Descriptor XGBoost: Using the engineered graph descriptors without the neural network achieved the lowest relative error, with AARD = 5.90% ( $R^2 = 0.9840$ ).
ThermoGraphMLP: The physics-informed neural model achieved $R^2 = 0.9741$ and AARD = 6.83%. While it demonstrated competitive relative-error behavior, it did not outperform the boosted tree models in squared-error metrics ( $R^2$ or RMSE).
Physics Consistency: Diagnostics confirmed that the model adhered to expected qualitative trends: solubility increased with pressure and decreased with salinity (salting-out effect).
Extrapolation Limitations: The study noted that high-salinity extrapolation remains a significant challenge, with holdout tests showing high error rates (e.g., AARD of 41.6% for ExtraTrees in high-salinity holdouts).

Significance and Claims
The paper modestly frames its contribution as a theory-driven and reproducible architecture study rather than a definitive replacement for state-of-the-art tabular models.

Primary Finding: Graph-based chemical descriptors can improve relative-error behavior (lowering AARD) compared to raw tabular inputs, suggesting that encoding ionic interactions is beneficial.
Current Limitation: The current physics-informed neural model does not yet surpass strong tabular baselines (LightGBM/XGBoost) in overall accuracy.
Path Forward: The authors argue that for the work to be publishable as a major advance, future iterations must demonstrate superior performance in leave-one-source-out or leave-one-salt-system-out validation, and specifically improve extrapolation to high-salinity and mixed-brine regimes. The current results establish that the graph representation is a viable direction, but the full message-passing implementation and improved extrapolation capabilities are required to realize the full potential of thermodynamics-informed learning in this domain.