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

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:This paper introduces a novel graph-based methodology that integrates thermodynamic constraints for CO2 solubility prediction, offering a new direction for ML in CCUS.

<title>Abstract</title> <p>Accurate prediction of CO2 solubility in saline water is essential for estimating dissolution trapping in carbon capture, utilization, and storage projects. Although recent gradient-boosting models achieve high accuracy on compiled CO2 solubility databases, most models treat brine composition as a flat vector and do not explicitly represent ion-ion, ion-water, and CO2-brine interactions. To address this limitation, this work presents a thermodynamics-informed graph learning workflow for predicting CO2 solubility in pure water, single-salt brines, and mixed brines. The model converts each brine into a chemical graph containing CO2, H2O, cations, and anions, and augments the graph descriptors with soft thermodynamic constraints for posi tivity, pressure consistency, salting-out behavior, and charge-aware composition. The workflow was tested on 3323 cleaned experimental records extracted from a published supplementary database. In a preliminary local run, tabular XGBoost gave the highest coefficient of determi nation, with R2 = 0.9856, RMSE = 0.00098, and AARD = 6.23%. Graph-descriptor XGBoost gave the lowest relative error, with AARD = 5.90%. The physics-informed graph neural model achieved R2 = 0.9741 and AARD = 6.83%, showing competitive relative-error performance but not yet outperforming boosted trees in squared-error metrics. These results show that graph based chemical descriptors can improve relative-error behavior, while high-salinity extrapolation remains the key challenge for a publishable final model.</p>

• Research Square https://doi.org/10.21203/rs.3.rs-10200370/v1first seen 2026-07-03 04:28:22