Advanced Computational Modeling of CO2 Solubility in Ionic Liquids Using an XAI Methodology
TL;DR
A high-fidelity Quantitative Structure-Property Relationship (QSPR) model is developed, realized through an Artificial Neural Network (ANN) topology, to forecast the degree of CO2 solubility across 14 different IL chemical families and a wide thermodynamic range, relevant for the oil and gas industry.
Abstract
The heightened global imperative for climate change mitigation necessitates the adoption of highly efficient carbon capture and storage (CCS) technologies. While Ionic Liquids (ILs) offer highly modifiable physicochemical properties for selective CO2 absorption, an accurate, high-performance prediction of CO2 solubility across diverse operating conditions remains a significant computational challenge. This research addresses this predictive constraint by developing a high-fidelity Quantitative Structure-Property Relationship (QSPR) model, realized through an Artificial Neural Network (ANN) topology, to forecast the degree of CO2 solubility across 14 different IL chemical families and a wide thermodynamic range (9.7–100120 KPa, 298–373 K). The developed QSPR model achieved exceptional predictive results (R2 = 0.989; MSE = 5.41e-4). Three innovations distinguish this work for real-world deployment: First, the model is presented in a "white-box" framework, which is critical for enhancing reproducibility and enabling swift integration into commercial software apps. Second, an explainable AI (XAI) methodology, utilizing Garson's algorithm, provided essential mechanistic insights, establishing dominance of the system's input features in the order: pressure > critical temperature > molecular weight > acentric factor > critical pressure > temperature. This analysis guides future solvent synthesis by confirming the primary thermodynamic control over the absorption process. Third, the model demonstrates an ultra-low computational footprint, requiring only 33 multiply-accumulate operations (MAC) and consuming 264 bytes of memory, making it highly suitable for low-latency, real-time prediction within digital twin environments. Given the critical role of CO2 solubility in solvent selection for carbon capture, this research is particularly relevant for the oil and gas industry, offering a tool to enhance natural gas quality by removing CO2, improving calorific value, and reducing pipeline corrosion.