Graph Machine Based-QSAR Approach for Modeling Thermodynamic Properties of Amines: Application to CO2Capture in Postcombustion

Abstract

Amine scrubbing is usually considered as the most efficient technology for CO<sub>2<sub/> mitigation through postcombustion Carbon Capture and Storage (CCS). However, optimization of the amine structure to improve the solvent properties requires to sample a large number of possible candidates and hence to gather a large amount of experimental data. In this context, the use of QSAR (Quantitative Structure Activity Relationship) statistical modeling is a powerful tool as it performs a mapping of a set of input vectors (i.e. the characteristics or the properties of the molecules under consideration) to a set of output vectors (i.e. their targeted properties). In this work, we used a high throughput screening experimental device to measure CO<sub>2<sub/> solubility data on a set of 46 amine aqueous solutions. Absorption isotherms are represented using a thermodynamic model based on two thermodynamic constants, pKa<sup>*<sup/> and pKc<sup>*<sup/> , accounting for the main chemical reactions occurring in the liquid phase between amine and CO<sub>2<sub/>. Then, we used a statistical approach named Graph Machines at the same time to cluster the molecules and to model the variation of the acidity constant pKa<sup>*<sup/> as a function of the molecular structure. The originality of our approach is the use of graphs to represent molecules in multidimensional spaces and simultaneously construct predictive models of their physicochemical properties based on these graphs. This approach is applied in this paper to predict the thermodynamic properties of a set of 5 new molecules.