PC-SAFT PARAMETER MODELING WITH GRAPH NEURAL NETWORK (GNN)
Machine Learning; Thermodynamic modeling; PC-SAFT; Graph Neural Network.
The modeling of thermodynamic properties is essential to allow process optimization through simulation. In this regard, the equations of state have demonstrated great capacity in modeling the most diverse types of molecules. The Perturbed-Chain Statistical Associating Fluid Theory (PC-SAFT) equation of state is one of the most used models for this purpose, being capable of modeling polar and non-polar, associative and non-associative, and even ionic molecules. Deep Learning models, in turn, despite being unfeasible for modeling thermodynamic properties such as equations of state, are very robust in finding complex patterns. With this in mind, in the present study two Deep Learning models of the Graph Neural Network type were developed to predict PC-SAFT parameters from molecule graphs, eliminating the need for experimental data. The first model demonstrated excellent performance on simpler molecules, with an average percentage error of 2.27% for density calculation and 10.25% for vapor pressure calculation. Model 2 was developed targeting more complex molecules and presented an average percentage error of 2.50% for density and 4.88% for vapor pressure.