Banca de DEFESA: WILDSON BERNARDINO DE BRITO LIMA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : WILDSON BERNARDINO DE BRITO LIMA
DATE: 27/10/2023
TIME: 09:00
LOCAL: Via Google Meet
TITLE:

Use of Graph Neural Networks to predict the PC-SAFT pure-components parameters: evaluation of vapor pressure and density of non-associable and associable components

KEY WORDS:

Machine learning; Prediction of Thermodynamic Properties; Non-Associable Terms; Equation of State.

PAGES: 72
BIG AREA: Engenharias
AREA: Engenharia Química
SUBÁREA: Processos Industriais de Engenharia Química
SPECIALTY: Processos Bioquímicos
SUMMARY:

The modeling of thermodynamic properties is essential for the evaluation and optimization of processes through simulation tools. 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 non-associative, associative, polar and even ionic species. However, to use PC-SAFT, a first necessary step is to perform parameterization with experimental data, with data on saturated liquid density and vapor pressure being minimum prerequisites to ensure the equation of state produces robust results in an application, such as prediction of Liquid/Vapor and Liquid/Liquid Equilibrium. Within the context of bioprocesses, there is greater difficulty in obtaining these experimental data due to both physical and economic constraints. This is the case of ionic liquids and deep eutectic solvents that have attracted much attention from the scientific community in various processes, such as protein extraction. These compounds, however, have negligible vapor pressure, which is one of the properties that make them so attractive. Deep learning models, in turn, despite being unfeasible to model a wide variety of thermodynamic properties, as the equations of state do, are very robust in finding complex patterns that may exist, for example, between molecules and their pure-component parameters. In the present study, two deep learning models of the Graph Neural Networks type were developed to predict the PC-SAFT pure-component parameters, for the hard-chain and dispersion energy, from the molecule graphs, eliminating the experimental data prerequisite. The models demonstrated excellent performance on the test set, composed of non-associative, associative and ionic species. Model 1 presented an mean absolute percentage error of 10.96% and 25.94% for liquid and vapor density and vapor pressure, respectively. Model 2 performed better with 5.03% and 19.22%, respectively. The performance of the parameters predicted by the models was demonstrated with water, furfuranol, ethanol, 1-butyl-3-methylimidazolium bis(trifluoromethylsulfonyl)imide and 1-butyl-3-methylimidazolium tetrafluoroborate. The need to predict parameters related to the association and polarity of the PC-SAFT was also identified, in order to use the equation of state at its maximum power.

COMMITTEE MEMBERS:
Presidente - 1346198 - EVERALDO SILVINO DOS SANTOS
Interno - 1149554 - OSVALDO CHIAVONE FILHO
Externa ao Programa - 1224101 - VANJA MARIA DE FRANCA BEZERRA - UFRNExterno à Instituição - SAMUEL JORGE MARQUE CARTAXO - UFC