Paramétrage de l’équation d’état PC-SAFT par apprentissage machine

The use of biomass appears as a promising alternative for the synthesis of a various families of strong value-added chemicals, which can be used in the manufacture of products. Product issued from the biomass are mainly composed of oxygenated molecules, but the models and methods traditionally used in the petrochemical industry have historically been developed to restore properties of hydrocarbons. The design of new manufacturing processes requires methods for rapid and accurate estimation of physicochemical properties of interest in the chemical industry.
Thermodynamic models - equation of state (EoS), in particular the EoS PC-SAFT (for “Perturbed Chain Statistical Associating Fluid Theory”) and its derivatives are widely used to restore the properties of fluids for targeted conditions of temperature and pressure. The application of this equation requires knowledge of parameters specific to the fluid to be studied. Group contributions have been developed to predict some of these parameters, thus highlighting the link between the structure of the considered fluids and the values of associated parameters.
In this PhD subject, we propose to implement methods based on data science to develop new approaches to parameterize EoS, and thus extend their fields of application. The envisaged combining between machine learning and thermodynamic models aims to improve the predictions of each of these models taken individually. Several approaches will be investigated: (i) approaches consisting in predicting the EoS parameters by means of QSPR (quantitative structure-property relationship) based models, and (ii) an approach based on the TGDS (Theory-guided data science).