The chemistry of simple molecules can be extremely complex. Thermal decomposition, combustion, and pyrolysis processes often involve hundreds of intermediate species and hundreds of thousands of elementary reactions between those species. In the last decade, computer-aided kinetic modeling software packages have been developed to deal with those large kinetic models. They have demonstrated success in automatically developing kinetic models for the pyrolysis of renewable oils, combustion of biofuels, etc. The number of significant species and reactions in gas-phase kinetic models increases exponentially with the number of heavy atoms in the fuels. Today’s challenges for automatic kinetic modeling software are how to deal with detailed elementary-step kinetic models for molecules with more than ~6 heavy atoms, for molecules with heteroatoms and for surrogate mixtures. Besides searching for the most important reaction steps, a new challenge comes with assigning thermodynamic and kinetic parameters. 

 

These kinetic models contain too many thermodynamic and kinetic parameters (e.g. k’s, Keq’s) to determine all of them experimentally. Instead, fast calculation of those parameters is done automatically using structure-activity relationships. The few thermodynamic and kinetic parameters that are most sensitive towards the concentration of certain desired products are typically refined with high-level theoretical calculations or experimental measurements. With this procedure one can achieve high accuracy in modeling the combustion or pyrolysis of simple fuels. However, when applied to fuels with unique molecular structures or processes occurring in a different phase conventional methods are lacking. Recent advances in machine learning for applications in chemical engineering have opened a new route for the fast and more accurate prediction of such chemical properties. This majority of this talk will center around the application of machine learning using message passing neural networks for the fast predictions of thermodynamic and kinetic properties in the framework of computer-aided chemical kinetic model development. More specifically, this talk will focus on the use of transfer learning and hybrid modeling that combine quantum chemistry, thermodynamic state functions, and experimental measurements for the prediction of solubility limits.