The optimization of various chemical processes, such as fuel combustion in engines and pyrolysis of hydrocarbons in a steam cracker, relies heavily on highly detailed kinetic models. Such models require accurate kinetic and thermodynamic data. Ideally, these are provided by high-level-of-theory <i>ab initio</i> calculations. However, for the time being, performing such calculations “on-the-fly” is prohibitively expensive. Combining existing data with group additivity is the current state-of the art for thermodynamic and kinetic property estimation. One major disadvantage of this approach is the large amount of expert-user intervention. Essentially, the task of this expert-user is recognizing the important (functional) groups that constitute the molecule and that contribute to its thermodynamic behavior.<br/>Artificial neural networks have been successfully applied to several problems involving similar pattern- and element recognition tasks, especially in image processing and natural language processing. As neural networks have also been frequently applied in several QSPR problems, they are seen as a promising alternative for group additivity.<br/>In this work, a generally applicable artificial neural network approach is presented. It allows the estimation of various thermodynamic properties such as enthalpy- and entropy of formation and specific heat capacities at different temperatures. The neural network is trained on a large number of C,H,N,O-containing hydrocarbons. While the current results do not yet achieve the desired chemical accuracy of group additive methods, the trends are captured well. With some further improvements on the molecular representation in the input, this approach can become competitive with group additivity as it is applicable to a wider range of hydrocarbons than any individual group additive method.

Artificial neural network-based approach to thermodynamic property estimation

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.