(37g) Message Passing Attention Networks for Reaction Rate Estimation | AIChE

(37g) Message Passing Attention Networks for Reaction Rate Estimation


Sirumalla, S. K. - Presenter, Northeastern University
West, R. H., Northeastern University
Complex chemical systems like combustion of hydrocarbons, halogenated hydrocarbons (HHC), and heterogeneous catalysis, are well understood by detailed kinetic models. Detailed kinetic models differ from global kinetic models by including thousands of elementary reactions between hundreds of intermediate species. Building detailed kinetic models by hand is cumbersome, which led to development of Reaction Mechanism Generator (RMG) [1], an automatic kinetic model generation framework. Currently, thermokinetic parameters characterizing the species and reactions are quickly estimated using empirical estimation methods like Benson’s group additivity (GA) [2] during the model building process. But GA is known to poorly estimate the thermokinetic parameters of halogenated hydrocarbons [3] impeding the extension of RMG to HHCs.

Over the past few years machine learning, in particular deep learning, has outperformed several empirical methods in chemical property estimation. In this talk, we present a novel deep learning method, Message Passing Attention Networks (MPAN) [4,5], that directly operate on attributed chemical graphs to estimate thermokinetic parameters. MPAN consists of three phases, a message passing phase where the information is aggregated, an update phase based on attention mechanism to capture the important information at the graph level, and a readout phase which produces a single fixed length real valued vector that represents the graph in whole which can be described as molecular fingerprint. Molecular fingerprints are then used to estimate thermokinetic parameters i.e. enthalpy, entropy, heat capacities for intermediate species and reaction rate coefficients for elementary reactions. We trained the MPAN model with high level quantum chemistry calculations with various targets of interest. We achieved mean absolute error of 4 kcal/mol for reaction barrier height estimation of H abstraction reaction family.


  1. Gao, C. W., Allen, J. W., Green, W. H., & West, R. H. (2016). Reaction Mechanism Generator: Automatic construction of chemical kinetic mechanisms. Computer Physics Communications, 203, 212-225
  2. Benson, S. W., Cruickshank, F. R., Golden, D. M., Haugen, G. R., O'Neal, H. E., Rodgers, A. S., & Walsh, R. (1969). Additivity rules for the estimation of thermochemical properties. Chemical Reviews, 69(3), 279-324.
  3. Purnell Jr, D. L., & Bozzelli, J. W. (2018). Thermochemical Properties: Enthalpy, Entropy, and Heat Capacity of C2–C3 Fluorinated Aldehydes. Radicals and Fluorocarbon Group Additivity. The Journal of Physical Chemistry A.
  4. Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., & Dahl, G. E. (2017, August). Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning-Volume 70(pp. 1263-1272). JMLR. org
  5. Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., & Bengio, Y. (2017). Graph attention networks. arXiv preprint arXiv:1710.1090