Most chemical reactions of interest do not occur as a single elementary step and in fact occur as a part of complex reaction networks. Simple kinetic models, ignorant of the full reaction network, cannot predict many important phenomena (such as minor byproducts or fouling) that can disturb performance in a commercial reaction process. Accurately predicting the behavior of these reaction networks is vital to optimizing existing reaction processes, screening possible new reaction processes, developing better design principles, and discovering less obvious but important side processes. However, parametrizing processes in these reaction networks can require expensive high accuracy quantum chemistry calculations or experimental data to reach predictive accuracy making it most practical to combine different computational approaches with different levels of accuracy. Doing this in the most effective way requires us to fuse “2D” graph-based thinking with “3D” molecular coordinate-based thinking. Where 2D tools such as network ODE simulations, graph-based chemical property prediction, and graph-based reaction templates, are very fast and machine learning algorithms are easy to train. In contrast, 3D tools such as quantum chemical and molecular dynamics calculations, are much more expensive and challenging to run, but are more accurate representations of the physics.
I will present a workflow combining a family of tools: RMG, RMS, ARC, Pynta, and PySIDT to enable fusion of 2D and 3D techniques leveraging automatic mechanism generation, automatic simulation analysis, automated quantum chemistry and machine learning to efficiently automate kinetic model construction. I will also discuss some of my more recent work using Pynta and PySIDT to fuse 2D and 3D machine learning enabling highly efficient automatic calculation of coverage dependent rate coefficients on surfaces.
