Generalized Bondnet: A Generalized Graph Neural Network Framework to Predict Reaction Thermodynamics and Kinetics in Developing Reaction Networks and High-Throughput Screening Runs
Santiago Vargas, University of California, Los Angeles
We have extended previous work in the Persson group on graph neural networks for predicting reaction thermodynamics to predict reaction kinetics via a transfer learning approach. This approach also leverages new quantum mechanical descriptors such as QTAIM and NBO to get better predictions of reaction properties. This improved network will be introduced into a high-throughput framework for constructing reaction networks in UV-lithography, cyclical polymerization and battery SEI contexts. New developments also see 4x improvements in training time as well for training on immense datasets.
Abstract Author(s): Santiago Vargas, Rishabh Guha, Evan Spotte-Smith, Sam Blau, Kristin Persson