Accurately Extending the Organic-Trained RexGen Graph Convolutional Neural Network to Inorganic Reaction Prediction

Computational chemists have taken great interest in machine learning in recent years, as techniques are being developed to produce faster predictions with higher accuracy. In 2019, Coley, et al, proposed a graph convolutional neural network (GCNN) model which predicts organic chemical reactions with 85\% accuracy. By creating graphs with atoms as nodes and bonds as edges, and exploiting the standard rules of organic chemical bonding, the trained model was able to output graphs corresponding to the actual product of the given reaction. This is a promising model, and would be even more powerful if also able to predict metal-based, inorganic chemical reactions. We explore the accuracy of this model for inorganic reaction predictions in this paper. Testing on 53,000 data points not included in the original paper, we were able to demonstrate 85\% accuracy in predicting chemical reactions, both organic and inorganic, with the pre-trained, organic-based GCNN. On exclusively inorganic reactions, the pre-trained model performed with 82\% accuracy. Re-training the model with our own training and validation datasets comprised of inorganic reactions, we were able to achieve 96.0\% accuracy, and 97.2\% accuracy with dropout regularization.