General Chemically Intuitive Atom- and Bond-Level DFT Descriptors for Machine Learning Approaches to Reaction Condition Prediction

We demonstrate the usefulness of general atom- and bond-level density functional theory (DFT) descriptors to enhance the performance of neural networks for general reaction condition prediction. We treat condition prediction as a multiclass classification task and report the performance of neural networks and random forests as evaluated by 5-fold cross-validation on a 69,935 reaction data set with 296 distinct single-component reaction condition classes and varying input embedding compositions. We show that by combining structural and general DFT descriptors, models with up to 71% fewer trainable parameter than their purely structural counterparts can provide comparable or superior weighted precision, top-1 and top-3 accuracies. Moreover, we report improvements of up to 5, 10, and 11% in weighted precision, top-1 accuracy and F1 score, respectively, for neural networks trained on hybrid representations which combine general DFT and structural descriptors, when compared to structural models with equivalent architectures and input sizes. Remarkably, the best performing neural network trained on hybrid embeddings outperforms the best purely structural model investigated despite the latter benefiting from of an embedding strategy with 267 times more data points than the one used for generating and embedding hybrid descriptors, with both strategies being unsupervised learning algorithms that share considerable conceptual and architectural similarities.