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 DFT descriptors to enhance the performance of neural networks for general reaction condition prediction. We treat reaction condition prediction as a multi-class classification task and report the performance of neural networks trained on 59,512 reactions with 283 distinct reaction condition classes and varying input embedding compositions. We show that by combining structural and general DFT descriptors in optimized ratios, models with input size up to 15% smaller than their purely structural counterparts can provide comparable recall, top-1 and top-3 accuracies. Moreover, we report improvements of up to 6%, 7% and 9% in weighted F1 score, top-1 accuracy and weighted recall, respectively, for neural networks trained on combined general DFT and structural descriptors when compared to purely structural models with equivalent architectures and input sizes. Remarkably, these results were achieved using a training set containing 267 times fewer data points than the one used for generating and embedding structural descriptors, despite both embedding strategies being similar unsupervised learning algorithms.