Nuclear magnetic resonance (NMR) chemical shifts of molecular crystals are a valuable observable in distinguishing between polymorphs. In addition to experimental shifts, ab initio chemical shift predictions are included to disentangle convoluted solid-state NMR spectrum and parse through candidate structures. To circumvent this computational bottleneck, machine learning (ML) algorithms that reproduce density functional theory (DFT) predicted shifts have been produced for crystals that contain HCNO atom types. Yet, there are still large errors in atoms other than 1H due to a lack of training data and efficient atomic representations. In this work, we report improved chemical shift predictions for atoms other than 1H and 13C compared to existing ML chemical shift model precedents for molecular crystals through the use of graph neural networks (GNNs). We first show that our GNN models learn more efficiently than descriptor-based ML implementations. We then demonstrate the first use of convolution and attention graph layers for chemical shift prediction. And finally, we exploit ensemble averaging to significantly improve overall prediction performance relative to DFT and to extract uncertainty information. Overall, our GNNs show great promise and can be easily extended to more diverse stoichiometries without too much feature engineering (i.e. without specialist input). Our GNN code is based on pytorch/pytorch-geometric and is a modification of the existing MatDeepLearn package.

Improved machine learning-based nuclear magnetic resonance chemical shift prediction accuracy via graph neural networks

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.