<br/>Nuclear magnetic resonance (NMR) chemical shifts play an important role in the structural determination of materials, pharmaceuticals, and biologically relevant molecules. In combination with X-ray diffraction and chemical shift prediction models, one can routinely resolve atomic positions beyond the limits of diffraction alone. The goal of chemical shift prediction is to then refine or validate candidate structures that closely align with experimental shifts. While popular quantum chemistry methods such as density functional theory (DFT) work well for structure optimization and chemical shift prediction, the rapidly increasing computational demands become a significant bottleneck in large systems or when many candidate structures must be tested. More recently, low-cost machine learning (ML) chemical shift prediction models have started transforming the NMR crystallography of organic materials.<br/>Unfortunately, existing NMR ML models are limited in their elemental diversity (H, C, N, and O) and often perform noticeably worse for NMR parameters of atom types other than <sup>1</sup>H. To improve accuracy, we explore an improved model based on Δ-ML (DML) using artificial neural networks (NNs) that reproduce DFT-quality NMR parameters for small molecules containing HCNO. We leverage atom-centered symmetry functions (ACSFs) fed into an ensemble of NNs to improve accuracy and gauge the uncertainty of individual predictions through the standard deviation of the ensemble. Finally, we show that DML NMR reproduces experimental chemical shifts in real-world test cases at the computational cost of an inexpensive calculation.

Machine learning methods for reproducing density functional theory-based nuclear magnetic resonance chemical shifts

ACS Spring 2022

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