UC Berkeley

The nuclear magnetic resonance (NMR) chemical shift tensor is a highly sensitive probe of the electronic structure of an atom and furthermore its local structure. Recently, machine learning has been applied to NMR in the prediction of isotropic chemical shifts from a structure. Current machine learning models, however, often ignore the full chemical shift tensor for the easier to predict isotropic chemical shift, effectively ignoring a multitude of structural information available in the NMR chemical shift tensor. Here, we begin with a proof of concept model and demonstrate that the anisotropy in highly symmetric silicon Q4 sites can, in fact, be modeled as a function of local geometry. We use a combination of simple geometric features as well as symmetry based hand-crafted features as input into a symbolic regression algorithm to determine a functional form of this relationship. After showing the these sites can be modelled, we seek a generalizeable model and turn to equivariant graph neural networks (GNN) to predict full 29Si chemical shift tensors in silicate materials. The equivariant GNN model predicts full tensors to a mean absolute error of 1.05 ppm and is able to accurately determine the magnitude, anisotropy, and tensor orientation in a diverse set of silicon oxide local structures. When compared with other models, the equivariant GNN model outperforms the state-of-the-art machine learning models by 47%. The equivariant GNN model also outperforms historic analytical models by 57% for isotropic chemical shift and 91% for anisotropy. We then develop a geminal J-coupling GNN and use both the shift tensor and J coupling GNNs in a Gauss-Newton optimization procedure to refine the structure of siliceous zeolite ZSM-12. This machine learning based refinement was performed on a single CPU and required only 2 hours to complete which is a factor of 36 times faster than the previous generation of DFT based refinements.