Graph-neural-network predictions of solid-state NMR parameters in silica from spherical tensor decomposition

Area of Science:

• Materials Science
• Computational Chemistry
• Spectroscopy

Background:

• Nuclear magnetic resonance (NMR) is crucial for determining local atomic structure.
• Computational predictions of NMR parameters aid in interpreting experimental data and validating structural models.
• Machine learning (ML) offers an efficient method for these NMR predictions.

Purpose of the Study:

• To systematically investigate graph neural network (GNN) approaches for representing and learning tensor quantities in solid-state NMR.
• To assess the impact of ML model accuracy on predicting key NMR properties.
• To apply ML models to diverse material systems, including amorphous silica and dynamic processes.

Main Methods:

• Utilized graph neural networks (GNNs) to learn anisotropic magnetic shielding and electric field gradient tensors.
• Evaluated prediction quality for chemical shifts, quadrupolar coupling constants, tensor orientations, and 2D NMR spectra.
• Applied ML models to amorphous SiO2 configurations and the dynamics of cristobalite α-β inversion.

Main Results:

• Demonstrated the effectiveness of GNNs in predicting solid-state NMR tensor quantities.
• Showcased the translation of numerical ML accuracy into high-quality predictions of experimentally relevant NMR properties.
• Successfully applied ML models to complex, large-scale material structures and dynamic processes.

Conclusions:

• The study advances ML-driven NMR predictions for both static and dynamic material behaviors.
• This work bridges the gap between first-principles modeling and experimental NMR data.
• Streamlined ML approaches are paving the way for more efficient analysis of complex materials using NMR spectroscopy.