NMR Crystallography in the Big Data Era: New Methods and Applications Powered by Machine Learning

Structure determination of materials is key to understanding their physical properties. While single-crystal X-ray diffraction is the gold standard for structures displaying long-range order, many materials of interest are polycrystalline and/or disordered, which is a challenge for diffraction methods. On the other hand, nuclear magnetic resonance (NMR) spectroscopy probes the local environment around nuclei, and does not require long-range order. It is thus the method of choice for investigating the structure of disordered solids. In particular, NMR crystallography based on chemical shifts has proven able to determine the atomic-level structure of various materials through the combination of solid-state NMR experiments, crystal structure prediction (CSP) protocols, and density functional theory (DFT) computation of chemical shifts. However, several drawbacks prevent the widespread use of NMR crystallography, especially for disordered materials, such as the computational resources required to perform DFT chemical shift computations. In this thesis, we use machine learning to help alleviate these drawbacks. We extend the capabilities of ShiftML, a previously introduced model of chemical shifts of molecular solids, and incorporate the model into CSP protocols, in order to drive the generation of candidate crystal structures towards the experimentally observed structure. We also predict chemical shifts using ShiftML on a large database of crystal structures, and leverage the resulting database of chemical shifts to help assign measured chemical shifts to atomic sites without prior knowledge about the three-dimensional structure of the molecule under study. The database is also used to construct chemical shift-dependent interaction maps in molecular solids. The maps generated can in turn be used to score candidate crystal structures without performing any additional DFT-level chemical shift computation, and to construct structural constraints to drive CSP protocols. Another challenge tackled in this thesis is the resolution of 1H NMR spectra of solids. Dipolar coupling between spins lead to broadened lineshapes, which can (partially) be removed by spinning the sample at the magic angle. However, at finite spinning rates, these interactions are not completely removed. We develop a convolutional recurrent neural network to obtain the spectra that would be obtained at infinite spinning rates from a set of spectra measures at variable spinning speeds. The model is applied both to one-dimensional 1H spectra and two-dimensional 1H-1H correlation experiments. Finally, we investigate the structure of amorphous molecular solids by NMR crystallography, by replacing DFT chemical shift computations by ShiftML. This allows the computation of chemical shifts for ensembles of large structures generated by MD, that we compare to experimental values in order to extract preferred conformations and noncovalent interactions in amorphous compounds. A general method to de