Theoretical Studies of Nuclear Magnetic Resonance Chemical Shifts

Abstract

Nuclear magnetic resonance (NMR) spectroscopy is among the most important analytical methods for the determination of the chemical structure of matter. Its versatile applicability leads to the necessity of reliable computational techniques for simulating NMR spectra and parameters. Depending on the field of application, these should be general and yield sufficiently accurate results with limited computational demands. Therefore, this thesis contains five articles which deal with the assessment of existing methods for NMR chemical shift calculation, the development of new approaches to achieve a higher accuracy, and the application of these new methods to provide answers to actual open research questions. <br /> The calculation of accurate NMR chemical shifts can be done with density functional theory (DFT) in a relatively efficient way. It is, however, not straightforward to estimate the performance of each density functional approximation (DFA) without rigorous tests, of which there are too few in the literature, especially for nuclei other than ¹H and ¹³C. Therefore, this work focuses on the comprehensive evaluation of DFT methods regarding their suitability for the calculation of ²⁹Si and ¹¹⁹Sn NMR chemical shifts. In the scope of these studies, the SiS146 and the SnS51 benchmark sets are compiled and used for the assessment of various DFAs. Correspondingly, the use of computationally more demanding hybrid functionals does not generally lead to improved results compared to more efficient (meta-)generalized gradient approximation (GGA) DFAs. However, a relativistic treatment (ideally including spin-orbit effects) reveals being indispensable if heavy atoms are close to a light NMR nucleus (²⁹Si) or if the nucleus is heavy itself (¹¹⁹Sn). <br /> Afterward, two machine learning-based correction methods (Δ-ML) are presented for the improved description of ¹H and ¹³C NMR chemical shifts. The NMR prediction performance of DFT can generally be improved with the Δ_corr-ML method, which adds a correction term based on highly accurate coupled cluster reference data. Besides that, the Δ_SO-ML method is built on reference data obtained from spin-orbit relativistic DFT calculations of organic molecules containing heavy atoms and represents an efficient alternative to relativistic calculations. Both methods are evaluated using the underlying training and test data sets as well as benchmark sets beyond that data and show to consistently outperform the routinely applied linear regression approach. Applying both methods in a multilevel workflow can significantly reduce the deviation to experimental data of ¹³C NMR chemical shifts. Finally, a joint study with experimentalists on platform molecules for surface decoration is presented, that showcases the usage of the Δ_corr-ML method to verify computed ¹H NMR chemical shifts, which leads to important new insights into the involved chemical reactions.