JultikaUniversity of Oulu repository

Abstract

Application of nuclear magnetic resonance (NMR) experiments for nuclei other than proton, such as ¹³C, can be limited by low signal-to-noise ratio (SNR) (Although, it can be a problem for protons also). Special techniques to increase the SNR can suffer from short signal enhancement lifetimes, usually limited by the spin-lattice relaxation constant T₁. Long-lived spin states (LLS) can increase the polarization lifetimes significantly. LLS in this work is a naphthalene derivative singlet state that has a pair of ¹³C spins in it’s center, with a small chemical shift difference (dCS). A small dCS is required to access the spin-state with NMR-methods. This small dCS can be difficult to find experimentally so the aim of this work is to get an idea if a change of temperature increases this difference. In principle, the ensemble averages of the dCSs could be computed with quantum chemistry (QC) methods. However, computing enough accurate dCS values with these methods is impractical due to large computational load. Also, a large number of calculations would be needed to get a meaningful estimate of the dCS ensemble average even for one temperature. Therefore, the computational load is largely circumvented here by building a quantum-chemistry property surface (QCPS) model. The model uses QC-computed shielding constant values and interpolates them as a function of spatial degrees of freedom (DOF) of the molecule. This can possibly allow for a meaningful reduction in the number of required QC computations. The model can then be used to analyze on molecular dynamics (MD) trajectories simulated at different temperatures and a change of the ensemble average of dCS can be determined. QC computations are a bottleneck that limits the QCPS model dimensions. Therefore, this work also investigates which degrees of freedom the model should contain. Also a second QCPS model is optimized with smaller basis set of QC computed training set in order to compare the effect of a less accurate training sample to the model’s performance. The resulting ddCS/dT-slope is dominated by statistical error due to model’s low dimension of 28 DOF. This model lacks predictive range and the obtained dCS correlation coefficient with explicitly QC-computed test set of 0.7049 is low. The correlation coefficient of the model optimized with less accurate training samples is 0.7050, so both models have similar correlations with their corresponding test sets. In order to increase the model performance, an investigation on the required model dimension is suggested with, possible addition of solvent molecules.