Development of new experimental and data processing methods at critical signal-to-noise conditions in nuclear magnetic resonance

The Nuclear Magnetic Resonance(NMR) phenomena can be used to investigate a wide range of structural properties of various materials. The major drawback of NMR is the long measurement time needed to overcome the intrinsic low sensibility and Signal-to-Noise Ratio (SNR). This thesis aims to combine dissolution Dynamic Nuclear Polarization(d-DNP) with Digital Signal Processing (DSP) to improve the quality of NMR experiments when the acquisition time is constrained. The conversion of the Hypersense to work at the 1H Larmor frequency is reported along with the polarization build-up curves for two water-based sample formulations for cardiac magnetic resonance imaging perfusion experiments. Also, a +18% in the solid-state polarization level of a 13 C-urea sample is observed when the microwave irradiation is frequency modulated. The statistic of the noise after the use of a low-rank approximation filter is described and used to define the Maximum Likelihood Estimator (MLE) for the signal parameters. This MLE on the filtered data provides a better estimation of parameters, in terms of mean square error, once compared with both the MLE on the unfiltered data and the least-square estimator on the filtered data in a Montecarlo simulation. Two new estimators for the Population Weighted Rate Average (PWRA) of a multi-exponential decay curve, the Polynomial Fitting, and the Neural Network Ensemble (NNE), have been described and tested against the multi-exponential fitting. A Montecarlo simulation was used to compare the estimation performances of the three methods. The PF provides an almost unbiased estimation of the PWRA with an Absolute Percentage Error (APE) lower than 10% for 75% of the three exponential decay test sets. The NNE provides a slightly biased estimation of the PWRA with an APE lower than 15% for 95% of the whole test sets.