Ultraclean Pure Shift Spectroscopy with Fast Acquisition Based on Deep Neural Network

Compared to conventional nuclear magnetic resonance spectroscopy, pure shift techniques can greatly enhance the resolution of proton spectra. The lengthy acquisition time resulting from the specific two-dimensional acquisition mode and the chunking sidebands caused by concatenating data chunks to construct pure shift signals, however, provide obstacles to the wider use of pure shift techniques. To tackle these two problems, we provide a deep-learning-based reconstruction method to obtain an ultraclean and high-quality pure shift spectrum. Notably, we design a time-to-frequency (T2F) network structure, where the input of the network is the original time-domain signal and the output is the reconstructed spectrum. A specific input processing module is proposed to process signals with different data points, substantially increasing the network's adaptiveness in varying experimental conditions. Results on practical samples illustrate that the network can not only effectively suppress chunking sidebands and noise in the fully sampled pure shift data, but also efficiently remove severe undersampling artifacts from the undersampled pure shift data with a considerable acceleration of acquisition time.