Hybrid Domain Enhancement Network for Ultraclean Pure Shift Spectroscopy With Fast Acquisition

Pure shift technique can greatly improve the resolution of proton spectra compared to conventional nuclear magnetic resonance (NMR) spectroscopy but at the cost of longer acquisition time. To accelerate data acquisition in pure shift experiments, nonuniform sampling (NUS) is usually incorporated along with advanced data processing methods to reconstruct the spectrum. While some deep learning (DL) approaches have been proposed for this purpose, their performance declines when dealing with data under varying experimental conditions, primarily because they solely rely on frequency-domain information. To overcome this limitation, we develop a novel hybrid domain enhancement (HDE) network that combines time- and frequency-domain signals as input. We introduced an adaptive-domain transform unit (ADTU) to extract time-domain features that capture segmental correlations while retaining some temporal sequential information. Our experiments demonstrate that the proposed network effectively suppresses chunking sidebands and undersampling artifacts, showing exceptional robustness in complex scenarios, particularly under phase distortion. On real experimental data, it achieves a 5%–33% higher precision than state-of-the-art (SOTA) methods, indicating fewer artifacts in the reconstructed spectrum. This approach holds great potential for broader application in NMR spectroscopy, offering clearer and more accurate spectral interpretation.