Physics-informed deep learning enables fast ultrahigh-resolution nuclear magnetic resonance spectroscopy

Pure shift NMR spectroscopy has found extensive applications in exploring the structure, function, and interactions of molecules in an ultrahigh-resolution manner. However, time-consuming data acquisition resulting from additional time dimension for pure shift evolution impedes its further applications. In this study, a general and robust AI-assisted NMR methodology combining non-uniform chunk sampling with physics-informed deep learning (DL) reconstruction is proposed for fast implementation of pure shift NMR spectroscopy. The proposed DL protocol enables the suppression on sparsely sampling artifacts, faithful recovery of weak signals, as well as high-fidelity reconstruction on peak intensities, thus implementing accelerated pure shift NMR while maintaining spectral quality. The well-trained model shows broad applicability across one-dimensional, two-dimensional, even multi-dimensional pure shift NMR. In addition, ablation experiments are further performed to provide mechanistic insights into deep learning reconstruction on sparse sampled pure shift NMR spectra. Moreover, its application potentials have been further demonstrated through in-situ monitoring of 1-butanol electrooxidation on Pt/C and PtRu/C catalysts. As a result, this study establishes a robust AI-assisted NMR framework for disentangling molecular structure and dynamics information for complex sample systems with high temporal and spectral resolution, and could find wide applications across multiple chemistry disciplines.