Abstract
Despite commonly applied corrections, known as shimming, the magnetic field in an NMR spectrometer is never perfectly homogeneous. This undesired effect distorts the lineshapes, degrades the resolution, and lowers the signal-to-noise ratio in the collected spectra. As a remedy, numerical techniques have been developed to correct the spectra after acquisition, with reference deconvolution being the most popular example. However, these methods require a precarious parameter guess from the user, making them inconvenient and prone to errors. In particular, the imperfections of the reference deconvolution manifest themselves as strong oscillations at the spectral baseline. We propose a postacquisition shimming tool named ShimNet based on a convolutional neural network with the attention mechanism. The model learns the distortion characteristics of a given spectrometer from a series of calibration measurements. Once trained, it can correct any spectrum from the same machine in a fully automatic way. It achieves slightly better reconstruction quality than the existing methods and is considerably faster. This makes ShimNet an excellent tool for laboratories performing routine and massive NMR measurements for chemists. As an exemplary application, we demonstrate that the model properly corrects liquid-state spectra of small molecules, such as azarone, styrene, Cresol Red, or sodium butyrate. The open-source Python code is freely available from https://github.com/center4ml/shimnet and as the web service at https://huggingface.co/spaces/NMR-CeNT-UW/ShimNet.