Noise reduction of nuclear magnetic resonance spectroscopy using lightweight deep neural network

Abstract

Nuclear magnetic resonance (NMR) spectroscopy serves as a robust non-invasive characterization technique for probing molecular structure and providing quantitative analysis, however, further NMR applications are generally confined by the low sensitivity performance, especially for heteronuclear experiments. Herein, we present a lightweight deep learning protocol for high-quality, reliable, and very fast noise reduction of NMR spectroscopy. Along with the lightweight network advantages and fast computational efficiency, this deep learning (DL) protocol effectively reduces noises and spurious signals, and recovers desired weak peaks almost entirely drown in severe noise, thus implementing considerable signal-to-noise ratio (SNR) improvement. Additionally, it enables the satisfactory spectral denoising in the frequency domain and allows one to distinguish real signals and noise artifacts using solely physics-driven synthetic NMR data learning. Besides, the trained lightweight network model is general for one-dimensional and multi-dimensional NMR spectroscopy, and can be exploited on diverse chemical samples. As a result, the deep learning method presented in this study holds potential applications in the fields of chemistry, biology, materials, life sciences, and among others.