A robust, deep learning-based analysis of time-domain signals for NMR spectroscopy

Abstract When analyzing the Free Induction Decay (FID) signal produced by nuclear magnetic resonance (NMR) spectroscopy, Fourier transforms (FT) are used to decompose time-domain signals arising from nuclear interactions. This transformation enables the extraction of frequency-domain information, allowing for the recognition of patterns within the generated NMR spectra. Most modern NMR processing software applies FT to generate the final spectra. Researchers process FID using various techniques, such as phase correction, windowing, and FT, to enhance the interpretation of the obtained spectra. This processing step requires careful consideration of the characteristics of the original data and can also be influenced by the researchers' experience, often making it time-consuming to produce reliable results. However, recent advancements in artificial intelligence, particularly deep learning, have demonstrated superior pattern recognition capabilities compared to humans in complex scenarios. These developments have been successfully applied to various aspects of NMR spectroscopy. In this study, we demonstrate that neural networks can replace FT in NMR spectroscopy, enabling robust and rapid prediction of spectra and peak lists from FID signals. Our results confirm that deep learning can efficiently process NMR data to generate final spectra. As a proof of concept, we present the resulting spectra, along with peak lists predicted by supplying only FID input to the deep learning algorithm. The generated peak lists can be considered as spectra with infinite resolution.