High-quality pure shift NMR spectra by deep learning using multi-spectral input and joint loss functions

Pure shift methods greatly improve the resolution of nuclear magnetic resonance (NMR) spectra, aiding in the subsequent spectral analysis. However, existing approaches typically compromise sensitivity, introduce artifacts, or distort the signal integrals needed to quantify the amount of each component in a sample. Here, we propose a neural network, the Spin Echo to Pure Shift Network (SE2PSNet), that generates high-quality pure shift spectra with accurate integrals. Its input combines a set of spin echo spectra acquired at different echo times with a chemical shift binary spectrum produced by the existing SE2CSNet. From these, the network learns how NMR signals evolve across the spin echo spectra, while an attention mechanism with residual connections and a joint loss function amplify genuine signals, suppress noise, and extract integral-related features more accurately, thereby preserving both overall spectral quality and the weak signals essential for quantification. Across several representative samples, SE2PSNet resolved overlapping signals without introducing detectable artifacts, achieved sensitivity comparable to conventional single-pulse proton spectra, and provided accurate quantitative information. A neural network called SE2PSNet turns simple spin echo NMR data into high-quality pure shift spectra, resolving overlapping signals while preserving the sensitivity and accurate integral information, with broad potential applications.