Neural Network Method for Diffusion-Ordered NMR Spectroscopy

Diffusion-ordered NMR spectroscopy (DOSY) presents an essential tool for the analysis of compound mixtures by revealing intrinsic diffusion behaviors of the mixed components. For the interpretation of the diffusion information, intrinsically designed algorithms for a DOSY spectrum reconstruction are required. The estimated diffusion coefficients are desired to have consistency for all the spectral signals from the same molecule and good separation of signals from different molecules. For this purpose, we propose a novel method that adopts a coordinated multiexponential fitting to ensure the consistency of diffusion coefficients and apply a sparse constraint to enhance the robustness. A lightweight neural network is applied as an optimizer to solve this highly nonlinear and nonconvex optimization problem. The proposed method provides estimated diffusion coefficients with excellent distinguishment between species and outperforms the state-of-the-art reconstruction algorithms, such as the Laplacian inversion and the multivariate fitting methods.

Automated summary

Diffusion-ordered NMR spectroscopy (DOSY) is an important tool for analyzing compound mixtures by revealing their diffusion behaviors. This paper proposes a novel method for reconstructing DOSY spectra, which ensures consistency of diffusion coefficients and enhances robustness using coordinated multiexponential fitting and sparse constraint. A lightweight neural network is applied as an optimizer to solve this nonlinear and nonconvex optimization problem. The proposed method provides estimated diffusion coefficients with excellent species distinction and outperforms state-of-the-art reconstruction algorithms.