Neural network-based inversion of NMR dispersion profiles for enhanced analysis of food systems

Fast Field-Cycling Nuclear Magnetic Resonance (FFC-NMR) relaxometry is a powerful non-destructive technique used to study molecular dynamics and structures in various systems, including food products. This study introduces a novel machine learning framework to address the Quadrupole Relaxation Enhancement (QRE) phenomenon in FFC-NMR analysis. The proposed method leverages a pre-trained feed-forward neural network within a coordinate descent optimization algorithm to extract quadrupolar parameters and fit NMR Dispersion (NMRD) profiles. The neural network is trained using a unique model loss function combining \(L_2\) $L_2$ loss and predicted quadrupolar component of the NMRD profile accuracy. The approach is validated against a robust optimization method, showing strong concordance and potential for expedited analysis of extensive datasets. This advancement opens new avenues for assessing food quality and authenticity using FFC-NMR relaxometry.