AI-based Spectra Processing and Analysis for NMR

Abstract

Nuclear Magnetic Resonance (NMR) spectroscopy is a powerful analytical technique for obtaining atomic-level information across various scientific fields. However, direct interpretation of raw NMR data is impractical due to its complexity. While traditional signal processing methods are widely used, they face limitations in handling advanced tasks. Consequently, advanced computational approaches are required to optimize spectral reconstruction and improve analytical precision. Artificial Intelligence (AI), particularly deep learning, has demonstrated significant potential in addressing these challenges. This thesis explores AI-based signal processing in NMR spectroscopy, focusing on spectral reconstruction, resolution enhancement, and quality assessment.

The first study introduces Low-Rank Decoupling (LRD), a method leveraging prior knowledge of J-coupling for homonuclear virtual decoupling. LRD enhances spectral resolution while maintaining sensitivity and minimizing artifacts, outperforming conventional decoupling approaches.

The second study presents Magnetic Resonance processing with Artificial intelligence (MR-Ai) as an alternative to conventional nonlinear NMR processing. A 1D WaveNet-based NMR Network (1D WNN) is developed to address non-uniformly sampled (NUS) reconstruction as a pattern recognition problem, surpassing traditional methods in stability and accuracy. MR-Ai is also adapted for virtual decoupling, demonstrating robustness against variations in J-coupling values.

The third study extends MR-Ai beyond traditional NMR processing. A 2D WNN architecture is designed to reconstruct Echo (or Anti-Echo) spectra, correcting phase-twist distortions in incomplete quadrature detection. Additionally, MR-Ai introduces a reference-free evaluation metric, estimating uncertainty in spectral reconstructions for direct quality assessment without external references.

The fourth study introduces Peak Probability Presentation (P3), a novel AI-driven spectral visualization technique. To achieve this, an nD WNN architecture is developed for pattern recognition in nD NMR spectra. Unlike traditional intensity-based representations, P3 assigns a probability score to each spectral point, providing artifact-free, ultimate-resolution spectral interpretation. The results demonstrate P3’s superior performance in peak detection, spectral clarity, and noise differentiation. Additionally, P3 is integrated into Targeted Acquisition (TA) to develop a quantitative spectrum quality score for real-time spectral quality assessment and optimized data acquisition.

Overall, this thesis demonstrates that AI-driven NMR processing not only enhances existing methods but also introduces fundamentally new approaches to spectral reconstruction, resolution enhancement, and quality assessment. As AI evolves, its integration into NMR workflows is expected to revolutionize the field, making high-quality spectral data more accessible, interpretable, and efficient.