Date of Award

Spring 5-7-2025

Author's School

McKelvey School of Engineering

Author's Department

Electrical & Systems Engineering

Degree Name

Master of Science (MS)

Degree Type

Thesis

Abstract

Nuclear Magnetic Resonance (NMR) spectroscopy is a powerful analytical technique widely used for molecular structure elucidation in chemistry, biology, and medicine. However, spectral accuracy is often degraded by noise—particularly in low acquisition time settings—resulting in reduced resolution and obscured chemical features. While traditional noise reduction techniques such as signal averaging can improve spectral quality, they require longer acquisition times, limiting their utility in real-time and high-throughput applications.

This thesis presents a deep learning-based denoising framework designed to enhance the quality of complex-valued NMR spectra. The proposed model, built upon a U-Net architecture, incorporates both real and imaginary components of the signal to preserve phase information and spectral fidelity. A core innovation of this work is the integration of a noise level estimator that predicts the noise intensity present in the input signal. This estimated noise level is then used to condition the denoiser, enabling the model to adapt dynamically to a wide range of noise scenarios.

The framework was evaluated across two primary settings: (1) synthetically generated noisy spectra created by adding machine-derived noise to clean Free Induction Decay (FID) signals, and (2) real experimental spectra obtained from Bruker spectrometers under uncontrolled acquisition conditions. Quantitative metrics such as Signal-to-Noise Ratio (SNR), Peak Height-to-Noise Ratio, and Normalized Mean Square Error (NMSE) were used to assess performance. Results demonstrate that the adaptive framework consistently improves spectral clarity, outperforming traditional denoising baselines and unconditioned deep learning models.

By combining noise-aware learning with complex-valued signal processing, this framework offers a robust and scalable solution for enhancing NMR spectra without increasing acquisition time. These findings contribute to the growing intersection of deep learning and spectroscopic analysis, advancing the potential for automated, high-throughput NMR applications in both research and industry.

Language

English (en)

Chair

Ulugbek Kamilov, Electrical & Systems Engineering; Computer Science & Engineering

Committee Members

Marcus Foston, Hong Hu

Available for download on Thursday, April 23, 2026

Share

COinS