CMD-Net: A Hybrid CNN-Mamba Denoising Network for NMR Spectroscopy

Abstract Nuclear magnetic resonance (NMR) spectroscopy is essential for chemical and biological analysis, yet its inherent insensitivity frequently yields noise-corrupted spectra that complicate weak signal detection and quantitative measurements. Here, we propose CMD-Net, a hybrid CNN-Mamba denoising network, to recover clean NMR signals closely approximating the high-SNR reference spectrum acquired with a large number of scans. The architecture integrates CNN-based local feature extraction with Mamba-2 global context modeling within a U-shaped encoder–decoder framework, enabling direct noise reduction in the frequency domain. Trained solely on synthetic 1D NMR data, CMD-Net can be applied to 1D heteronuclear spectra and 2D/3D NMR data sets. Extensive validation demonstrates that CMD-Net consistently outperforms state-of-the-art denoising methods across multiple noise levels, achieving superior weak peak preservation and artifact suppression. With a balance between performance and model size, CMD-Net holds promise for low-SNR spectral analysis and protein structural studies in chemistry, biology, materials science, and life science.