Abstract
<jats:p>Molecular spectroscopy is a cornerstone of analytical chemistry, yet the analysis of spectral data remains challenging due to noise, compressive sampling artifacts, and the difficulty of fusing multiple spectroscopic modalities with structural information. We propose M3Spec-Net, a unified deep learning framework for multimodal molecular spectroscopy analysis that integrates spectral reconstruction, 3D conformation modeling, and functional group detection within a single architecture. M3Spec-Net introduces three key innovations: a dual-attention spectral reconstruction module that combines local convolutional and non-local self-attention for high-fidelity spectrum recovery from noisy or compressed measurements; a 3D Gaussian conformation encoder that represents molecular structure as a continuous Gaussian field, enabling efficient structure-to-spectrum mapping; and a pace-synchronized multimodal fusion module that aligns structural and spectral feature spaces by synchronizing their encoding paces before cross-attention fusion. A hybrid localization head further detects functional group fingerprints from the fused representations. Experiments on the QM9 dataset, GEOM drug-like molecule dataset, and NIST IR spectral database demonstrate that M3Spec-Net consistently outperforms existing methods across three key tasks: spectral reconstruction (37.45 dB PSNR at 20% compression), structure-to-spectrum prediction (0.019 MAE), and functional group detection (0.919 F1score). Ablation studies confirm the effectiveness of each component, and human evaluation by expert chemists validates the perceptual quality of reconstructed spectra. Keyword: Multimodal Spectroscopy, 3D Gaussian Representation, Pace-Synchronized Fusion, Dual-Attention Reconstruction, Functional Group Detection</jats:p>