Physics-constrained pairwise representation learning for Raman spectra matching

Raman spectroscopy is a powerful and reliable tool for analyzing the chemical information of substances. Practical Raman spectra are often degraded by fluorescence background, stray light, environmental noise, and acquisition variations, leading to peak distortion, baseline drift, and reduced spectral consistency, which significantly increase the difficulty of reliable spectral matching. Existing deep learning methods have demonstrated promising performance by automatically learning spectral representations. Nevertheless, most of them are designed for classification tasks, making the learned features insufficiently robust to spectral degradation and diverse real-world Raman acquisition conditions. To address these limitations, this paper proposes a Physics-constrained Multi-Scale Network for Raman Spectroscopy (PCMS-Net) for robust spectral matching. A spectral simulation strategy based on physical guidance was developed to generate degradation-aware training pairs with diverse but physically reasonable spectral variations. Based on the Siamese matching framework, PCMS-Net directly learns the spectral similarity feature representations. PCMS-Net maintains the physical consistency of spectral features during the degradation process, constrains the feature representations, and simultaneously captures global spectral semantics and local multi-scale peak details to enhance the feature representations. This design enables the network to learn feature representations with discrimination and robustness for spectral matching. Experimental results show that PCMS-Net consistently outperforms traditional and representative deep learning methods in Raman spectral matching, demonstrating strong robustness under diverse real-world acquisition conditions.