PhysRaman-Net: A Physics-Informed Neural Network for Cross-Device Raman Spectral Analysis

Abstract

Raman spectroscopy as a non-destructive optical technique, is extensively employed for material characterization. However, inherently weak scattering signals and device-specific noises result in significant spectral variability across instruments, which severely limits the cross-device generalization of deep learning models. Since molecular vibrational signatures remain intrinsically device-invariant, we propose PhysRaman-Net, a novel physics-informed encoder-decoder framework for extracting device-invariant spectral representations. To mitigate the scarcity of labeled experimental data, we introduce a data augmentation strategy designed to model the noise differences among different Raman instruments. By modeling dominant noise sources according to their physical origins, we synthesize large-scale training samples that capture realistic spectral degradation patterns from theoretical spectra in the Computational Raman Database. Architecturally, we design a Permutation Invariant Encoder (PIE) based on the Transformer architecture to effectively capture the intrinsically unordered relationships among molecular vibrational modes. Distinctively, the proposed PhysDecoder is formulated as a differentiable physical model that explicitly encodes Raman scattering laws, thereby constraining the network to learn Raman tensors that represent the intensity information of Raman spectral peaks rather than abstract latent features. Extensive experiments on cross-device benchmarks demonstrate that PhysRaman-Net achieves superior generalization capability, attaining a classification accuracy of 96.8% for cross-device material classification, substantially outperforming state-of-the-art Convolutional Neural Network(CNN) and autoencoder baselines.