Physics-Guided Self-Supervised Dual-Stream Transformer for Robust Optoelectronic Spectral Analytics

Optoelectronic sensing, such as reflectance or absorbance spectroscopy, enables non-contact measurement and provides informative signals for material characterization, quality inspection, and process monitoring. However, practical optoelectronic spectral analytics is often limited by scarce labels and distribution shifts caused by illumination variation, device drift, and measurement noise. To address these issues, a physics-guided self-supervised dual-stream Transformer framework is developed for robust learning from optoelectronic spectra. First, radiometric-consistent calibration is performed and physically meaningful spectral views are constructed, including calibrated spectra and derivative-enhanced representations. Second, a dual-stream Transformer encoder is designed to jointly model the complementary views, where cross-attention and gated fusion are adopted to adaptively aggregate spectral features. Third, self-supervised pretraining is introduced through masked spectral modeling and condition-invariant contrastive learning, enabling label-efficient representation learning. In addition, physics-regularized objectives, including illumination-invariance consistency and spectral smoothness priors, are incorporated during fine-tuning to improve generalization under cross-condition evaluation. Experimental results on optoelectronic spectral datasets demonstrate that the proposed method consistently improves predictive accuracy and robustness compared with representative baselines, particularly under low-label settings and cross-device testing.