DWSA: Dynamic-Weighting and Self-Attention Intelligent Framework Enables High-Sensitivity SERS Discrimination of Ultratrace Volatile Organic Compounds

Abstract Real-time on-site discrimination of ultratrace volatile organic compounds (VOCs) via surface-enhanced Raman scattering (SERS) is fundamentally challenged by spectral complexity and the limited generalization of single models. Here, we develop a high-performance 3D porous material (PM) based on biopolymer-silver nanocomposites, providing a reliable platform for trace gas enrichment and signal enhancement. Moving beyond material optimization, we pioneer a novel dynamic-weighting and self-attention intelligent framework (DWSA) to integrate convolutional neural network (CNN)-based global feature extraction with k-nearest neighbors (KNN)-based local similarity reasoning. The hybrid algorithm overcomes the bottlenecks of traditional static ensemble methods; its core innovation lies in a confidence-driven dynamic weighting strategy, which adaptively balances the contributions of both models based on prediction reliability rather than fixed ratios. This underlying mechanism is augmented by a self-attention module that adaptively recalibrates feature weights to focus on key molecular vibrational fingerprints, effectively filtering noise and enhancing interpretability. Based on this hybrid framework, we propose a detection strategy with potential for portable SERS applications, featuring rapid spectral acquisition, intelligent identification, and QR-code-based traceability. This study promotes the practical application of SERS technology in on-site environmental monitoring by integrating ultra-high gas adsorption efficiency with the development of a novel intelligent algorithm.