Interpretable Microplastic Detection in Environmental Samples via Raman Spectroscopy and LLM-Based Semantic Feature Fusion

Abstract Microplastic pollution presents significant environmental and health risks, with increasing concern over its presence in aquatic, terrestrial, and atmospheric ecosystems. Despite advances in analytical methods, reliable identification and quantification of microplastics remain challenging, particularly under real-world conditions. Traditional Raman-based methods often struggle with accurate detection due to interference from environmental background materials, such as minerals and biological debris, leading to both overestimation and underestimation of microplastic concentrations. This study introduces a novel framework for microplastic identification by integrating Raman spectra with multimodal semantic features derived from large language models (LLMs). Unlike conventional methods relying on raw spectral vectors, the proposed approach combines expert-level semantic descriptors, including peak positions, intensities, widths, and symmetry, with original spectral data to overcome limitations caused by environmental noise and background interference. By employing an attention-based feature fusion strategy, the framework enhances the robustness, accuracy, and interpretability of microplastic detection, ensuring reliable performance across varying environmental conditions and instrument settings. Evaluations on two independent Raman datasets demonstrate consistent improvements in classification accuracy and generalization. External validation on a real-world environmental monitoring dataset containing both plastic and non-plastic samples further confirms the robustness and applicability of the method in open-set screening conditions. In addition to improving classification performance, this study emphasizes interpretability by linking model decisions to observable spectral features, which is essential for expert validation and decision-making. Overall, this work presents an interpretable and scalable strategy for Raman-based microplastic identification, with potential applications in environmental monitoring, recycling systems, and studies of microplastic exposure and health impacts.