Unveiling uncertainty in microplastic quantification: Artificial intelligence integrated Raman analysis

Accurate microplastic (MP) analysis is essential for mitigating global MP pollution. While μ-Raman spectroscopy provides high spatial resolution and molecular specificity for MP detection, its low scattering efficiency and time-intensive analysis process limit practical applications. Sub-sampling strategies, analyzing only a fraction of the filter, have been proposed to reduce burden, leading to quantification bias. Such bias in environmental monitoring may distort exposure assessments and lead to misinterpretation of MP-related health implications. In this study, we comprehensively optimized Raman acquisition conditions and developed an AI model to enhance the reliable identification of the chemical and physical properties of MPs in environmental samples. Our model successfully identifies MPs often overlooked during human inspection, accurately distinguishes low-quality spectra from background, and enables size and morphology analysis. Using this validated framework, we quantitatively assessed the uncertainties introduced by sub-sampling under realistic conditions. Quantitative simulation based on AI-integrated Raman mapping showed that for heterogeneous distributions, typical of real-world samples, achieving a normalized root-mean-square error of extrapolated MP count below 10% requires higher filter coverage (40-70%) than for homogeneous distributions (3-35%), revealing that assuming homogeneous MP distribution can bias results. We also found that frequent and small-block sampling is an efficient strategy for reducing the required coverage. Overall, this study elucidates the inherent uncertainties of sub-sampling strategies and highlights the potential of AI-integrated μ-Raman spectroscopy for accurate and efficient MP analysis in environmental monitoring.