Field readiness of Raman-AI pipelines for microplastic analysis across environmental matrices

Microplastic monitoring increasingly depends on workflows that can inform decisions beyond controlled laboratory datasets. Raman spectroscopy coupled with AI/ML or chemometrics can assist polymer identification, but high benchmark accuracy does not show whether a workflow remains reliable across matrices, instruments, preprocessing choices, and field use. This protocol-guided critical review evaluates the evidence needed before Raman-AI microplastic analysis can be considered field-ready. A Scopus search retrieved 151 records and retained 61 studies, coded by sample matrix, Raman modality, acquisition metadata, preprocessing, model family, analytical task, validation design, reporting assets, and deployment claim. The corpus is recent and classification-led. In all, 43 studies (70.5%) were published in 2023-2025; classification/identification accounted for 42/61 studies (68.9%); water-related or controlled/laboratory matrices dominated. Micro-Raman was the main backbone (36/61, 59.0%), followed by SERS-Raman (10/61, 16.4%), whereas quantification/regression (6/61, 9.8%) and imaging/mapping (3/61, 4.9%) remained limited. Reporting gaps were substantial. Spectral resolution appeared in 20/61 studies, normalization/scaling in 33/61, cosmic-ray removal in 16/61, leakage-control discussion in 9/61, uncertainty analysis in 6/61, interpretability in 4/61, and data/code availability in 8/61. We present a readiness-oriented framework that evaluates the full Raman-AI workflow, with classifier performance as one component. The MCDA-informed synthesis points to open benchmarks with code, grouped and external validation, uncertainty-aware outputs, peak-level interpretability, traceable preprocessing, and cross-matrix/cross-instrument testing. Until these elements are routine, field-readiness claims should remain conditional.