Application of machine learning in non-targeted analysis for environmental organic pollutants

The vast diversity of organic pollutants in the environment, coupled with the scarcity of authentic standards and reference spectra, poses major challenges to the reliable identification and quantification of these compounds. Machine learning (ML), a powerful tool for data processing, has garnered growing attention for its value in non-targeted analysis. This review summarizes recent advances in the application of ML to pollutant identification and quantification. In qualitative analysis, ML accelerates key steps in structure elucidation, including predicting MS/MS spectra to expand in silico libraries, inferring molecular formulas and substructures from spectral data, generating novel candidate structures beyond existing databases, and predicting retention time and collision cross section values to enhance identification confidence. In quantitative analysis, ML facilitates standard-free methodologies through optimized surrogate compound selection and prediction of ionization efficiency and response factors. Despite these achievements, salient challenges remain, including model generalizability across instrumental platforms, improved interpretability, and the establishment of unified frameworks for uncertainty quantification and reproducibility. Future advances toward more accurate, integrated, automated, and transferable ML systems are expected to enable scalable and intelligent analysis of environmental pollutants.