Polycyclic aromatic hydrocarbons (PAHs) are pervasive environmental pollutants with significant health risks due to their carcinogenic, mutagenic, and teratogenic properties. Traditional methods for PAH identification, primarily relying on gas chromatography-mass spectrometry (GC-MS), utilize spectral library searches together with other techniques, such as mass defect analysis. However, these methods are limited by incomplete spectral libraries and a high false positive rate. Here, we present PAH-Finder, a data-driven workflow that integrates machine learning with high-resolution mass spectrometry (HRMS). PAH-Finder introduces a novel approach to evaluate the fragment distribution of PAH backbones in MS spectra by normalizing fragment m/z values to a 0-100% range relative to the molecular ion peak. Seven machine learning features capture PAH fragmentation characteristics, and a random forest model trained on 98 PAH spectra and 1003 background spectra achieved an F1 score of ∼0.9 in 5-fold cross validation. Additionally, PAH-Finder leverages the presence of doubly charged fragments and molecular formula prediction to enhance the identification accuracy. In a case study, PAH-Finder identified 135 PAHs, including 7 types of previously unreported PAH formulas in particulate matter samples, demonstrating a 246% increase in annotation efficiency compared to the NIST20 library search. It also identified 32 heteroatom-doped PAHs not included in the training data set, showcasing its robustness of generalization. PAH-Finder's high accuracy in detecting a broad spectrum of PAHs facilitates efficient data processing and interpretation for nontargeted analysis, enhancing our understanding of air pollution and public health protection. PAH-Finder is freely available at Github (https://github.com/FangLabNTU/PAH-Finder).