Deep learning-integrated SERS platform for accurate identification of diverse phthalate ester subtypes

Phthalate esters (PAEs) are widely employed as plasticizers to enhance the durability, flexibility, and processability of polymer-based materials. Because PAEs do not form covalent bonds with polymer matrices, they can readily migrate into the environment. This raises significant ecological issues and human health concerns such as endocrine disruption and developmental toxicity. In this study, we developed a surface-enhanced Raman spectroscopy (SERS)-based analytical platform integrated with deep learning algorithms for the rapid and accurate identification and classification of seven representative PAEs. Plasmonic gold nanopillar (AuNP) substrates were engineered to form vertically and horizontally oriented nanogap structures that generate intense localized electromagnetic hotspots, substantially amplifying the Raman signals of adsorbed PAEs. Comprehensive spectral datasets were collected through SERS detection using a portable Raman spectrometer. Deep neural network (DNN) models trained on these data achieved robust classification performance with an accuracy of 99.4 % across all PAE species. Furthermore, SHapley Additive exPlanations (SHAP) analysis provided interpretable insights into the most discriminative Raman spectral features driving the model predictions. The platform was successfully applied to detect PAEs in commercially available consumer products at concentrations near the regulatory threshold of 0.1 % (w/w) established by regulatory bodies such as the European Union's RoHS Directive and the US CPSIA. These findings demonstrate that the integration of nanostructure-enhanced SERS and deep learning constitutes a powerful, high-throughput, and field-deployable approach for reliable PAE detection, with broad applicability in environmental monitoring, consumer product safety evaluation, and regulatory compliance.