Machine learning assisted transparent bending-insensitive SERS substrate for reliable quantitative analysis of thiram...

Transparent and flexible surface-enhanced Raman scattering (SERS) substrates are highly suitable for rapid pesticide detection on fruit surfaces, owing to their capability for in situ detection on curved surfaces. However, achieving reproducible and stable SERS signals from flexible SERS substrates under bending conditions remains a major challenge. Additionally, integrating machine learning into Raman spectral analysis can further enhance the practicality of SERS technology. Herein, we develop a novel sensing platform that combines a transparent, bending-insensitive SERS substrate with machine learning for the rapid quantitative determination of thiram residues on fruit surfaces. The flexible SERS substrate was constructed by depositing silver nanoparticles (AgNPs) onto a polydimethylsiloxane (PDMS) twisted nanowrinkle array. Endowed with high SERS activity, exceptional transparency, and bending insensitivity, this substrate enabled reliable and consistent detection of thiram residues on apple and cherry tomato peels, with a sensitivity as low as 10^-7 M. Furthermore, a back propagation neural network (BPNN)-based machine learning model was developed to predict thiram concentrations, achieving remarkable accuracy with a coefficient of determination (R²) of 0.9954 and a root mean square error (RMSE) of 0.0985. Thus, the proposed sensing platform provides a rapid, accurate, and reliable approach for the identification and quantification of thiram in fruits.