*3.5. Discussion*

In recent years, many studies on in situ detection of targets by using flexible SERS substrates have been widely reported. Alyami et al. [20] fabricated novel AgNP/PDMS composites by self-assembly of organic AgNP solutions on flexible PDMS surfaces; CV and thiram concentrations as low as 1 × <sup>10</sup>−<sup>7</sup> M and 1 × <sup>10</sup>−<sup>5</sup> M were measured on contaminated fish skin and orange peel, respectively. Chen et al. [40] detected three-pesticide residues on tomato peel based on the SERS and flexible tape. Although these methods achieved in situ detection, the sensitivity was low due to the coffee ring effect caused by the weak hydrophobicity of the PDMS and tape surface. Moreover, with the adhesive tape it was easy to destroy the structure of the nanoarray during the "paste and peel off", resulting in low stability and reproducibility. In this study, we designed the flexible SERS substrate by assembling β-CD@AuNPs on PTFE film coated with perfluorinated liquid, effectively reducing the coffee ring effect and generating a large number of hot spots. The sensitivity and stability of SERS in situ detection were competitive with the strongest results reported by the above work, and the detection process is faster and more convenient, within 1 min.

In addition, DL methods, such as CNNs, recurrent neural networks (RNNs), and generative adversarial networks (GANs), with their strong self-learning ability and excellent fitting ability, were gradually used in spectral analysis to obtain fast and intelligent quantitative or qualitative analysis [28,41]. In particular, CNNs are widely used in the modelling of spectral data by virtue of their advantages with less preprocessing and easy expansion of

network architecture. Erzina et al. [42] proposed the advanced route for express and precise recognition of normal and cancer cells by using SERS combined with a CNN, with 100% prediction accuracy. Yu et al. [43] obtained the accurate identification of six representative Vibrio species by combining label-free SERS technology with a CNN, achieving a high accuracy rate of 99.7%. However, these higher accuracy rates were obtained on the basis of building deeper and more complex networks, resulting in an increase in the number of parameters and memory footprint. In this study, the lightweight network developed based on a CNN was used for the first time to construct identification models of various PAHs and the accuracy rate was as high as 97.6%, indicating that this method could improve the computing speed and reduce the memory consumption while ensuring the model accuracy.
