Determination of Dicofol in Tea Using Surface-Enhanced Raman Spectroscopy Coupled Chemometrics

Dicofol is a highly toxic residual pesticide in tea, which seriously endangers human health. A method for detecting dicofol in tea by combining stoichiometry with surface-enhanced Raman spectroscopy (SERS) technology was proposed in this study. AuNPs were prepared, and silver shells were grown on the surface of AuNPs to obtain core–shell Au@AgNPs. Then, the core–shell Au@AgNPs were attached to the surface of a PDMS membrane by physical deposition to obtain a Au@AgNPs/PDMS substrate. The limit of detection (LOD) of this substrate for 4-ATP is as low as 0.28 × 10−11 mol/L, and the LOD of dicofol in tea is 0.32 ng/kg, showing high sensitivity. By comparing the modeling effects of preprocessing and variable selection algorithms, it is concluded that the modeling effect of Savitzky–Golay combined with competitive adaptive reweighted sampling–partial least squares regression is the best (Rp = 0.9964, RPD = 10.6145). SERS technology combined with stoichiometry is expected to rapidly detect dicofol in tea without labels.


Introduction
Tea is the second-most-popular drink around the world after water and is known for its unique aroma and texture [1,2]. As the birthplace of tea, China has a rich tea culture and is one of the world's largest tea producers and major exporters [3]. Tea is diverse and contains a variety of organic compounds and inorganic mineral elements [4]. Studies showed that regularly drinking tea can combat several chronic devastating disorders such as cancer, diabetes, cardiovascular diseases, and neurological diseases [5,6].
During the growth of tea, people use different kinds of pesticides to fight pests and diseases. Organochlorine pesticides are the most widely used class of synthetic organic pesticides so far [7]. However, this type of pesticide is chemically stable and has a long residue time in the environment, causing serious environmental pollution [8,9]. Dicofol is a highly effective organochlorine pesticide that has been widely used in agricultural pest control [10]. However, it is not easy to degrade under natural conditions. It is also known to be toxic and harmful, can cause human liver dysfunction and carcinogenic symptoms, and has long been banned in China. However, as the residue period increases, pesticides have a cumulative effect in the soil and are sometimes detected in certain water quality olecules 2023, 28, x FOR PEER REVIEW 3 of 16

Optimization of AuNPs and Au@AgNPs
The enhancement of the SERS signal is mainly caused by the local electric field enhancement of the plasma on the metal surface. Therefore, the controllability and uniformity of the particle size and distribution in nanoparticle systems is the key to constructing SERS substrates with strong stability, high sensitivity, and a uniform signal [32]. This study prepared AuNPs with different particle sizes and grew silver shells with different thicknesses on the surface of AuNPs to obtain core-shell Au@AgNPs. Figure S1a-f are TEM characterizations of AuNPs prepared by adding different volumes (150, 200, 250, 300, 350, and 400 μL) of trisodium citrate to HAuCl4 solution. It can be observed that when 150 and 200 μL of trisodium citrate are added, the synthesized AuNPs are uneven in size, with particle sizes of 80 nm and 50 nm, respectively. When the volume of trisodium citrate increased from 250 μL to 300, 350, and 400 μL, the shape of the AuNPs gradually approached a spherical shape, and the particle size decreased from 34 nm to 25, 23, and 19 nm, respectively.
The thickness of the silver shell is determined by the volumes of the AgNO3 and ascorbic acid. Figure S1g-l shows the TEM images of core-shell Au@AgNPs prepared by dropping different amounts of AgNO3 and ascorbic acid in AuNPs (34 nm) solution. It can be observed that the core-shell Au@AgNPs' center is darker, and the edges are brightly colored, which accords with the characteristics of the core-shell nanostructure [33,34]. The thickness of the Ag shell increased from 3 nm to 4.5, 5.5, 9, 11, and 15 nm when AgNO3 and ascorbic acid were added at 30, 45, 60, 75, 90, and 105 μL, respectively.

Characterization of AuNPs and Core-Shell Au@AgNPs
Trisodium citrate (150, 200, 250, 300, 350, and 400 μL) was added to HAuCl4 to prepare AuNPs with different particle sizes. The UV results are shown in Figure 2a. With the increase in sodium citrate, the peak position of the localized surface plasmon resonance (LSPR) moved blue. The LSPR band is located at 535 nm for the 80 nm AuNPs and 523 nm for the 19 nm AuNPs, with a blue shift of 12 nm. The color of the AuNPs solution

Optimization of AuNPs and Au@AgNPs
The enhancement of the SERS signal is mainly caused by the local electric field enhancement of the plasma on the metal surface. Therefore, the controllability and uniformity of the particle size and distribution in nanoparticle systems is the key to constructing SERS substrates with strong stability, high sensitivity, and a uniform signal [32]. This study prepared AuNPs with different particle sizes and grew silver shells with different thicknesses on the surface of AuNPs to obtain core-shell Au@AgNPs. Figure S1a-f is TEM characterizations of AuNPs prepared by adding different volumes (150, 200, 250, 300, 350, and 400 µL) of trisodium citrate to HAuCl 4 solution. It can be observed that when 150 and 200 µL of trisodium citrate are added, the synthesized AuNPs are uneven in size, with particle sizes of 80 nm and 50 nm, respectively. When the volume of trisodium citrate increased from 250 µL to 300, 350, and 400 µL, the shape of the AuNPs gradually approached a spherical shape, and the particle size decreased from 34 nm to 25, 23, and 19 nm, respectively.
The thickness of the silver shell is determined by the volumes of the AgNO 3 and ascorbic acid. Figure S1g-l shows the TEM images of core-shell Au@AgNPs prepared by dropping different amounts of AgNO 3 and ascorbic acid in AuNPs (34 nm) solution. It can be observed that the core-shell Au@AgNPs' center is darker, and the edges are brightly colored, which accords with the characteristics of the core-shell nanostructure [33,34]. The thickness of the Ag shell increased from 3 nm to 4.5, 5.5, 9, 11, and 15 nm when AgNO 3 and ascorbic acid were added at 30, 45, 60, 75, 90, and 105 µL, respectively.

Characterization of AuNPs and Core-Shell Au@AgNPs
Trisodium citrate (150, 200, 250, 300, 350, and 400 µL) was added to HAuCl 4 to prepare AuNPs with different particle sizes. The UV results are shown in Figure 2a. With the increase in sodium citrate, the peak position of the localized surface plasmon resonance (LSPR) moved blue. The LSPR band is located at 535 nm for the 80 nm AuNPs and 523 nm for the 19 nm AuNPs, with a blue shift of 12 nm. The color of the AuNPs solution changed from dark purple to light orange. Figure 2b shows that the Raman intensity of AuNPs gradually increases when the volume of trisodium citrate increases from 150 to 250 µL and then gradually decreases with the increase in the addition amount. AuNPs (34 nm) synthesized with a volume of 250 µL trisodium citrate had the highest Raman intensity. Figure 2c is the histogram of the particle size distribution of the 34 nm AuNPs, which shows that the distribution of the synthesized AuNPs is relatively uniform. The upper-right corner of the histogram shows the TEM image of the particle, which demonstrates that the prepared AuNPs were relatively uniform. Therefore, 34 nm AuNPs were selected for the core-shell Au@AgNPs' synthesis.

ER REVIEW 4 of 16
intensity. Figure 2c is the histogram of the particle size distribution of the 34 nm AuNPs, which shows that the distribution of the synthesized AuNPs is relatively uniform. The upper-right corner of the histogram shows the TEM image of the particle, which demonstrates that the prepared AuNPs were relatively uniform. Therefore, 34 nm AuNPs were selected for the core-shell Au@AgNPs' synthesis. To explore the effect of silver-shell thickness on the core-shell Au@AgNPs' LSPR characteristics, we collected the UV spectra of the core-shell Au@AgNPs' solutions with different silver-shell thicknesses, as shown in Figure 2d. There were two peaks in the Au@AgNPs' Raman spectrum when AgNO3 and ascorbic acid were added at 30 μL. The peak at 374 nm is the Ag shell peak, and the peak at 528 nm is the Au nucleus peak. With the thickening of the Ag shell, the LSPR of the Ag shell gradually strengthens, and the LSPR of the Au nucleus gradually weakens or is even completely shielded. It can be seen from Figure 2d that the peak around 528 nm gradually disappears, while the peak around 374 nm gradually becomes obvious and redshifted. When the Ag shell thickness exceeds a certain value (9 nm), the Au nuclear LSPR disappears, leaving only the LSPR of the Ag To explore the effect of silver-shell thickness on the core-shell Au@AgNPs' LSPR characteristics, we collected the UV spectra of the core-shell Au@AgNPs' solutions with different silver-shell thicknesses, as shown in Figure 2d. There were two peaks in the Au@AgNPs' Raman spectrum when AgNO 3 and ascorbic acid were added at 30 µL. The peak at 374 nm is the Ag shell peak, and the peak at 528 nm is the Au nucleus peak. With the thickening of the Ag shell, the LSPR of the Ag shell gradually strengthens, and the LSPR of the Au nucleus gradually weakens or is even completely shielded. It can be seen from Figure 2d that the peak around 528 nm gradually disappears, while the peak around 374 nm gradually becomes obvious and redshifted. When the Ag shell thickness exceeds a certain value (9 nm), the Au nuclear LSPR disappears, leaving only the LSPR of the Ag shell. The core-shell Au@AgNPs' solution changes from orange to light yellow. Figure 2e shows that when the addition of ascorbic acid and AgNO 3 increased from 30 to 75 µL, the Raman intensity of the core-shell Au@AgNPs gradually increases and then decreases with the increase in the addition amount. When the amount of ascorbic acid and AgNO 3 is 75 µL, the thickness of silver shell is 9 nm, and the Raman strength of the core-shell Au@AgNPs is the highest. Figure 2f shows that the size distribution of the core-shell Au@AgNPs' solution's particles at this thickness is relatively uniform. The TEM image at the upper right shows that the synthesis of the particles is relatively uniform. Subsequent operations use the size of the core-shell Au@AgNPs.

Characterization of the Au@AgNPs/PDMS Substrate
Before collecting the Raman spectra of the target substance, the Raman spectra of the blank tin foil, blank PDMS, and blank Au@AgNPs/PDMS substrate were collected, and the results are shown in Figure S2a,b. The intensity of the blank tin foil, blank PDMS, and blank Au@AgNPs/PDMS substrates is less than 100, and the influence on the subsequent detection of the target substances is negligible. The TEM results of the Au@AgNPs/PDMS substrate are shown in Figure 2g, where the core-shell Au@AgNPs is spherical and uniformly distributed on the PDMS membrane. The diameter of a single core-shell Au@AgNP is about 45 nm. To evaluate the SERS-enhanced effect of the Au@AgNPs/PDMS substrate, 4-ATP solution (10 −6 mol/L) was dripped onto the tin foil and SERS-enhanced substrate, and the Raman spectrum was collected ( Figure S2b). The band intensity of the 4-ATP located around 1078 cm −1 is 278 on the tin foil and 41,702.3 on the reinforced substrate. It can be seen that the Au@AgNPs/PDMS substrate has a good SERS-enhancement effect. In addition, the original Raman spectra of the 4-ATP on the tin foil showed significant Raman peaks at 473, 876, 1087, and 1588cm −1 . They belong to the C-C out-of-plane coupled vibration (473 cm −1 ), the C-S out-of-plane coupled vibration (876 cm −1 ), the C-S in-plane tensile vibration (1087 cm −1 ), and the C-C in-plane tensile vibration (1588 cm −1 ), respectively [35,36]. When 4-ATP4-ATP is attached to Au@AgNPs/PDMS, a Ag-N bond is formed, the out-of-plane vibration is weakened, and the Raman peaks at 473 cm −1 and 876 cm −1 disappear [37].

Reproducibility and Stability Analysis
To verify the reproducibility of the Au@AgNPs/PDMS substrate, 5 µL 4-ATP (0.01 mol L −1 ) solution was dropped onto the SERS substrate, and 15 Raman spectra were collected. The relative standard deviation was calculated at the spectral intensity of 1078 cm −1 . The relative standard deviation (RSD) is 3.94% (shown in Figure 2h), indicating the good reproducibility of the SERS substrate. After storing the Au@AgNPs/PDMS substrate for 10 days, 20 days, and 30 days, 5 µL of 4-ATP (0.01 mol L −1 ) solution was dropped to acquire the Raman spectra. Figure 2i shows that compared with the newly prepared substrate, the Raman intensity of the SERS base (1078 cm −1 ) slightly decreased after 30 days of storage. The results indicated that the substrate stability was excellent (RSD = 3.08%). The results showed that the stability of the Au@AgNPs/PDMS substrate is good and can be stored for at least one month.

Sensitivity of Au@AgNPs/PDMS
To further explore the sensitivity of the Au@AgNPs/PDMS substrates, the Raman spectra of 4-ATP solutions with gradient concentrations of 10 −6 , 10 −7 , 10 −8 , 10 −9 , 10 −10 , and 10 −11 mol/L were collected. Figure 3a shows that as the 4-ATP concentration decreases, the Raman intensity also decreases, and the SERS signal is still visible when the 4-ATP concentration is as low as 10 −11 mol/L (the SERS intensity is 822). This indicates that the Au@AgNPs/PDMS substrate has a high sensitivity. Figure 3b shows the relationship between the Raman intensity and logarithmic concentration of 4-ATP at 1078 cm −1 . A high correlation between the 4-ATP Raman intensity and concentration is indicated (R 2 = 0.996). In addition, Table 1 compares the results of the 4-ATP detection by different corenanoparticles' SERS. In general, precious metals of different shapes such as Ag and Au, metallic oxides such as TiO2 and SiO2, and polymer PDMS are used as the SERS's active to detect 4-ATP. The limit of detection (LOD) of these substrates is not less than 10 −10 m In addition, Table 1 compares the results of the 4-ATP detection by different core-shell nanoparticles' SERS. In general, precious metals of different shapes such as Ag and Au, non-metallic oxides such as TiO 2 and SiO 2 , and polymer PDMS are used as the SERS's active base to detect 4-ATP. The limit of detection (LOD) of these substrates is not less than 10 −10 mol/L, while the introduction of metal backbones can be up to 10 −12 mol/L. The LOD was calculated according to the following formula: LOD = 3σ/k. The SERS spectrum of the blank substrate was repeated 10 times, and the standard deviation of the band intensity was calculated, which was σ. k was the slope of the standard curve The Au@AgNPs/PDMS substrate prepared in this study has a high sensitivity, and the LOD of 4-ATP is as low as 0.28 × 10 −11 mol/L. Simultaneously, the substrate preparation process is simpler and faster than that of the metal skeleton, which has better commercial application prospects.  Figure S2c shows the Raman spectra of the dicofol samples on tinfoil and PDMS film, and the difference between the two can be ignored. There are few related studies on dicofol, so the reference materials are limited. Density-functional theory (DFT) was applied to calculate the standard Raman spectra of dicofol to confirm the theoretical distribution of the Raman peaks and to verify the accuracy of the experimental spectra [45]. The molecular configuration of dicofol was constructed using Gauss View 6.0. The theoretical Raman spectra of dicofol were calculated using the B3LYP/6-311 G (D, P) group of the B3LYP method of DFT. The calculation results of the dicofol Raman frequency were checked on the Gauss View 6.0 software platform, and the corresponding correction factors were used for correction. The standard Raman spectrum of dicofol calculated by DFT is shown in Figure S2c. The spectral curves of eight gradient concentrations of dicofol are shown in Figure 3c. The higher the content of dicofol is, the higher the Raman intensity of the tea sample is. The Raman spectra of dicofol on tin foil and Au@AgNPs/PDMS show that Raman peaks exist at 489, 888, 1055, 1094, and 1463 cm −1 and 489, 619, 709, and 1261cm −1 , respectively. The corresponding vibration forms are explained in Table S1. The standard curve was constructed with the band intensity at 489cm −1 . Figure 3d shows that the coefficient of determination of the dicofol content and Raman intensity is 0.9791. The LOD of dicofol on the Au@AgNPs/PDMS substrate prepared in this study was 0.32 ng/kg.
Environmental fluctuations and tea sample complexity have a certain impact on spectral intensity, so the spectral data are processed by stoichiometry to improve the model reproducibility and prediction accuracy. Figure 3e shows all the raw Raman spectra of dicofol under eight concentration gradients (15 spectra per collected gradient). Figure 3f shows the PLS modeling results of the original Raman spectrum. On this basis, five algorithms (SG, SNV, MSC, first derivative, and second derivative) were used to preprocess the Raman spectral data. The spectral shifts were in the range of 300-2000 cm −1 , and each spectrum included 667 variables. Table 2 shows the results of the different pretreatment models of dicofol. SNV and MSC show poor modeling results. The noise removal ability of these two preprocessing algorithms is not well-demonstrated in this experiment. The derivative treatment did not significantly optimize the original spectra or even slightly reduce them. After smoothing, the PLS modeling results were the best (Rp = 0.9906, and RMSECP = 4.8120), and the residual predictive deviation (RPD) value was 7.2341, indicating that the prediction ability of the SG-PLS model was better. Therefore, the subsequent variable selection algorithm was based on the SG smoothing preprocessing.

Variable Selection and Models Development
After data preprocessing, we carried out variable selection to improve the prediction accuracy of the model. In this study, four variable selection algorithms (CARS-PLS, Si-PLS, SPA-PLS, and UVE-PLS) were used to process the data.

CARS-PLS Model
After the SG smoothing of the raw spectral data, the CARS-PLS algorithm was used for spectral variable selection. Figure 4a(I) shows the periodic variation rule of RMSECV, which reaches the minimum value of 3.008 in the 30th cycle. After removing the irrelevant variables from the model, the RMSECV value gradually increases. Figure 4a(II) shows the periodic variation rule of the regression coefficient. The regression coefficient values for the first 15 cycles were close to zero. After that, the regression coefficient of the strong competition variable increases, while that of the weak competition variable decreases. Figure 4a(III) shows the trend of the number of variables, which decreases faster in the first six periods and gradually slows down afterward. This change indicates that the refinement of the variable selection is realized by quickly eliminating any redundant variables, which are reduced from 667 variables to 21. The variable-selecting and modeling results of the CARS-PLS model are shown in Figure 4a,b.

Si-PLS Model
The model construction process of Si-PLS is as follows. The spectral data are divided into several equidistant intervals, and then the correlation information of the interval variables and analytes is recombined to establish a stable prediction model. In this study, the entire spectrum is divided into 10 subintervals. By recombining the 2nd, 3rd, 4th, and 10th subintervals, the optimal PLS model is obtained (Rp = 0.9910), which is reduced from 667 variables to 267. The variable-selecting and modeling results of the Si-PLS model are shown in Figure 4c,d.

SPA-PLS Model
SPA randomly selects a variable in the data matrix, calculates the projection of the random variable on the remaining variables, and selects the variable with the least-redundant information content to improve the modeling effect by projecting the information. The variable selection results of SPA are shown in Figure 4e, nine characteristic variables are selected, and finally the prediction model of dicofol in tea is established. Compared with the model prediction accuracy of PLS, that of SPA-PLS is reduced, and the modeling results are shown in Figure 4f.

UVE-PLS Model
The UVE algorithm eliminates invalid variables by adding artificial noise variables to the sample variables and cross-verification based on PLS regression coefficients, thus improving the modeling effect. The maximum modeling principal component number is 20, and the artificial noise variable is 160. Figure 4g shows the UVE variable selection plot for dicofol, which is reduced from 667 variables to 45. The modeling results of the UVE-PLS model are shown in Figure 4h. Table S2 clearly shows the modeling results of the four variable selection algorithms. The RPD value is used to evaluate the prediction ability of the four models, and the RPD value is always greater than 4. CARS-PLS achieves the best prediction (Rp = 0.9964, and RMSEP = 2.9268), followed by Si-PLS, UVE-PLS, and PLS, while SPA-PLS has the worst prediction. The number of variables is reduced from 667 to 21, and the model is obviously simplified. The CARS-PLS prediction effect is the best, which proves that the variable selection algorithm can eliminate redundant variables and improve the modeling effect.

Discussion
The residue of dicofol in tea leaves is harmful to human health. A method for detecting dicofol in tea by combining stoichiometry with surface-enhanced Raman spectroscopy (SERS) technology was proposed in this study. The core-shell structure can combine the advantages of the two metals, while the PDMS film has good flexibility and strong corrosion resistance. Therefore, we prepared core-shell Au@AgNPs with excellent enhancement properties, high stability, and high sensitivity. Then, core-shell nanoparticles were attached to the surface of the PDMS by physical deposition, and the Au@AgNPs/PDMS substrate was prepared.
A Au@AgNPs/PDMS substrate has excellent stability and reproducibility and can detect a 10 −11 mol/L 4-ATP signal, showing high sensitivity. This may be due to the high enrichment of the core-shell Au@AgNPs on the PDMS membranes resulting in more SERS "hot spots". This substrate was used for the SERS detection of dicofol in tea, and a standard curve was established. R 2 = 0.9791, indicating a good correlation between the Raman intensity and sample concentration.
Considering the influence of noise and sample complexity on spectral intensity, the spectral data were processed by chemometrics. The algorithm modeling effect of SG smoothing combined with CARS-PLS was the best (Rp = 0.9964, RMSEP = 2.9268, and RPD = 10.6145). The use of chemometric algorithms effectively removed the influence of redundant information and improved the modeling effect.
In summary, the highly efficient and sensitive SERS detection method for dicofol in tea proposed in this study had good repeatability and was easy to operate. Therefore, it has broad application prospects in medicine, food safety, the life sciences, and other fields.

Chemicals Reagents and Materials
The gold nanoparticles were prepared from chloroauric acid (AuCl 3 ·HCl·4H 2 O) and trisodium citrate solution. Silver nitrate (AgNO 3 ), ascorbic acid (AA), and prepared gold nanoparticles were used to synthesize gold and silver core-shell nonsolutions. Polydimethylsiloxane (PDMS) films were retreated with Piranha solution (H 2 SO 4 :HNO 3 = 3:7 (v/v)) and APTES solution (purchased from Zhongke Experimental Materials Co., Ltd., Hefei, China). 4-ATP is a signaling molecule. Dissolve dicofol in an ethanol solution. Reagents used are of analytical grade purity and are intended for direct use unless otherwise stated. Wash each glassware with ultra-pure water before usage. Green tea was purchased from a supermarket in Jingkou District, Zhenjiang City.

Instrumentation
The substances used in the experiment were weighed using a microanalytical balance (Sartorlus, Shanghai, China), and the green tea powder sample was digested with a microwave digester (MARS 6) (CEM, Matthews, NC, USA). AuNPs and core-shell Au@AgNPs were characterized with ultraviolet spectrophotometer (UV-Vis) (Agilent Technologies Inc., Santa Clara, CA, USA), laser particle size analyzer (Particle Size Analyzer, Malvern, UK), scanning electron microscope (SEM) (JEOL Ltd., Tokyo, Japan), and JEM-2100 HR transmission electron microscopy (TEM) (JEOL Ltd., Japan). Spectral data of 4-ATP and dicofol were collected and processed by confocal micro-Raman spectroscopy (XploRA PLUS, HORIBA, Palaiseau, France) The excitation wavelength was 638 nm, and the objective lens was 10×. The laser power was 400 mW, the cumulative number of sample collection was 1, and the collection time was 1s. MATLAB 2016b (MathWorks, Natick, MA, USA) was used to process the spectral data. Gauss View 6.0 (Gaussian Inc. Chicago, IL, USA) was used to construct the molecular configuration of dicofol.

Synthesis of Flexible Substrates for Au@AgNPs/PDMS
First, the gold nano solution was prepared by seed growth method: take 30 mL of ultrapure water in a 100 mL beaker, add 425 µL of HAuCl 4 (5 g/L) solution, heat to boiling state, hold for 2 min under the condition of magnetic stirring at 1200 rpm, then add 150, 200, 250, 300, 350, and 400 µL of trisodium citrate solution with a mass fraction of 1%, and continue to heat until the color of the liquid no longer changes. Allow the liquid to cool to room temperature and store in a refrigerator at 4 • C. At this point, gold nanoparticle solutions of different diameters were prepared. Gold nanoparticles (34 nm) with the best SERS-enhancement effect were selected for subsequent experiments.
Then the preparation of gold and silver core-shell nanoparticles was carried out: 6 mL of optimized gold nanoparticle solution (34 nm) was taken in a test tube, and 30, 45, 60, 75, 90, and 105 µL of ascorbic acid solution with a concentration of 10 mmoL/L were added after 10 min of ultrasonication. After shaking at 1000 rpm for 2 min, add the same concentration of AgNO 3 solution. Refrigerate at 4 • C after 7 min of continuous shaking.
After that, the amination functionalization of the PDMS membrane was carried out. Piranha solution (H 2 SO 4 :HNO 3 = 3:7 (v/v)) was added to the PDMS film. Shake for 10-30 s, take out the PDMS film, clean with ultrapure water, blow dry with high-purity nitrogen, repeat this step 3~4 times, and then put it in a clean surface dish. Hydroxylated PDMS films were soaked in APTES ethanol solution (5%, v/v) for 3 h (70 • C). Amine-functionalized PDMS film was obtained after washing with ultra-pure water and drying.
Finally, the core-shell Au@AgNPs were deposited in aminated PDMS by physical method to form Au@AgNPs/PDMS film.

Feasible Characterization of the Composite Raman Substrate
In order to determine whether the substrate material would affect the subsequent detection of substances, the Raman spectra of blank tin foil, blank PDMS, and blank Au@AgNPs/PDMS substrates were collected. The original Raman spectra of 0.1 mol/L dicofol standard solution on tin foil and on blank PDMS film were collected. In order to prove that the SERS substrate had strong SERS-enhanced effect, 10 −6 mol/L Raman signaling molecule 4-ATP was added into dicofol solution. The Raman spectra of the mixed solution on the tin foil and SERS substrate were collected, and the changes in Raman spectral intensity before and after using the enhanced base were compared.

Sample Preparation
A bag of ordinary green tea was purchased from Jiangsu University supermarket. The powdered green tea samples were stored in the refrigerator at 4 • C for the subsequent experimental operation. First, 0.200 g tea powder was added with 9 mL concentrated HNO 3 and different concentrations of dicofol standard solution, standing for digestion for 12 h. The tea samples were digested according to the general operating rules of microwave digestion instrument. The digestion conditions were set as follows: the digestion temperature was 190 • C, the heating time as 20 min, and the constant temperature time was 10 min. After digestion, the digestion tank should be taken out after cooling to room temperature. The digestion solution was clear and transparent without precipitation [20]. The digestion solution was filtered by 0.22 µm membrane and stored at 4 • C.

SERS Detection of 4-ATP and Dicofol
SERS spectra of experimental samples were acquired using a confocal Raman microscopy system, and spectral data were processed using LabSpec6 software (Lab Spec_6_5_2) inside the system. The excitation power used in the experiment was 400 mW, the cumulative number of sample acquisitions was 1, and the acquisition time was 1 s. All Raman spectra were obtained under a 10× objective with an excitation wavelength of 638 nm. Using 4-ATP as the signaling molecule, the enhanced effects of the prepared gold nanoparticles, core-shell Au@AgNPs, and Au@AgNPs/PDMS were investigated.
To analyze the SERS-enhanced effect of gold nanoparticles of different sizes, 10 µL of 4-ATP (0.1 mmol/L) solution was mixed with 990 µL of AuNPs. After 10 min, take 5 µL of the mixed liquid for Raman detection. In order to explore the effect of silver-shell thickness on the SERS performance of core-shell Au@AgNPs, 10 µL of 4-ATP (0.1 mmol/L) solution was mixed with 990 µL of core-shell Au@AgNPs. After 10 min, take 5 µL of the mixed liquid for Raman detection. Raman spectra of 4-ATP in six AuNPs with different particle sizes and six core-shell Au@AgNPs with different silver-shell thicknesses were collected.
An Au@AgNPs/PDMS substrate with a size of 6 mm × 6 mm was placed on tin foil for SERS testing. To investigate the sensitivity of the Au@AgNPs/PDMS substrate, different concentrations of 4-ATP solution (10 −6 , 10 −7 , 10 −8 , 10 −9 , 10 −10 , 10 −11 , and 10 −12 mol/L) were added to the substrate for SERS detection. To investigate the reproducibility of the Au@AgNPs/PDMS substrate, 10 µL of 4-ATP (0.1 mmol/L) solution was dropped on the substrate, and 15 points were uniformly collected in mapping mode in a region of 3 mm × 2 mm, and the standard deviation was calculated. To explore the stability of the Au@AgNPs/PDMS substrates, three substrates were stored in closed Petri dishes, and Raman measurements were performed dropwise with 5 µL of 4-ATP (0.1 mmol/L) solution at 10, 20, and 30 days. Each substrate was uniformly collected from 25 points in the mapping mode in a region of 3 mm × 3 mm, and the intensity of the 25 points was averaged. Finally, the standard deviation of the three substrates was calculated.
The SERS spectra of dicofol standard solution on tin foil were collected, and the SERS spectra of eight gradient concentrations (1,10,20,30,50,70,90, and 100 ng/kg) spiked tea samples on the Au@AgNPs/PDMS substrate were collected. The calculation method of dicofol content in tea samples is described in the supplementary materials.
Due to the complexity of tea samples and the variability of the external environment, the acquired spectra were affected by fluorescence interference and had a certain degree of baseline drift. Therefore, it was necessary to perform baseline processing and data smoothing on the acquired spectra to improve the spectral quality. The de-baseline parameters selected for this study were set as follows: background fit was polynomial, degree = 7, and maximum points = 256. In addition, control the data smoothing effect by changing the function type and parameters [46]. The function type selected for this study was Denoise, and the parameters were set as follows: degree = 1, and size = 1.

Spectra Preprocessing
To minimize baseline drift and interference of redundant background information, Raman spectra were preprocessed by Savitzky-Golay (SG), spectral multivariate scattering correction (MSC), variable normalization (SNV), first derivative, and second derivative. These preprocessing methods can better remove the noisy data in the spectrum and eliminate the baseline drift, which is significant to improve the prediction effect and generalization ability of the model. In this study, five methods were selected to preprocess the original Raman spectral data of dicofol and establish PLS model.

Variable Selection
Multivariate calibration methods can improve spectral modeling by selecting spectral variables. In this study, four variable selection algorithms were used to process the spectrum data of dicofol by synergy interval partial least squares method (SI-PLS), uninformed variable elimination algorithm (UVE-PLS), successive projection algorithm (SPA-PLS), and competitive adaptive reweighted sampling (CARS-PLS). SI decomposes all variables into several equidistant subintervals and then combines 1, 2, or more subintervals to construct the model. The joint interval can compensate for the lack of information of a single interval and improve the accuracy of the model [47,48]. The UVE algorithm is based on PLS and can eliminate uninformative variables from the data [49]. SPA is a forward characteristic variable selection method, which can select the combination of variables with the leastredundant information content and least collinearity [50]. CARS treats each variable as an independent individual through adaptive weighted sampling and retains the variable with a larger weight as a new subset, which can effectively remove irrelevant variables and reduce the influence of collinear variables [51,52].

Model Assessment
Five pretreatment algorithms based on PLS model and four variable selection algorithms were combined to establish the best prediction model for dicofol concentration. The 120 spectral data (667 independent variables) were divided into 72 calibration sets and 48 test sets. The calibration set was used to build the partial least squares model, and the test set was used to verify the generalization ability of the evaluation model. The model was evaluated by correlation coefficient of calibration (Rc), root mean square error of calibration (RMSEC), correlation coefficient of prediction (Rp), and root mean square error of prediction (RMSEP). The smaller the RMSEC and RMSEP values were, the closer R was to 1, and the more accurate and stable the model was. Algorithmic processing of all SERS data took place in MATLAB 2016b.

Conclusions
This study proposed a method for detecting dicofol in tea by combining stoichiometry with SERS technology. The core-shell Au@AgNPs were prepared using ascorbic acid as the reducing agent and then attached to the aminationized PDMS membrane using physical deposition methods to obtain a Au@AgNPs/PDMS substrate. Under the condition of an excitation wavelength of 638 nm, the Raman spectra of the 4-ATP and dicofol gradient concentrations on the SERS substrate were collected, and a linear relationship between the spectral intensity and gradient concentration was established. The LOD for 4-ATP was 0.28 × 10 −11 mol/L for this substrate and 0.32 ng/kg for dicofol in tea, which is much lower than the safe limit of dicofol in tea (200 ng/kg). The modeling effects of five preprocessing algorithms and four variable selection algorithms based on PLS model were compared. The algorithm combining SG and CARS-PLS modeled the dicofol data the best (Rp = 0.9964, and RPD = 10.6145). Combined with stoichiometry, SERS technology can be used for the efficient and sensitive detection of dicofol in tea.

Data Availability Statement:
The data presented in this study are available on request from the corresponding authors.