Fast Quantification of the Mixture of Polycyclic Aromatic Hydrocarbons Using Surface-Enhanced Raman Spec-troscopy Combined with PLS-GA-BP Network

Yan Xia, shi xiaofeng, and ma jun

Fast Quantification of the Mixture of Polycyclic Aromatic Hydrocarbons Using Surface-Enhanced Raman Spec-troscopy Combined with PLS-GA-BP Network

Yan Xia, shi xiaofeng*, and ma jun

Optics and Optoelectronics Laboratory of Qingdao, Ocean University of China, Qingdao 266100, China

To realize the fast and accurate quantitative analysis of the mixture of polycyclic aromatic hydrocarbons (PAHs), surface-enhanced Raman spectroscopy (SERS) coupled with multivariate calibrations were employed. In this study, three kinds of calibration algorithms were used to quantitative analysis of the mixture of naphthalene (Nap), phenanthrene (Phe), and pyrene (Pyr). Firstly, partial least squares (PLS) algorithm was used to select characteristic variables, then the global search capability of genetic algorithm (GA) was used for the determining of the initial weights and thresholds of back propagation (BP) neural network so that local minima was avoided. The PLS-GA-BP model exhibited superiority to quantify PAHs mixture, which achieved2=0.9975, 0.9710, 0.9643, ARE=10.07%, 19.28%, 16.72% and RMSE=13.10, 5.40, 5.10nmolL−1for Nap, Phe, Pyr (in the PAHs mixture) concentration prediction respectively. The forecast error, ARE and RMSE have been reduced more than 50% and 60% respectively compared with the whole spectral BP model. The study indicates that accurate quantitative spectroscopic analysis of the mixture of PAHs samples can be achieved through the combination of SERS technique and PLS-GA-BP algorithm.

polycyclic aromatic hydrocarbons (PAHs); surface enhanced Raman spectral (SERS); back propagation (BP) algorithm; multi-component quantitative analysis

1 Introduction

Polycyclic aromatic hydrocarbons (PAHs) are one of the persistent organic pollutants, which attract the atten- tion of regulators duo to their highly carcinogenic, tera- togenic, mutagenic and lipophilic (Menzie., 1992). PAHs can do harm to human health seriously even with low concentration in environment. Moreover, since any incomplete combustion of carbohydrates may produce PAHs, they are widespread in the environment (Simcik., 1999; Zhang., 2009, Akdoan., 2020). There- fore, detection and monitoring of these hazardous compounds are imminent.

Several spectral measurements such as high performance liquid chromatography (HPLC) (Joe., 1982; Wenclawiak., 1992), gas chromatography-mass spectrometry (GC-MS) (Orecchio., 2009), fluorescence spectrometry (Zhang., 2009; Han., 2018), and so on, have been applied to PAHs detection and identification. However, these analytical techniques either need very laborious sample handling process (GC-MS, LC) or have difficult in mixture distinguish (fluorescence spectroscopy), which hinder their field application and quantitative analysis of PAHs.

As a high precision trace detection technique, SERS is regarded as the fingerprint spectrum of substances (Qian., 2008; Dasary., 2009; Langer., 2019), and it has great potential for application in various fields due to its unique advantages, such as excellent molecular specificity, simple sample pretreatment and ultra-high sensitivity as well, especially in organic pollutants field detection such as PAHs (Shi., 2015). However, SERS is more frequently used as a qualitative tool rather than quantitative one because the SERS signals of analyte not only depend on the analyte concentration, but also the physical properties of enhanced substrate, such as the particle size and shape of colloids, the degree of aggregation, the intensity and focusing of laser excitation source (Chen., 2015).

Over the years, lots of efforts have been addressed to the issue of SERS detection, and most concerning is to improve the sensitivity of SERS substrates (Tripp., 2008; Gu., 2013; Yin., 2019), but fewer studies focus on the accuracy of rapid quantitative analysis of organic pollutant mixtures. Moreover, due to the influence of the collection efficiency, changes in the detection conditions and competitive adsorption between various substances, it is difficult to obtain a repeatable SERS signal (Lin., 2018). In general, the SERS technique is full of challenges for rapid quantitative analysis (Zhao., 2015), especially for multi-component organic substances. Nevertheless, for on-site real-time detection, ra- pidly analysis of the data is necessary. Therefore, to reduce the influence of human factors and improve the efficiency and accuracy of data analysis, using advanced data mining techniques to provide support for scientific decision-making has attracted increasing research attention. For most study, there is nonlinear relationship between concentration and spectrum intensity. To improve the quantitative accurate under in this situation, many methods are adopted, such as optimized artificial neural network (ANN) (Li., 2019; Wang., 2019), optimized partial least squares (PLS) (Chen., 2019) and support vector regression (SVR) (Fan., 2015),.

BP neural network is one of the most popular ANN methods (Ma., 2005; Liu., 2017),which shows a great promise in the quantitative analysis of spectral data (Jun., 2019). However, the existing BP neural network has some drawbacks, such as low convergence rate (Chun., 2018; Jia., 2019)and being difficult to devise suitable network structure (Xie., 2007). In addition, for actual SERS regression analysis, high precision is the prerequisite for quantitative analysis, and the characteristic variables of model input data play an important role for quantitative accuracy.Therefore, to obtain high precision prediction, some optimization algorithms such as PLS, GA are needed to select the input parameter and optimize the algorithm.The original data can be compressed according to the output variables by PLS, which can improve the prediction accuracy of the nonlinear model compared with principal component analysis (PCA) (Li., 2007). And the precision and prediction robustness of a model can be improved by combining GA and BP algorithm (Li., 2015).

In this study, SERS spectra of 31 mixed solutions of three PAHs at different concentrations were obtained using a portable Raman system. An advanced high-preci- sion algorithms were established and used in the concentration prediction of the mixture. In which,according to the value of average relative error (ARE), root-mean-squa- re error (RMSE) and correlation coefficient (2), PLS was used to select the number of principal components of each substance with one hidden layer and one neuron BP network, and many experiments were conducted to determine the deep structure of BP neural network. And to accelerate the convergence speed and improve the accuracy for multicomponent PAHs SERS spectra quantitative analysis, GA was introduced into the PLS-BP algorithm, with full use of its global search capability and the local search capability of BP algorithm. The experiments demonstrated that the PLS-GA-BP algorithm performed well in the regression analysis of PAHs mixture.

2 Materials and Methods

2.1 Chemicals and Instrumentation

Nap (C10H8, 97%), Phe (C14H10, 97%), Pyr (C16H10, 97%) were supplied by Sigma-Aldrich Co. (USA). Hydrogen tetrachloroauric (III) trihydrate (HAuCl4, 99.9%), trisodium citrate (Na3C6H5O7·2H2O, 99.5%), methanol (CH3OH, 99.7%), sodium hydroxide (NaOH, 99%), hydrochloric acid (HCl, 99%) and sulfuric acid (H2SO4, 98%) were from Sinopharm Group Chemical Reagent Co. Ltd. (China).

The portable spectrometer (QE65000) was purchased from Ocean Optics (USA). The excitation light source, 785nm semiconductor laser (FC-785-500-MM) was ma- nufactured by Xilong Optoelectronics Technology Co., Ltd. (China).

2.2 Preparation of Gold Colloid

The gold colloid was synthesized by reducing chloroauric acid with sodium citrate as a reductant using the method of Frens (1973). The specific process can refer to the reference(Shi., 2012).

2.3 Sample Preparation and SERS Measurement

2.3.1 Sample preparation

The analytes used in this study, Nap, Phe, and Pyr are three main PAHs in the surface seawater of Qingdao coastal area (Li., 2012). Stock solutions of them were prepared at a concentration of 40μmolL−1(Nap), 50 μmolL−1(Phe) and 40μmolL−1(Pyr) respectively with methanol as the solvent. Subsequently, a series of PAHs solutions at different concentrations (500, 100, 50nmol L−1for Nap; 100, 50, 20, 10, 4nmolL−1for Phe; 100, 50, 20, 10, 4, 1nmolL−1for Pyr) were obtained by diluting with deionized water. Thirty- one PAHs mixed solutions with different concentrations were prepared by mixing the prepared three PAH solutions with different concentrations of Nap, Phe and Pyr in equal volumes. Considering in the human error of the prepared solution, two mixed solution of each concentration was prepared. Thus, sixty- four samples were obtained.

2.3.2 SERS measurement

Existing studies have shown that the aggregation degree of gold nanoparticles (Au-NPs) in the colloids solution can be increased by adjusting the pH value of the gold (Au) colloids, which could obtain more high-sensitivity SERS detection. With Au-NPs adjusted by NaOH as enhanced substrate, SERS technology coupled with chemometric algorithms such as full spectrum BP algorithm, PLS-BP, PLS-GA-BP was endeavored to quantify multi-component PAHs.

The quantitative strategy of the proposed method for PAHs mixture is schematically depicted in Fig.1. Briefly, PAHs molecules were combined with Au-NPs to form Au-NPs-PAHs by adding PAHs (Nap, Phe and Pyr) molecules to the uniformly distributed Au NPs colloid. When NaOH was added to adjust the pH of the solution, the aggregation of Au-NPs and PAHs molecules was formed. Thus the number of SERS hot spots was increased, and the detection sensitivity was improved.

SERS detection was performed as follows: PAHs mixture and Au-NPs (size about 57nm, pH=6) were mixed with themixture:Au-NPs=3:1. Then, adjust the pH of the mixed solution to 13 with NaOH. Finally, the average SERS spectra of three replicated spectra with an integration time of 10 s and the laser power of 180mW (on sample) were recorded.

Fig.1 Schematic illustration of fast quantification of PAHs mixture by SERS coupled GA-PLS-BP algorithm.

2.4 Algorithms

PLS is a multivariate statistical analysis method, which combines the advantages of the three algorithms: principal component analysis (PCA), canonical correlation analysis (CCA), and multiple linear regression (MLR), and it focuses on the effect of the principal component factor and spectral treatment (Chen., 2015; Kutsanedzie., 2017).

ANN is an algorithm which offers an alternativeway to simulate complex and ill-defined problems. BP neural network is a typical ANN algorithm that has been widely used in many fields, such as medical signal analysis and quantitative analysis of spectra. With proper parameters, BP network can approximate any nonlinear function (Yu., 2014).

GA is a simple, fault-tolerant and fast simulated evolutionary process algorithm, which follows the principle of evolution and takes the good individual evolution as the optimal solution. It can achieve the optimal initial weight and threshold for the BP network.

3 Results and Discussion

3.1 Characterization of Analytes

The typical SERS spectra of Nap, Phe, Pyr and their mixture in the same concentration are shown in the Fig.2. As the picture shows, the SERS peaks intensity and location for the mixture has changed compared with the peaks of single product, although their concentration are same. The changed ratio of each signal was different even for one constituent. Compared with the Nap solution, the Raman intensities of the mixture at 753cm−1and 1374 cm−1(for Nap) reduced 58.70% and 69.21% respectively. And since different solubilities and adsorption properties between different kinds of PAHs, this reduction is even more different among different issues. The Raman shifts at 1374cm−1(for Nap), 702cm−1(for Phe) and 588cm−1(for Pyr) reduced 69.21%, 41.34% and 27.73% correspondently, which testified that tetracyclic Pyr (with poorer water solubility) had better combination with Au-NPs compared with Nap and Phe (consisted of two and three benzene rings respectively). Therefore, the peaks intensity of Pyr reduced least in the mixture compared with the pure solution. Furthermore, the peaks at 1598cm−1(for Phe), 1606cm−1(for Pyr) overlapped to 1600cm−1 in the mixture.All these differences make the quantitative analysis of mixture more difficult.

Fig.2 Typical SERS spectra of 100nM Nap, Phe, Pyr and their mixture. (a), Nap; (b), Phe; (c), Pyr; (d), mixture.

Due to the competitive adsorption between substances, the changes of one other substance in the solution will cause the spectrum change of the fixed substance. SERS spectra of PAHs mixture (Nap, Phe, Pyr) are shown in Fig. 3(A), where the concentration of Pyr changes while the concentration of Nap and Phe does not change. The dependences of the intensities at 588, 1054, 1230 and 1498 cm−1bands on the concentration of Nap are plotted in Fig. 3(B). As can be seen, the SERS intensities of Pyr increased with the increasing concentration. However, the relationship between concentration and peak intensities was no longer linear. The regression equations and correlation coefficients are presented in Table 1. The non-linear correlation between Raman intensity and concentration was noted.

Fig.3 (A) SERS spectra of different concentrations of Pyr with the same concentrations of Nap (167nmolL−1) and Phe (33 nmolL−1). (B) Dependence of the Raman intensities at 588, 1054, 1230, 1498cm−1 bands on the Pyr concentration.

Table 1 Regression equations between Raman intensity and concentration (2, 3, 7, 17, 33nmolL−1) of Nap and their correlation coefficients with the same concentrations of Nap (167nmolL−1) and Phe (33nmolL−1)

Fig.4 Raman intensity of Nap (167nmolL−1), Phe (33nmolL−1) in five parallel experiments with different concentrations of Pyr.

With the different concentration of Pyr, the Raman intensity at 542, 707 (for Phe), 753, 1374cm−1(for Nap) are shown in the Fig.4. It can be seen that the spectral characteristics of Nap and Phe were affected by the changing of Pyr. The relative error (RE) of 542, 707, 753 and 1374 cm−1bands intensity are 27.95%, 27.67%, 31.32%, 27.67% respectively, which increase the difficulty of quantitative analysis. In this study, the quantitative me- thod of the SERS spectrum of three-component PAHs solution was studied, and to get the optimized accuracy, three kinds of multivariate calibration models were used.

3.2 Data and PLS-GA-BP Algorithm

In this paper, SERS signals from 352 to 1800cm−1were selected for the subsequent quantitative analysis. Sixty- four mixture samples of PAHs (Nap with the concentration of 167, 33, 17nmolL−1; Phe with 33, 17, 7, 3, 2nmolL−1; and Pyr with 33, 17, 7, 3, 2, 0.3nmolL−1) were selected. The sixty-four PAHs mixture samples were selected and divided into two sets, fifty-four for the calibration set and ten for the test set.

The robustness and prediction precision of the established models are often influenced, by some useless issues of the full spectral data. For the quantitative study of SERS spectroscopy, the selection of wave band is one of the keys for the accuracy of network prediction. In this study, PLS was used to choose the input variables.

The initial weights and thresholds of traditional neural network are randomly generated. However, network connection weights and thresholds of the whole distribution will affect the prediction accuracy. Improper initial parameters will lead to no convergence or fall into local extremum which will worsen the accuracy of the final prediction. To improve the accuracy of SERS spectra quantitative analysis of PAHs mixture, GA was adopted to optimize the initial weights and threshold values. GA algorithm can effectively reduce the randomness of the initial parameters, overcome the local optimization defects of the BP algorithm and make the prediction effect more stable. Fig.5 shows the PLS-GA-BP neural network algorithm flow.

Fig.5 Flow chart of PLS-GA-BP.

The input variables of the network were obtained through PLS dimensionality reduction, and as the fitness function of GA algorithm, the prediction error of BP algorithm was provided by MATLAB neural network toolbox using the gradient descent algorithm as the training function. Gradient descent algorithm for medium-sized BP neural network is the default training function of the toolbox, and it also has the fastest convergence speed. The initialization parameter which are the thresholds and weights, can be obtained after the GA process.

3.3 Number of Principal Components

For regression model, the input data will seriously affect the prediction accuracy of network. To improve the operation efficiency and rate, PLS method was used for data compression. To determine the number of principal component, the BP network with one hidden layer and one neuron was selected. The number of network training, learning rate, minimum error, additional momentum factor were set as 1×104, 0.01, 1×10−3, 0.95, respectively.

Table 2 Estimation of precision for PAHs samples by PLS- BP model with different number of principal components

For simplicity, only some typical results are shown in the Table 2. Three principal components were selected as optimum number for the regression of Nap according the lowest value of ARE, RMSE and the largest value of2. Similarly, the optimal number of PLS components were six and nine in the calibration set of Phe and Pyr. The PLS-BP model was recorded2=0.9955, ARE=15.04%, RMSE=24.79nmolL−1for Nap,2=0.9652, ARE=31.67%, RMSE=7.99nmolL−1for Phe,2=0.9565, ARE=19.03%, RMSE=8.08nmolL−1for Pyr.

3.4 Hidden Layer Structure

Choosing appropriate number of hidden layer and nodes is particularly significant to improve the prediction accuracy of BP network. In this study, the structure of BP network was selected according2, ARE and RMSE. With the compressed points in 3.1, different hidden layer structures of BP network were studied and the results are shown in Table 3.

Table 3 Quantification of PAHs samples by PLS-BP model with the compressed points

The result clarified that the different topologies of BP network had significance influence on the prediction accuracy. For most study, a three layer BP network with one hidden layer can be satisfied, but in order to obtain higher accuracy, the structure of network could be deepened. According to2, ARE and RMSE, the optimum hidden layer for the quantitative analysis of Nap, Phe, Pyr in the PAHs mixture samples were 2, 2, 1 respectively. Take Nap for example, network structure of 2 hidden layers reduced ARE from 15.04% to 13.51% compared with 1 hidden layer, which expounded the necessity to regulate the structure. With the optimum BP network structure for each PAH in mixture, the model was recorded2=0.9967, ARE=13.51%, RMSE=19.12nmolL−1for Nap,2=0.9495, ARE=26.99%, RMSE=9.2507nmolL−1for Phe,2=0.9565, ARE=19.03%, RMSE=8.08nmolL−1for Pyr.

3.5 Results of Quantitation

In this study, three different multivariate calibrations models including BP, PLS and GA were employed systematically and comparatively in order to obtain an optimum model for rapid quantification of multicomponent PAHs (Nap, Phe and Pyr) in water. The establishment process of the model is shown in Fig.5. The quantitative analysis results of multicomponent PAHs (Nap, Phe, Pyr) with three different multivariate calibrations models are shown in Fig.6.

Fig.6 A linear regression plot of the reference values versus SERS predicted values of the validation data set of PAHs mixture in three different multivariate calibrations models.

Table 4 Comparison of different analysis methods in the pre- diction of concentration of PAHs mixture (Nap, Phe, Pyr)

Table 4 exhibits the quantitative analysis results of three models (BP, PLS-BP, PLS-GA-BP) for Nap, Phe, Pyr in the multicomponent PAHs samples.

The optimum number of hidden layer and neurons are important for BP model performance. In this study, one hidden layer and four neurons were selected as the optimum structure according to the value of2, ARE and RMSEfor each tissue in the three components mixture. For the concentration prediction of PAHs mixture,the BP model was recorded with2=0.9505, ARE=29.32%, RMSE=55.65nmolL−1for Nap,2=0.9597%, ARE=37.84%, RMSE=12.85nmolL−1for Phe,2=0.9517, ARE=37.26%, RMSE=17.00nmolL−1for Pyr. In the PLS-BP model, the characteristic variables were selected using PLS by studying the regression modeling of dependent variables on independent variables, thus the effective spectral variables were achieved. The model was recorded with2=0.9961, ARE=16.82%, RMSE=23.31 nmolL−1,2=0.9007, ARE=28.38%, RMSE=8.40nmolL−1, and2=0.9606, ARE=19.06%, RMSE=8.57nmolL−1for Nap, Phe and Pyr, respectively. Compared with the whole spectral BP model, the ARE for PLS-BP model had been reduced 42.63%, 25.00%, 48.85%, and the RMSE had been reduced 49.00%, 34.63%, 49.59%. GA is a global probability search algorithm, which was used to optimize solutions in an uncertain space and search for optimal variables in the whole combined space meanwhile avoid falling into local minimum. Understandably the PLS-GA-BP model provided the best result with the finest predictive ability, stability and the smallest variables. The method was based on individuals, genetic iterations, generation gap crossover and mutation probabilities set to 40, 50, 0.95, 0.7, 0.01 respectively, and was recorded with2=0.9975, ARE=10.07%, RMSE=13.10nmolL−1for Nap,2=0.9710, ARE=19.28%, RMSE=5.40 for Phe,2=0.9643, ARE=16.72% and RMSE=5.10nmolL−1for Pyr. Compared with PLS-BP model, the ARE for PLS-GA-BP model had been reduced 40.13%, 32.06%, 12.27%, respectively, and the RMSE had been reduced 53.84%, 35.71%, 40.49% successively. And compared with the BP model, the ARE had been reduced 65.65%, 49.05%, 55.13% and the RMSE had been reduced 76.46%, 57.98%, 70.00%, respectively.

According to the three model results, the PLS-GA-BP model offers a simple, convenient and highly precision method for rapid quantification of multicomponent PAHs in water.

4 Conclusions

The present work established a high-precision quantitative method for SERS spectrum of PAHs mixture. The result showed that with the poorest water solubility, Pyr coupled with Au-NPs optimally, and the SERS peaks of Pyr were affected least by other substances. Conversely, the peaks of Nap were affected mostly. These phenomena laid a foundation for the quantitative analysis of SERS spectroscopy of multi-component PAHs. And to obtain a best model for rapid and accurate quantification of PAHs mixture in water, three different multivariate calibrations models including BP, PLS-BP, PLS-GA-BP were studied. In which, with three principal components and two hidden layers for Nap, six principal components and two hidden layers for Phe, nine principal components and one hidden layer for Pyr, the PLS-GA-BP performed best, and was recorded with2=0.9975, 0.9710, 0.9643, ARE= 10.07%, 19.28%, 16.72% and RMSE=13.10, 5.40, 5.10 n molL−1for Nap, Phe, Pyr, respectively. It indicated that the PLS-GA-BP showed high potential in quantitative analysis of multicomponent PAHs in water, and it might be used for fast, high precision, field quantitative analysis of PAHs in environmental water where has been severely contaminated by PAHs.

Acknowledgements

This research was supported by National Natural Science Foundation of China (No. 41476081), the Major Research and Development Project in Shandong Province (No. 2019GHY112027), the Shandong Provincial Natural Science Foundation (No. ZR2020MF121).

Akdoan, A., and Gursoy, O., 2020. Determination of four priority polycyclic aromatic hydrocarbons (4PAHs) by GC-MS in traditional Turkish yoghurts., 37 (3): 391-400.

Chen, Q. S., Hassan, M. M., Xu, J., and Zareef, M., 2019. Fast sensing of imidacloprid residue in tea using surface-enhanced Raman scattering by comparative multivariate calibration., 211: 86-93.

Chen, X. J., Liu, K., Cai, J. B., Zhu, D. H., and Chen, H. L., 2015. Identification of heavy metal-contaminatedusing infrared spectroscopy., 7 (5): 2172-2181.

Chen, Y., Chen, Z. P., Jin, J. W., and Yu, R. Q., 2015. Quantitative determination of ametryn in river water using surface- enhanced Raman spectroscopy coupled with an advanced chemometric model., 142 (15): 166-171.

Chun, L., and Zhou, J. J., 2018. An improved algorithm of BP neural network and its application., 11: 60-64.

Dasary, S. S. R., Singh, A. K., Senapati, D., Yu, H. T., and Ray, P. C., 2009. Gold nanoparticle based label-free SERS probe for ultrasensitive and selective detection of trinitrotoluene., 131 (38): 13806- 13812.

Fan, Y. X., Lai, K. Q., Rasco, B. A., and Huang, Y. Q., 2015. Determination of carbaryl pesticide in Fuji apples using surface-enhanced Raman spectroscopy coupled with multivariate analysis., 60 (1): 352-357.

Frens, G., 1973. Controlled nucleation for the regulation of the particle size in monodisperse gold suspensions., 241 (105): 20-22.

Gu, X. F., Tian, S., Zhou, Q., Adikins, J., Gu, Z. M., Li, X. W.,., 2013. SERS detection of polycyclic aromatic hydrocarbons on a bowl-shaped silver cavity substrate., 3 (48): 25989-25996.

Han, T. T., and Zhang, H. R., 2018. Determination of total polycyclic aromatic hydrocarbons in Buddha incense smoke by fluorescence spectrometry., 46 (6): 75-77.

Jia, K., Ni, Z. W., Li, J. M., and Zhu, X. H., 2019. Research on IBABC-BP parallel integration learning algorithm and its application., 39 (3): 477-494.

Joe, F. L., Jean, S., and Thomas, F., 1982. High performance liquid chromatography with fluorescence and ultraviolet detection of polynuclear aromatic hydrocarbons in Barley Malt., 65 (6): 395-402.

Jun, T., Xiang, L., and Han, K. L., 2019. Rapid analysis of main components in lavender essential oil by near infrared spectroscopy combined with bp neural network., 47 (3): 189-192, 195.

Kutsanedzie, F. Y. H., Chen, Q., Hassan, M. M., Yang, M. X., Sun, H., and Rahman, M. H., 2017. Near infrared system coupled chemometric algorithms for enumeration of total fungi count in cocoa beans neat solution., 240 (1): 231- 238.

Langer, J., Aberasturi, D. J. D., Aizpurua, J., 2019. Present and future of surface enhanced raman scattering., 14 (1): 86-87.

Li, B. P., Qin, H. J., Luo, X., Cao, S. W., and Wang, J. D., 2007. Determination of four contents of feedstuff powder using near infrared spectroscopy by PLS-BP model., 27 (10): 2005-2009.

Li, H. H., Liu, S. S., Hassan, M. M., Ali, S., Ouyang, Q., Chen, Q. S.,., 2019. Rapid quantitative analysis of Hg2+residue in dairy products using SERS coupled with ACO-BP- AdaBoost algorithm., 223: 117281-117290.

Li, H. S., Lai, L., Chen, L., Lu, C., and Cai, Q., 2015. The prediction in computer color matching of dentistry based on GA+BP neural network., 2015 (2): 1-7.

Li, X. G., Deng, W., Zhou, X., Tang, X. L., Guo, X. Y., and Wang, Y., 2012. Distribution of PAHs in surface seawater of Qingdao coast area and their preliminary apportionment., 33 (3): 741-745.

Lin, S., Lin, X., Liu, Y. L., Zhao, H. L., Hasi, W., and Li, W., 2018. Self-assembly of Au@Ag core-shell nanocubes embedded with internal standard for reliable quantitative SERS measurements., 10: 4201-4208.

Liu, J., and Li, H. L., 2017. PID control algorithm of BP neural network of dynamic quantitative weighing packaging system., 38 (5): 78-81.

Ma, F. M., and Li, G. J., 2005. Application of BP neural net work in quantitative recognizing pipeline defect., 7 (4): 16-18.

Menzie, C. A., Potocki, B. B., and Santodonato, J., 1992. Exposure to carcinogenic PAHs in the environment., 26 (7): 1278-1284.

Orecchio, S., Ciotti, V. P., and Culotta, L., 2009. Polycyclic aromatic hydrocarbons (PAHs) in coffee brew samples: Analytical method by GC–MS, profile, levels and sources., 47 (4): 819-826.

Qian, X. M., and Nie, S. M., 2008. Single-molecule and single- nanoparticle SERS: From fundamental mechanisms to biomedical applications., 37 (5): 912-920.

Shi, X. F., Liu, S., Han, X. H., Ma, J., Jiang, Y. C., and Yu, G. F., 2015. High-sensitivity surface-enhanced raman scattering (SERS) substrate based on a gold colloid solution with a PH change for detection of trace-level polycyclic aromatic hydrocarbons in aqueous solution., 69 (5): 574-579.

Shi, X. F., Ma, J., Zheng, R. E., Wang, C. Y., and Kronfeldt, H. D., 2012. An improved self-assembly gold colloid film as surface-enhanced Raman substrate for detection of trace-level polycyclic aromatic hydrocarbons in aqueous solution., 43: 1354-1359.

Simcik, M. F., Eisenreich, S. J., and Lioy, P. J., 1999. Source apportionment and source/sink relationships of PAHs in the coastal atmosphere of Chicago and Lake Michigan., 33 (30): 5071-5079.

Tripp, R. A., Dluhy, R. A., and Zhao, Y., 2008. Novel nanostructures for SERS biosensing., 3 (3-4): 31-37.

Wang, Y. T., Zhang, Y., Shang, F. K., Zhang, J. Z., Zhang, H., Sun, Y. Y.,., 2019. Measurement of polycyclic aromatic hydrocarbons in water by bp neural network combined with ATLD and three-dimensional fluorescence spectrometry., 39 (11): 3420-3425.

Wenclawiak, B., Rathmann, C., and Teuber, A., 1992. Supercritical-fluid extraction of soil samples and determination of polycyclic aromatic hydrocarbons (PAHs) by HPLC., 344 (10-11): 497-500.

Xie, Y. T., 2007. The determination of the number of the hidden cells and the variable selection of the neural network., 22 (6): 9-15.

Yin, X. P., Dong, H., Wang, S. Q., Liang, Y., Gao, N., Zhang, W. L.,., 2019. A multiple coupling approach to produce high-performance SERS substrates., 30 (01): 199-202.

Yu, F., and Xu, X. Z., 2014. A short-term load forecasting model of natural gas based on optimized genetic algorithm and improved BP neural network., 134 (1): 102-113.

Zhang, W., Zhou, N., Li, N., Xie, Y. S., Luo, H. D., and Li, Y. Q., 2009. Simultaneous determination of pahs in food by low- temperature constant-energy synchronous fluorescence spectrometry., 29 (10): 2806- 2809.

Zhang, Y., and Tao, S., 2009. Global atmospheric emission inventory of polycyclic aromatic hydrocarbons (PAHs) for 2004., 43 (4): 812-819.

Zhao, Y., Tian, Y., Ma, P. Y., Yu, A. M., Zhang, H. Q., and Chen, Y. H., 2015. Determination of melamine and malachite green by surface-enhanced raman scattering spectroscopy using starch-coated silver nanoparticles as substrate., 7: 8116-8122.

November 12, 2020;

April 14, 2021;

June 3, 2021

© Ocean University of China, Science Press and Springer-Verlag GmbH Germany 2021

. Tel: 0086-532-66781201

E-mail: shixiaofeng@ouc.edu.cn

(Edited by Ji Dechun)