Vortex assisted dispersive solid phase extraction of thallium followed by electrothermal atomic absorption spectrometry, Adsorption mechanism and soft computing algorithm prediction

ABSTRACT Vortex-assisted dispersive solid-phase extraction technique was used as a rapid and efficient method for preconcentration of ultra-trace levels of total thallium (Tl) followed by electrothermal atomic absorption spectrometry. Graphene oxide-modified polyvinylpyrrolidone nanocomposite was prepared as an adsorbent and characterised by the Fourier transform infrared spectrometry, field emission-scanning electron microscopy, energy-dispersive X-ray spectroscopy , and X-ray diffraction spectroscopy. To find the optimum conditions for extraction of ultra-trace levels of Tl (ӀӀӀ), the response surface methodology based on central composite design was used. Based on the results, pH = 6.6, amounts of adsorbent = 7.7 mg, extraction time = 27 min and desorption time = 5 min provide maximum extraction recovery for Tl (ӀӀӀ). Under the optimum conditions, the calibration curve was linear in the range of 0.08–1.5 µg L−1 Tl (ӀӀӀ) with the R2 value of 0.9985. The relative standard deviation was 4.0% (n = 7) and the detection limit was 0.019 µg L−1 (n = 8). Also, the enrichment factor which could be calculated from the slope of the calibration curve after preconcentration step to that without preconcentration was 95.8. To understand the adsorption mechanism, two-parameter and three-parameter adsorption isotherms were studied, and the obtained results show that the adsorption of Tl (ӀӀӀ) followed by the Freundlich isotherm and the maximum adsorption capacity was 142.8 mg g−1. Also, the results of adsorption kinetic show that the adsorption of Tl (ӀӀӀ) was followed by the pseudo-first-order kinetic model. Moreover, random tree (RT) and artificial neural network (ANN) are employed for prediction of adsorption performance based on effective parameters. The outcomes of soft computing demonstrated that RT (R2 = 0.95) and ANN (R2 = 0.87) have acceptable accuracy for estimation and prediction of extraction recovery of Tl (ӀӀӀ).


Introduction
Today application of metals for different purposes increases dramatically [1][2][3].Thallium (Tl) is one of the hazardous elements in the environment and exists in two oxidation states, Tl (Ӏ) and Tl (ӀӀӀ), which exhibit different toxicities and bioactivities.Tl induces cytotoxic, apoptosis, DNA damage and increases structural chromosomal aberrations.Although its low concentration exists in water samples, however, it has high toxicity even at trace levels.Tl is introduced into the environment mainly as waste from the production of zinc, cadmiumand lead and by combustion of coal.Anthropogenic sources of thallium include gaseous emissions from cement industries, copper smelting, petroleum refining, coal-based power plants, metal sewers and ore processing operations [4].According to the United States Environmental Protection Agency (USEPA), the maximum level of Tl in drinking water is 2 µg L −1 ; therefore, in order to obtain accurate results for determination of trace levels of Tl, sample preparation techniques should be used before direct analysis.Up to now, different analytical techniques including liquid phase microextraction (LPME) [5,6] and solid-phase extraction (SPE) [7][8][9][10][11][12][13] were used for preconcentration of Tl.Today, SPE is considered as a powerful sample preparation technique due to advantages such as synthesis of selective adsorbents [14][15][16][17][18][19], high enrichment factor, reusability of adsorbent and green approach [20].Graphene oxide (GO) is usually prepared from oxidation of graphite and has unique properties such as high surface to area, ease of synthesis and ability to surface modification due to the existence of epoxy, hydroxyl and carboxylic acid groups in its structure [21][22][23][24].In 2018, Eftekhari et al. used tannic acid for modification of GO.The synthesised tannic acidcoated graphene oxide (TA-GO) nanocomposite was then applied for preconcentration of trace levels of lead (Pb 2+ ) in different real samples.The obtained results show that TA-GO nanocomposite has the maximum theoretical adsorption capacity of 250 mg g −1 for Pb 2+ [25].In 2017, Molaei et al. synthesised SiO 2 -coated magnetic graphene oxide modified with polypyrrole-polythiophene (mGO/SiO 2 @coPPy-Th) for preconcentration of heavy metals from water and agricultural samples.The obtained results show that the maximum sorption capacities for Cu(II), Pb(II), Zn(II), Cr(III) and Cd(II) were 201, 230, 125, 98 and 80 mg g −1 , respectively [26].Polyvinylpyrrolidone (PVP) or povidone is an amorphous and synthetic polymer consisting of linear 1-vinyl-2-pyrrolidinone groups.It is a water-soluble polymer which has -N and -O groups in its structure so that it is expected to interact by heavy metals effectively [27,28].Herein, GO was synthesised by the Hummer method and then modified by PVP to synthesis of GO-PVP nanocomposite.By using SPE method, the synthesised GO-PVP nanocomposite was used for preconcentration of Tl (ӀӀӀ) followed by its determination using electrothermal atomic absorption spectrometry (ETAAS).Based on our knowledge, there is no research conducted for applying of GO-PVP nanocomposite for SPE of Tl (ӀӀӀ).The response surface methodology based on central composite design (RSM-CCD) technique was used to find the optimum conditions.Also, to evaluate the adsorption mechanism, two-and threeparameter adsorption isotherms were investigated and interpreted.Then, the adsorption kinetic models were studied by interpreting the kinetic results, and finally the proposed method was used for determination of ultra-trace levels of total Tl in different water and wastewater samples.

Instrumentation
Perkin Elmer Analyst 700 model 4100, ETAAS equipped with deuterium lamp as a background correction system was used for determination of Tl.Tl hollow cathode lamp at a wavelength of 276.8 nm was used as a radiation source.A pyrolytic-coated graphite tube with Lvov platform was used.Table 1 shows the instrumental parameters of ETAAS for determination of Tl.The pH of sample solution was measured with Metrohm 827 (Switzerland).Phase separation was assisted using a Centurion Scientific instrument Centrifuge (model Andreas Hettich D72, Tuttlingen, Germany).Field emission-scanning electron microscopy (FE-SEM) and energy-dispersive X-ray spectroscopy (EDX) analysis were performed by the TESCAN BRNO Mira3, LMU (Czech Republic) instrument.The Fourier transform infrared (FT-IR) spectrophotometer (Thermo Nicolet model AVATAR 370, USA) was used to determine the functional groups of GO and GO-PVP nanocomposite.Also, X-ray diffraction (XRD) analysis was performed by the X Pert PW 3040/60 Philips instrument (Netherlands).

Synthesis of GO and GO-PVP nanocomposite
GO was synthesised from graphite powder by the Hummer method [25].Briefly, to the 1 g of graphite powder, 23 mL of H 2 SO 4 was added and the mixture was stirred at 5°C for 30 min.Then, 0.5 g of NaNO 3 was added to the mixture and stirred for 30 min at 15-20°C.By addition of 3 g of KMnO 4 during 1 h to the mixture, it was stirred for 90 min at 15-20°C and heated up to 35°C followed by stirring for another 120 min at this temperature.Also, 100 mL of deionised water was added slowly followed by addition of 5 mL H 2 O 2 to the mixture to remove excess amounts of permanganate.After that, the synthesised graphite oxide was washed with deionised water until it reachs pH 4-5 and ultrasonicated at 500 W for 45 min to produce GO.The prepared GO was centrifuged at 3500 rpm and dried at vacuum oven at 60°C overnight.
For the synthesis of GO-PVP nanocomposite, 0.5 g of the synthesised GO was ultrasonicated in 100 mL of deionised water for 45 min.Then, 0.2 g of PVP polymer was dissolved in 20 mL of deionised water and added to the GO mixture.It was ultrasonicated for 20 min followed by stirring for 3 h at room temperature.Finally, the synthesised GO-PVP nanocomposite was centrifuged at 4000 rpm for 10 min, washed with deionised water for three times and dried at 60°C overnight.

Solid-phase extraction method
The SPE method is performed as follows: 10 mL of sample solution containing 1 µg L −1 Tl (ӀӀӀ) was adjusted at pH 6.6.Then, 7.7 mg of GO-PVP nanocomposite was added to the sample solution and vortex for 30 min at 2800 rpm.The obtained mixture was centrifuged at 4000 rpm for 5 min followed by discarding of supernatant using 2-mL syringe.By addition of 100 µL of 1 mol L −1 HNO 3 as desorbent solution and vortex for 5 min, it was centrifuged again for 5 min at 5000 rpm.Finally, an aliquot of 20 µL of supernatant was injected into the ETAAS for quantification of Tl.

Analysis of real samples
Different water and wastewater samples were collected, and after oxidation of Tl (Ӏ) to Tl (ӀӀӀ) by concentrated H 2 O 2 (30%) and heating for 10 min at 60-70°C to remove excess amounts of H 2 O 2 , the level of total thallium was determined according to the SPE procedure.

Characterisation of GO-PVP nanocomposite
There are two mechanisms for interactions of PVP polymer with GO nanosheets: (1) SN 2 reaction between N-and O-groups of PVP polymer as nucleophile with epoxy ring of GO, and (2) hydrogen bonding interaction between OH-groups of GO nanosheet with N-and O-atoms of PVP polymer.To characterise the synthesised GO-PVP nanocomposite, FT-IR spectrophotometry, FE-SEM, EDX and XRD analysis were performed.The results of FT-IR spectrophotometry are presented in Figure 1.As it shows, the peaks at 3400 cm −1 , 1730 cm −1 , 1650 cm −1 , 1220 cm −1 and 1050 cm −1 are related to OH-, C=O, C=C, epoxy and C-O groups, respectively, and clearly show that GO synthesised successfully [25,29].By modification of GO with PVP polymer, the reduce of epoxy peak at 1220 cm −1 was observed which may be related to the opening of epoxy ring due to its interaction with Nand O-atoms of PVP by SN 2 reaction [30].Also, the peaks around 1280 cm −1 and 1423 cm −1 in GO-PVP nanocomposite are related to the C-N bending vibration of the pyrrolidone structure and CH deformation modes of CH 2 group, respectively [31].
Also, the FE-SEM images of GO and GO-PVP nanocomposite are presented in Figure 2  (a, b).As it could be seen in Figure 2(a), GO nanosheet has a uniform structure; however, its modification by PVP causes to produce of non-uniform structure (Figure 2(b)) which clearly shows that GO-PVP nanocomposite was synthesised successfully.Figure 3 shows the EDX spectrum of GO-PVP nanocomposite, and according to the results, the N-peak in the EDX spectrum is related to the PVP polymer coated on GO.
The XRD spectrums of GO and GO-PVP nanocomposite are presented in Figure 4.As it could be seen, the peak at 2θ = 10.9° is related to the 0 0 2 plane of GO.However, after its modification with PVP polymer, the peaks around 2θ = 15-30° are related to the semicrystallinity of PVP polymer [32].Finally the BET analysis was employed to determine the surface area of the GO-PVP nanocomposite.The surface area of the adsorbent plays a critical role in adsorption phenomena so that by increasing the surface area the available active sites for interaction by analyte increase effectively.The BET spectrum of GO-PVP nanocomposite is presented in Figure 5.According to IUPAC classification, it exhibits type IV isotherm with an obvious hysteresis loop (at P/P0 = 0.50-0.95)which corresponds to the presence of mesoporesmicropore structure.Also, the hysteresis exists due to the presence of slit-shaped pores (H3).According to the results, the obtained surface area and total pore volume of GO-PVP nanocomposite were 75.12 m 2 g −1 and 0.352 cm 3 g −1 , respectively.

RSM-CCD optimisation technique
The RSM-CCD method was used to find optimum conditions for SPE of Tl (ӀӀӀ).The level of each variable is presented in Table 2.The optimisation responses of mathematical computing in the CCD are illustrated in Table 3.In this experiment, cubic, linear, 2FI and quadratic are evaluated as models and according to the predicted R 2 and R 2 coefficients; quadratic model was recommended to the laboratory data.For design of experiment with four parameters, the model for Tl (ӀӀӀ) is shown in Equations (1).Analysis of variance (ANOVA) is done to calculate the coefficients of the quadratic model.Also, the outcomes of ANOVA analysis for quadratic equation are presented in Table S1.
To appraise how the mathematical equation fulfils the hypothesis of ANOVA, a normal probability curve of the residuals calculated by the Design Expert software and the outcomes are depicted in Figure 6.As it appears, the focuses on this plot lie near the straight line which affirms that it complies with normal distribution.The studied parameters and their amounts were processed using nonlinear regression equation to compute and plot the 3D CCD surface curves (Figure 7(a-f)).Figure 7(a) shows the effects of pH of sample solution and the amounts of adsorbent (M) on the ER of Tl (ӀӀӀ).As it was predicted by ANOVA test, the pH of sample solution has high significant effect on the ER of Tl (ӀӀӀ).According to the results, by increasing the pH from 2 to 5-6.5, the ER of Tl (ӀӀӀ) increases which may be related to the deprotonation of N-and O-groups of PVP polymer at lower acidic pH values; therefore, efficient adsorption of Tl (ӀӀӀ) occurred at the pH around 6-6.5.However at high pH values (pH ˃ 7), the ER decreases gradually which may be related to the formation of Tl(OH) 3 in the sample solution.Also, the amounts of adsorbent were considered as a significant parameter according to the ANOVA test, and based on the results, the ER increases by increasing the amounts of adsorbent which could be related to the increasing of the active sites for interaction by Tl (ӀӀӀ).Figure 7b, c) shows the effects of pH, extraction time and desorption time on the ER of Tl (ӀӀӀ).As predicted by ANOVA test, the extraction and desorption times are considered as insignificant parameters which show that the adsorption of analyte onto the adsorbent or desorption of Tl (ӀӀӀ) (0.5 mol L −1 HNO 3 as desorbent solution) takes place immediately.Figure 7(d-f) also shows the effects of amounts of adsorbents, extraction and desorption times on the ER of Tl (ӀӀӀ) and completely discussed above.Finally, the optimum conditions to obtain maximum ER for Tl (ӀӀӀ) are pH 6.6, amounts of adsorbent 7.7 mg, extraction time 27 min and desorption time 5 min.

Effect of ionic strength
The effect of ionic strength on the ER of Tl (ӀӀӀ) was studied by analysing the sample solution containing different concentrations of KNO 3 in the range of 0-500 mg L −1 .The obtained results show that no serious change was observed in the ER of Tl (ӀӀӀ) up to 250 mg L −1 KNO 3 .

Effect of interfering ions
The effect of different ions on the ER of Tl (ӀӀӀ) was studied, and the results are presented in Table 4.An ion was considered as interfere if it changes the ER of analyte more than ± 5% [33,34].According to the obtained results, no serious interfere was observed for determination of Tl (ӀӀӀ).

Analytical figures of merit
The linear range for Tl (ӀӀӀ) was obtained in the range of 0.08-1.5 µg L −1 with a correlation coefficient of 0.9985.The relative standard deviation (RSD) for seven replicate analysis of 1.0 µg L −1 Tl (ӀӀӀ) was 4.0%, and the limit of detection was 0.019 µg L −1 (n = 8).The preconcentration factor was determined by division of the volume of sample solution (10 mL) to the volume of desorbent solution (0.1 mL) which equals to 100, and the calculated enrichment factor was 95.8 [35].

Analysis of real sample
The accuracy of the proposed method was checked by the analysis of CRM-TMDW containing 10 µg L −1 Tl, and the obtained result, 9.88 ± 0.76 µg L −1 , was in satisfactory agreement with a certified value (n = 6, 95% confidence limit).Also, the results for determination of trace levels of Tl (ӀӀӀ) in different water and wastewater samples are presented in Table 5.Based on the results, the level of Tl in tap and river water samples was lower than the recommended reference value according to EPA for drinking water samples (2 µg L −1 ) and the obtained recoveries were in the range of 91.0-97.0%[36].

Equilibrium study
By application of two-parameter [Langmuir, Freundlich, Temkin and Dubinin-Radushkevich (D-R)] and three-parameter (Toth, Khan, and Sips) isotherm algorithms [37], the adsorption mechanism of Tl (ӀӀӀ) onto the GO-PVP nanocomposite was evaluated.Likewise, the conceptual plan of equilibrium appraisal and outcomes is depicted in Figure S1.According to the results, the correlation coefficients of the Langmuir and Freundlich isotherms are equal to 0.97; however, for determination of exact adsorption mechanism, three-parameter algorithms are evaluated.According to three-parameter isotherm computations, all exponents in Toth (1/t), Khan (a K ) and Sips (β S ) isotherms are far from 1; therefore, heterogeneous surface adsorption (Freundlich equation) is the best model for adsorption of Tl (ӀӀӀ) onto the GO-PVP nanocomposite [38].Moreover, based on the D-R and Temkin isotherm calculations, E (D-R model) and b (Temkin model) parameters are less than 8 kJ mol −1 and show that the adsorption of Tl (ӀӀӀ) is a physisorption phenomenon.Also, according to the Langmuir isotherm, the maximum adsorption capacity of GO-PVP nanocomposite for Tl (ӀӀӀ) is 142.8 mg g −1 .

Kinetic study
For evaluation of the adsorption kinetic of Tl (ӀӀӀ) onto the GO-PVP nanocomposite, the pseudo-first-order, pseudo-second-order, intra-particle diffusion, Elovich and Boyd models are evaluated [38,39].The conceptual model of kinetic scrutinising and the obtained results are illustrated in Figure S2.According to the results, the delta value (q e(exp) -q e(cal) ) of the pseudo-second-order kinetic model is more than the pseudo-first-order kinetic model, and there are much differences between q e(exp) and q e(cal) in the pseudo-second-order kinetic model.Also, the correlation coefficient of the pseudo-first-order kinetic model (R 2 = 0.98) is higher than the pseudo-second-order kinetic model (R 2 = 0.75); Therefore, it could be concluded that the adsorption kinetic of Tl (ӀӀӀ) onto the GO-PVP nanocomposite is followed by the pseudo-first-order kinetic model.Moreover, by application of Elovich kinetic model, heterogeneity of surface is approved (R 2 = 0.99).Also, based on Figures 8 and 9, the intraparticle and Boyd lines do not pass from the origin of plots, thus, the adsorption of Tl (ӀӀӀ) onto the GO-PVP nanocomposite is implemented by the film diffusion process [29].Geometry method (Figure 10) was also used as a system for analysing interfering of adsorption and desorption rates [29,38].Based on the geometric results in Figure 11, it can be concluded that the adsorption and desorption processes do not have any interfering during the reaction time.

Soft computing algorithms
In the present research, random tree (RT) and artificial neural network (ANN) are applied for evaluation of adsorption behaviour during the reaction.For implementation of RT and ANN soft computing algorithms, WEKA 3.9.2 and MATLAB 2013b tools are utilised, correspondingly.A schematic structure of RT outcomes computing in WEKA software is illustrated in Figure 12 and Table 6.Moreover, the logical equations of RT algorithm for recovery percentages prediction based on the pH, amounts of adsorbent, extraction and  desorption times are depicted in Equation ( 2).According to Table 6, it can be concluded that the RT algorithm with correlation coefficient and root mean squared error equal to 0.95 and 8.54, respectively, are acceptable and useful for forecasting recovery percentage of Tl (ӀӀӀ).In the following, specifications of ANN model for prediction of recovery percentage are summarised in Figure 13 and Table 7.According to Figure 13, four parameters (pH, amounts of adsorbent, extraction and desorption times) are entered as an input and recovery percentage is considered as an output of mathematical model.Training, validation, test and overall results of ANN calculations are depicted in Figure 14.According to the results, ANN computing can estimate the recovery percentages based on the effective   parameters by correlation coefficient value of 0.87 which is acceptable.Likewise, variations of mean square error (MSE) for train, validation and test versus epochs are demonstrated in Figure 15 which tend to be constant amount.Finally, by comparison of correlation coefficients in RSM-CCD, RT and ANN algorithm, which are equal to 0.97, 0.95 and 0.87, respectively, it could be concluded that the RSM-CCD and RT techniques have the best accuracy for estimation of RP of Tl (ӀӀӀ) based on the effective factors in adsorption process.  2 Dispersive micro solid-phase extraction. 3Magnetic solid-phase extraction. 4Ultrasonic assisted-dispersive solid-phase microextraction. 5Inductively coupled plasma optical emission spectrometry. 6Ultrasound-assisted cloud point extraction (UA-CPE) and dispersive μ-solid phase extraction (D-μ-SPE). 7Mixed micelle cloud point extraction.

Comparison of the proposed method with other techniques
The comparison of the proposed SPE method with other analytical procedures such as dispersive micro solid-phase extraction [7], mixed micelle cloud point extraction [40] and ultrasound-assisted dispersive liquid-liquid microextraction [41] was performed for determination of Tl, and the results are presented in Table 8.As can be seen, the proposed method has low LOD (0.019 µg L −1 ), acceptable RSD (4%), high enrichment factor (95.8) and novelty (using GO-PVP nanocomposite) for determination of Tl (ӀӀӀ) in different water samples.However, since Tl species have different toxicities, the main drawback of the proposed method is its inability to speciation of Tl species.

Conclusion
Vortex-assisted dispersive solid-phase extraction (VADSPE) technique was used for preconcentration and determination of ultra-trace levels of total Tl followed by ETAAS.GO-PVP nanocomposite was used as an adsorbent and characterised by the FT-IR spectrophotometry, FE-SEM images, EDX and XRD analysis.In order to optimise the critical parameter, RSM-CCD methodology was used.To study the adsorption isotherm, two-and three-parameter adsorption isotherms were studied, and based on the results, the adsorption of Tl(ӀӀӀ) is a physical process and followed by the Freundlich adsorption isotherm.Also, the maximum adsorption capacity for Tl (ӀӀӀ) was 142.8 mg g −1 .The results of adsorption kinetic show that, the adsorption of Tl (ӀӀӀ) was followed by the pseudo-first-order kinetic model.Finally, RT and ANN were employed for prediction of adsorption performance based on the effective parameters.The outcomes of soft computing demonstrated that RT (R 2 = 0.95) and ANN (R 2 = 0.87) have acceptable accuracy for estimation and prediction of extraction recovery of Tl (ӀӀӀ).

Figure 1 .
Figure 1.FT-IR spectrum of GO and GO-PVP nanocomposite.

Figure 4 .
Figure 4. XRD spectrum of GO and GO-PVP nanocomposite.

Figure 7 .
Figure 7.The response surface model of the proposed method versus affecting parameters for Tl (ӀӀӀ).

Figure 12 .
Figure 12.Outcomes of RT algorithm for estimation of extraction recovery.

Figure 13 .
Figure 13.Structure of ANN algorithm for estimation of recovery percentage.

Figure 14 .
Figure 14.Outcomes of ANN algorithm for estimation of recovery percentage.

Table 1 .
ETAAS instrumental parameters for determination of Tl.

Table 2 .
The margins of parameters utilised for CCD of Tl (ӀӀӀ).

Table 4 .
Effect of different cations and anions on the extraction recovery of 1.0 ng mL −1 Tl (ӀӀӀ).

Table 5 .
The results of analysis of real samples for determination of Tl (ӀӀӀ) (results: mean ± standard deviation based on three replicate analysis).

Table 6 .
Statistical outputs of RT computation for extraction recovery prediction.

Table 7 .
ANN conditions for recovery percentage forecasting in adsorption process.

Table 8 .
Comparison of the proposed method with other SPE techniques for determination of Tl.