Introduction

Elements mass fraction in food matrices have been commonly determined by different spectroanalytical techniques based on plasma, such as inductively coupled plasma optical emission spectrometry (ICP OES) (Villa et al. 2015; Mir-Marqués et al. 2015; Mir-Marqués et al. 2016; Costa et al. 2019a) and inductively coupled plasma-mass spectrometry (ICP-MS) (Chevallier et al. 2015; Mir-Marqués et al. 2016) due to its high sensitivity and multi-elemental capability.

However, these analytical instrumental techniques usually require a sample pre-treatment procedure employing acid mineralization with the help of oxidizing reagents (hydrogen peroxide, for instance). These procedures are most used in the sample preparation, presenting some drawbacks, such as risk of contamination of the samples, low analytical frequency, and use of concentrated acid, which can affect the analytical blank and the accuracy of the results (Korn et al. 2008; Ferreira et al. 2010). All these observations make less attractive the use of acid mineralization for the determination of macro- and micronutrients in food matrices.

Energy dispersive X-ray spectroscopy (EDXRF) is commonly used for metals determination in different matrices, since liquid to solid samples (Pereira et al. 2006; Gupta et al. 2013; Peruchi et al. 2014; Bull et al. 2017; Brito et al. 2017; Andrey et al. 2018; Byers et al. 2019). Aside from direct solid sample analysis, EDXRF is non-destructive, selective and allows simultaneous determination; also has higher frequency, it is simpler in design, and cost less (Kelsey et al. 2016; Oyedotun 2018; Oliveira et al. 2020), which makes it suitable to overcome the drawbacks described previously. EDXRF presents a simple spectrum with few spectral interference possibilities. Some difficulties can be observed to determine light elements, because mostly are transparent to X-rays and the X-ray absorption effect is relatively lower than those presented by heavy elements (Van Grieken and Markowicz 2001).

Another technique commonly used for direct solid analysis is laser-induced breakdown spectroscopy (LIBS). This is a versatile technique that presents several advantages, such as multi-elemental capability and direct solid analysis, and high analytical frequency where hundreds of spectra are collected in a short period of time (Markiewicz-Keszycka et al. 2017; Costa et al. 2019b; Machado et al. 2020).

Both techniques (EDXRF and LIBS) have been traditionally used in combination with routine univariate calibration strategies (Costa et al. 2019c). However, depending on complexity of the sample, this option is not always effective, being necessary other approaches to minimize matrix effect, and some examples are well described in the scientific literature.

Therefore, the goal of this study was to evaluate multivariate calibration models employing direct solid sample analysis by EDXRF for the determination of Cu, K, Sr, and Zn in cocoa bean samples. Considering the potential information provided by EDXRF and LIBS, regarding the composition of cocoa beans, data fusion strategy was also evaluated to improve both analytical performance and predictive capability of the analytes.

Materials and Methods

Instrumentation

The measurements were carried out using a benchtop EDXRF spectrometer, Rigaku NEX QC+ (Austin, TX, USA), equipped with an Ag target X-ray tube and Be detector window that can be operated at a maximum voltage of 50 kV with a resolution of 0.024 keV (2048 channels) and a 25% detector dead time. During the spectrum acquisition was considered Kα transitions for all elements under investigation.

A commercial LIBS, J 200 model from Applied Spectra (Freemont, USA), was used as complementary technique. The instrument is equipped with a Nd-YAG laser operating at 1064 nm. The instrument contains a high efficiency particulate air cleaner (HEPA) to purge the ablated particles, and the movement of the sample is automated with a XYZ stage. It also contains a 1280 × 1024 complementary metal-oxide semiconductor (CMOS) color-camera imaging system. An optical fiber bundle is coupled to a 6-channel charged-coupled device (CCD) spectrometer to convert the plasma emission light into spectra, with a spectral range of emission signals between 186 and 1042 nm resulting in 12,288 variables. The gate width is the time that spectrometer collects the emission signals, which is fixed at 1.05 ms. For the identification of elements emission lines, Aurora Software Package (also from Applied Spectra) was used. The additional operation conditions used were 2699 J cm−2 laser pulse fluence and 1.9 μs delay time, and the laser pulse energy was about 50 mJ (Gamela et al. 2020a).

Inductively coupled plasma optical emission spectrometer (ICP OES) model iCAP 7400 (Thermo Fisher, Madison, WI, USA) was used as reference method after microwave digestion. The measurements were performed using argon gas (99.999%, White Martins-Praxair, Sertãozinho, SP, Brazil), using both axial and radial viewing modes. The instrumental conditions and the emission lines used are shown in Table 1.

Table 1 Instrumental parameters for ICP OES determinations
Full size table

The digestion procedure was performed using a single reaction chamber oven (UltraWaveTM, Milestone, Sorisole, Italy). Before digestion process, a volume of 150 mL of water and 5.0 mL of concentrated HNO3 were added into the single chamber reaction, which was pressurized with nitrogen gas (99.9%, White Martins-Praxair) until 40 bar.

Reagents and Solutions

Analytical-grade reagents and deionized water (Milli-Q system, 18.2 MΩ cm, Millipore, Bedford, MA) were used to prepare all solutions and standards. All glassware and vessels were soaked in a 10% v v−1 HNO3 solution during 24 h. After that, a rinsing step with deionized water was performed, and the materials were left to dry in a clean hood before use. The multielementar solutions used for external calibration were prepared from dilution of 1000 mg L−1 of Cu, K, Sr, and Zn standard solutions (Qhemis, São Paulo, SP, Brazil) using dilute HNO3, i.e., 1% v v−1 (Synth, Diadema, SP, Brazil), previously purified using a sub-boiling distillation system Distillacid™ BSB-939-IR (Berghof, Eningen, Germany).

Samples and Sample Preparation

Ten samples of cocoa bean were obtained in an area of the Cocoa Research Center—CEPEC, an entity linked to the Executive Committee of the Cacao Plan—CEPLAC of the Brazilian Ministry of Agriculture, Livestock and Supply (MAPA) located in Bahia State, Brazil.

Before analysis by EDXRF, the samples were only grinded to reduce the particle size to improve the homogeneity of the analytes in the samples. Then, approximately 500 mg of each sample was placed into polypropylene cells and afterwards covered with Mylar® film of a 6-μm thickness (Premier Lab Supply, Port St. Lucie, FL, USA). The samples analyzed using EDXRF spectrometer were the same for LIBS. However, previously to LIBS analysis, the samples were pressed using 60 t in−1 to form pellets (~ 12 mm diameter) in order to make easier the data collection.

To evaluate the efficiency of the digestion procedure of the sample, four certified reference materials (CRMs) of baking chocolate (NIST 2384), whole milk powder (NIST 8435), apple leaves (NIST 1515), and spinach leaves (NIST 1570a), all from National Institute of Standard and Technology (Gaithersburg, MD, USA), were also digested using concentrated HNO3.

Reference Method

The method used as reference was developed by Gamela et al. (2020b) and presents the following characteristics: a mass of 200 mg of cocoa beans was directly weighed in perfluoroalkoxy alkanes (PFA) vessels and, 5 mL of 3.75 mol L−1 HNO3 and 1.75 mL of 30% v v−1 H2O2 were added. The samples were submitted to the following heating program (temperature in °C/ramp in min/hold in min): (i) 180/5/5, (ii) 210/5/10, and (iii) 230/5/10. After that, the vessel contents were transferred to volumetric flasks, and the solution was diluted to 25 mL with ultrapure water before analysis.

Acquisition of LIBS Spectra and Chemometric Tools

For LIBS, each sample was analyzed in triplicate using the conditions described in “Instrumentation.” Approximately 600 spectra were collected, and all data treatment was performed using MATLAB® 2019a (The MathWorks Inc., Natick, MA, USA) laboratory-made routines, and an interface integrated MATLAB for first-order multivariate calibration (MVC1) (Olivieri et al. 2004). The collected spectra were normalized to minimize the microheterogeneity and signal fluctuation during the data acquisition (Castro and Pereira-Filho 2016). However, the normalization process was not necessary for EDXRF, since this technique has not shown this issue previously mentioned.

The samples were divided using Kennard and Stone algorithm into calibration and validation sets (Kennard and Stone 1969). A number of samples used in the calibration and validation sets were 7 and 3, respectively. After that, partial least squares regression (PLS) was used to model all data set, and appropriate number of latent variables was selected and predicted the concentration values in the sample test. This tool is considered first-order multivariate calibration, which minimizes the lack of selectivity. The mathematical algorithms are used to obtain information to predict the concentration in the samples, and the main advantage is the possibility to propose models in the presence of concomitants (Escandar et al. 2007; Olivieri 2015; Zhang et al. 2018).

The best results were considered those that present lowest standard error of cross-validation (SECV). In order to improve the capability of the prediction of the model, data fusion strategy was evaluated. Data fusion consists in the use of multiple sources of data to produce a single model (De Oliveira et al. 2019; Gamela et al. 2020c), and due to the versatile characteristic of LIBS, this technique was used as complementary.

Results and Discussions

Reference Method by ICP OES

For the evaluation of the proposed procedure, four different CRMs, such as apple and spinach leaves, baking chocolate, and whole milk powder, were submitted to the optimized conditions described in previous study by Gamela et al. (2020b), and the results are shown in Table 2. The results obtained using this procedure demonstrated that the concentrations of all analytes in CRM are in concordance with certified values according to Student’s t test at 95% of confidence level. Moreover, the results obtained employing the total digestion of the samples with concentrated HNO3 and with proposed procedure were also compared using a paired Student’s t test at 95% of confidence level (Gamela et al. 2020b). The results showed that there is no significant difference between the reference method and proposed procedure.

Table 2 Concentration values (mg kg−1) of Cu, K, Sr, and Zn obtained in certified reference materials using diluted HNO3 (mean ± standard deviation, n = 3)
Full size table

Multivariate Calibration and Data Preprocessing

In this first step, univariate calibration was used to build regression models based on EDXRF data under the area of specific region of the element. However, this strategy presented unacceptable errors for all determined elements due to variability of matrix between the samples, what make this strategy inadequate to estimate the concentration values in the samples.

Partial least squares (PLS) multivariate models were used to predict the concentration values of Cu, K, Sr, and Zn in cocoa bean samples. For modeling the data, the whole EDXRF spectra were used for the calculations and are shown in Fig. 1. The data were modeled using the calibration and validation samples, as described in “Acquisition of LIBS Spectra and Chemometric Tools” section. The reference concentration obtained from ICP OES was the y vector.

Fig. 1
figure 1

Representative spectra obtained from cocoa bean sample using EDXRF

Full size image

The MVC1 calculates important figures of merit, such as limit of detection (LOD), limit of quantification (LOQ), and SECV, which allows to evaluate the efficiency of the regression models (Olivieri et al. 2004). The obtained models presented good results using one latent variable (LV), where the LOD and LOQ obtained were 1 and 3 mg kg−1 for Cu, 345 and 1036 mg kg−1 for K, 0.5 and 2 mg kg−1 for Sr, and 2 and 6 mg kg−1 for Zn. Moreover, the SECV was below the lowest concentration values in the validation samples for all determined elements. These values were 7 mg kg−1 for Cu, 1506 mg kg−1 for K, 4 mg kg−1 for Sr, and 7 mg kg−1 for Zn.

The predicted concentration values for these elements ranged from 16 to 17 mg kg−1 for Cu, 6602 to 8034 mg kg−1 for K, 8 to 12 mg kg−1 for Sr, and 36 to 38 mg kg−1 for Zn. Figure 2 shows the obtained trueness values for each element in all samples, which ranged from 85 to 106% for Cu, 94 to 100% for K, 109 to 120% for Sr, and 90 to 106% for Zn.

Fig. 2
figure 2

Trueness values (%) obtained for (a) Cu, (b) K, (c) Sr, and (d) Zn using PLS

Full size image

Other strategy evaluated was data fusion between EDXRF and LIBS combined with multivariate calibration to enhance the regression model for K. In this sense, the data fusion model was built using the specific selected spectra of the transition energy Kα (3.314 keV) of K from EDXRF, and the spectra of two emission lines of this element (K I 766.49 nm and K I 769.89 nm) from LIBS, as depicted in Fig. 3. After that, the data were autoscaled and the calibration model was calculated using PLS. However, due to lower sensitivity of the LIBS, the data fusion was not successful for Cu, Sr, and Zn.

Fig. 3
figure 3

The selected spectra of K from (a) EDXRF and (b) LIBS used for data fusion

Full size image

The SECV value obtained for this element was 872 mg kg−1. Figure 4 shows the SECV values obtained using single model EDXRF, and data fusion model employing EDXRF and LIBS simultaneously. The results obtained for K using data fusion strategy presented lower SECV when compared with those obtained by a single source (EDXRF), showing a good capability to predict the concentration values of this element. In addition, the obtained LOD and LOQ were 1899 mg kg−1 and 5697 mg kg−1, respectively. The predicted concentration values found ranged from 6053 to 8339 mg kg−1, and the trueness was 76 to 120% as depicted in Fig. 5.

Fig. 4
figure 4

The standard error of cross-validation (SECV) obtained for K using single EDXRF and data fusion model (EDXRF and LIBS)

Full size image
Fig. 5
figure 5

Trueness values (%) obtained for K employing data fusion strategy and PLS

Full size image

Therefore, considering the essential species for human health, all analyzed samples presented high concentration values of K and can be considered good source of this nutrient. Regarding micronutrients, Zn were found in higher concentration values than Cu and Sr.

Conclusions

Multivariate calibration and EDXRF and LIBS data were able to predict the mineral content (Cu, Sr, and Zn) by means of direct solid analysis. Whole spectra and PLS were employed to calculate the calibration model, which showed simplicity and presented good results, as low SECV, LOD, LOQ, and acceptable trueness. Moreover, data fusion employing EDXRF and LIBS was used for the determination of K, and it demonstrated that with the use of this strategy and PLS for modeling, the SECV was improved. In this sense, the use of EDXRF and LIBS combined with multivariate calibration represents an alternative to minimize matrix effect.