Intermolecular interactions of substituted benzenes on multi-walled carbon nanotubes grafted on HPLC silica microspheres and interaction study through artificial neural networks

2.1 Materials and reagents

Pristine MWCNTs (o.d. 20–50 nm, i.d. 5–10 nm, length 10–20 μm, specific surface area 60 m² g⁻¹, purity > 95 wt%) were obtained from Cheap Tubes Inc. (Brattleboro, Vermont, USA). Silica microspheres (average particle size 5 μm, average pore diameter 100 Å, surface area 200 m² g⁻¹) were purchased from Agilent Technologies Italia S.p.A (Cernusco sul Naviglio, Italy). Activated carbon (Norit, 40–140 mesh) was supplied by Norit Italia S.p.a. (Ravenna, Italy). HCl (37%, w/w, ultrapure) was from Merck Chemicals (Milan, Italy). Triton X-100, chloroform, HPLC grade acetonitrile (ACN), polybutadiene (PB, M_n 1530–2070) were supplied by Sigma–Aldrich (Milan, Italy). Ultrapure water (resistivity 18.2 MΩ cm⁻¹) was produced by a Milli-Q system (Millipore, Milan, Italy). Analytical standards were purchased from Sigma–Aldrich and Carlo Erba Reagenti (Milan, Italy).

2.2 Liquid chromatography

MWCNT-silica and AC-silica were slurry packed into 3.9 × 300 mm HPLC columns, as described elsewhere (Speltini et al., 2013a). The chromatographic system consisted of a Shimadzu (Milan, Italy) LC-20AT solvent delivery module equipped with a DGU-20A3 degasser and interfaced with a SPD-20A UV detector. Injection volume was 20 μL. For elution details see figures caption. For the intermolecular interaction study, all runs were performed in ACN–water (40:60), flow rate 1 mL min⁻¹ and a commercial Hypersil C18 column (4.6 × 250 mm, 5 μm) from Varian (Turin, Italy) was used for comparison under the same elution conditions. Injections of standard solutions in ACN were always performed in triplicate.

2.3 Characterization of the synthesized materials

SEM images of the powders were acquired in high vacuum using a Zeiss EVO®-MA10-HR microscope. Particle size determination was performed on the raw powders, while morphology was studied on gold sputtered conductive samples. Particle size distribution curves were obtained by the Digital Micrograph (^™) 3.11.0 software. Specific surface area measurements on the samples were performed with a Sorptomatic 1990 by Thermo Fisher Scientific. About 0.350 g of powders was charged in the glass sample holder and degassed at 250 °C for 60 h. Subsequently, samples were cooled down at −196 °C and two adsorption–desorption cycles followed by a last adsorption run were performed (B.E.T. method, analyzing gas: N₂, 50 points for run; blank done in He before the 1st adsorption run).

2.4 Intermolecular interaction study

At first glance, a linear model was chosen for the correlation between k′ and analyte properties. We adopted a method based on the multiple linear regression having the formula: $$\log k^{\prime }=y_0+{\displaystyle \sum _{i=1}^I}{c}_i{X}_i$$ where X_i is the i-th property of each compound, c_i is the unnormalized correlation coefficient of the i-th property and y₀ is the residual. Otherwise, in the case of ANN, the relationship between experimental k′ and analyte properties was studied through three-layered ANNs. The neural networks used contain one linear input layer, one sigmoidal hidden layer and one linear output neuron (see figure in Supplementary data). Several topologies of networks were tested and no useful fittings were found with a linear hidden layer, this indicating a non-linear relationship of the input parameters with the retention factor. The neural networks were simulated by using the Joone freeware program (http://sourceforge.net/projects/joone/).

The theoretical capacity factor k′ of each analyte is given by the overall fit function of our neural networks with H hidden sigmoidal neurons, I inputs: $$\log k^{\prime }={\displaystyle \sum _{h=1}^H}\left[\sigma \left({\displaystyle \sum _{i=1}^I}{w}_{\mathit{ih}}{X}_i-{\beta }_h\right)\right]w_h$$ where the sigmoidal function σ is defined as: $$\sigma (a)=\frac{1}{1+e^{-a}}$$ and β_h are the biases of hidden neurons; w_h are the weights between the input neurons (the descriptors) and the hidden neurons; w_ih are the weights between the hidden neurons and the output neurons (log k′); X_j are the i-th properties of the analyte. All matrices and arrays present in the trained networks are listed in Supplementary Data. All inputs were chosen to be easily quantifiable by first-principle DFT calculations. The following properties are suitable for a good correlation with the experimental data: dipole moment (Debye), molecular mass (amu), rotational energy (cal mol⁻¹ K⁻¹), HOMO energy (Ha), LUMO energy (Ha), zero-point energy (ZPE, Ha particle⁻¹), and isotropic polarizability (Bohr³). The training of neural networks was obtained by minimizing the root-mean-square error (RMSE). The number of neurons in the hidden layer was kept the lowest possible compatibly with a low RMSE in order to reduce the overfitting. The training procedure was divided into several steps decreasing gradually the learning rate (the variation of the neuron weights in each step) and momentum (the contribution of the last weight change in the previous iteration to the weight change in the current iteration).

We found that two different topologies of networks (different in the number of hidden and input neurons) best fit the behavior of CNT and octadecyl phases (see Section 3.3).

3.1 Characterization and properties of the obtained materials

The radiation-induced bonding of MWCNTs to the PB-coated silica gave a final composite material consisting of microspheres. These simply act as support/spacer for the CNT bundles, enabling them to interact with the solutes migrating in the eluent flow. The morphology of the CNT-based hybrid material can be appreciated from the SEM images reported in Fig. 1, where it is well evident that the CNTs “hanks” cover some silica 5 μm microspheres and the interstices among them. The AC-silica sample SEM analysis (not reported) shows only the spherical morphology of the silica particles. The particle size of the PB coated particles as obtained by the analysis of the SEM images (see Table 1) is similar for the two modified kinds of microspheres and 5% higher than the pristine silica particles. Concerning the surface area, there is a decrease due to coating, with the MWCNT sample showing the lowest value (a 34.5% variation with respect to the pristine sample). This same sample presents a little higher pore width with respect to the AC sample (+12%), due to the different layering of the coating agent on the silica spheres.

Close

Figure 1

Scanning electron microscopy images of MWCNT-silica microspheres. The magnification reported in part b is to better show the morphology of the CNT structures.

Export to PPT

In the new hybrid material the chemical bonding of MWCNTs preserved the nanotube structure after the grafting, as indicated by the moderate decrease of the relative intensity of the Raman G-band in MWCNT-functionalized powders with respect to the pristine sample – a full discussion is reported in Speltini et al. (2013a). It was experimentally found a value of IG/ID ratio equal to 0.70 for pristine powders and 0.63 for MWCNT-silica; a similar low defect ratio has been already observed in the case of physical adsorption of SWCNTs onto amino-silica particles (Fujigaya et al., 2011), testifying the mildness of the proposed immobilization method. The preservation of CNT pristine structure is fundamental to evaluate its real interaction with the solute. Actually, the nanotubes here immobilized by gamma-radiation can be regarded as MWCNTs whose carbonaceous structure is highly preserved; it is worth noting that the potential influence of the metallic impurities present in the commercial material on molecules retention times is minimized due to the purification procedure based on HCl washings. As already discussed (Merli et al., 2010; Speltini et al., 2010), this treatment enables a substantial removal of the metal fraction without altering the CNT morphology.

3.2 Intermolecular interactions of substituted benzenes

Separation in chromatography is generally driven by selective association with a specific solute with the stationary phase followed by dissociation, resulting in different retention times depending on the analyte nature; this evidence can be indirectly exploited to gain important information about the sorption process. Indeed, the markedly different capacity factors (mean values of three injections, relative standard deviation, RSD 0.1–3.8%) obtained for a set of substituted benzenes on the CNT-column (see Supplementary data) prove selective interactions between the nanotubes and each analyte. The role of the CNTs can be better appreciated comparing these results with those obtained on a commercial C18 stationary phase, that clearly indicate a different elution order on the two columns (see Supplementary data). This is expectable on the basis of the different base-structure and properties of each sorbent, and is in line with the recent literature (Liang et al., 2010; Kwon and Park, 2006) where the different behavior among CNTs and more traditional reversed-phase materials was ascribed to the high number of interplays that CNTs are able to establish with the molecules with respect to the C18 alkyl strands. Indeed, while on the latter the adsorption affinity is recognized to be driven by mere hydrophobic interactions (Liang et al., 2010), on the contrary, π–π stacking, electrostatic (i.e. dipole–dipole) and dispersive forces, and molecular sieving also take place with CNTs: as a result, the combination of two or more interactions modulates the selectivity (André et al., 2009; Chambers et al., 2011). This is supported also by a recent paper proving that in the case of amino-acids, adsorption on SWCNTs is driven by a mechanism different from classical hydrophobic interaction, that only gives a limited contribution (Basiuk and Bassiouk, 2008).

It is well known that in liquid chromatography the role and the nature of the eluent are basilar (Basiuk and Bassiouk, 2008; Kwon and Park, 2006) and the retention time of a solute results from the combination of the repartition in the liquid eluent (dependent on solute physical–chemical properties) and the affinity for the sorbent, generally governed by hydrophobicity in the case of usual reversed-phase materials (Liang et al., 2010). For this reason, elution was performed maintaining the same mobile phase for the two columns, so that the different adsorption behavior of the solutes on the two sorbents was ascribable to the neat interaction with either the nanotubes or the C18 strands, and a direct comparison between the two data set of capacity factors was possible.

The key role of the nanotubes in the adsorption process was demonstrated by comparison on a blank control column filled with PB-silica (prepared by the same procedure omitting CNTs) and on a column packed with AC-silica particles. As verified for alkylbenzenes and aromatic acids (Speltini et al., 2013a), the contribution of the PB-silica was negligible, so that separation on the CNT phase can be ascribed essentially to the intermolecular interplays between solutes and nanotubes.

The role of the MWCNTs anchored by gamma-radiation has been further proved by substituting the nanotubes with the same amount amorphous carbon (5 wt%) grafted onto the silica microspheres following the same gamma-radiation procedure. Results unquestionably evidenced the different morphology and physical–chemical properties of the two carbon-made materials, as in the case of BTX, so strongly adsorbed on the carbon to require very high percentage of ACN for the desorption (up to 90% v/v) that occurred with a broadening band (Fig. 2). These results further prove the goodness of the method used for the preparation of a CNT-containing hybrid material really basing its properties and binding affinity on carbon nanotubes. This is very important in the framework of an intermolecular interaction study.

Close

Figure 2

Elution profiles obtained on (a) the MWCNT-column (3.9 × 300 mm) and on (b) the AC-column (3.9 × 300 mm). Peak identification: benzene (1), toluene (2), p-xylene (3), naphthalene (4). Mobile phase: (a) ACN–water (40:60), flow rate 1 mL min⁻¹; (b) isocratic ACN 50% for 9 min, linear gradient to ACN 90% in 11 min, isocratic ACN 90% for 40 min. Detection: 254 nm.

Export to PPT

Indeed, the preparation method is a crucial step, as not all the materials and methods proposed in the last years are suitable to selectively study the neat interaction between CNTs and solutes. For instance, in polymeric matrices, consisting of CNTs embedded in the polymer phase, this last plays a major role in the separation/adsorption process (Aqel et al., 2012; Chambers et al., 2011; Li et al., 2005), thus shading the actual function of the nanomaterial.

To verify which variables were significantly involved in the retention process on the CNT stationary phase and to provide a predictive model, a computational approach has been elaborated. The correlation was made by using the log k′ instead of k′, as reported in the literature (Gianis and Tsantili-Kakoulidou, 2013; Héberger, 2007; Kaliszan, 2007). The log k′ values for each column are reported in Table 2. With this method, two analytes were excluded for each stationary phase due to their low k′ values (phenol and 2,4-dinitrophenol for CNT; 2,4-dinitrophenol and 4,6-dinitro-ortho-cresol for C18).

All the descriptors were normalized between 0 and 1, in order to have a better understanding of the weights calculated by the model (see Supplementary data). Even if the elution order was different on the two studied stationary phases, a strong correlation exists between the log k′ values observed on the two columns (see Supplementary data), indeed only compounds having entries 3, 10 and 25 differ significantly.

For the ANN modeling, the total set was divided into two sets, the first was used as training set while the remaining data were used as validation set. At this purpose, some criteria were tested in order to choose the proper training set. The first criterion was to select the most retained compounds (highest k′), as it was supposed that for these molecules the interactions with the stationary phase are maximized. This criterion was abandoned because the validation set showed a bad match with the experimental results, thus indicating that the model created did not match the experimental data. The second criterion for the training set choice was to select only few compounds covering the total span of log k′ values. The trained networks, for both stationary phases, were created with 5 sigmoidal neurons in the hidden layer and showed interesting results with the fitting of the validation set.

In order to simplify the model and to avoid (or reduce) overfittings, a simplification of the network was achieved. The first simplification was made by decreasing the number of descriptors. The method used to find out the less relevant descriptor was based on the sum of absolute values of network multipliers (Despagne and Massart, 1998; Nord and Jacobsson, 1998; Olden and Jackson, 2002). For CNT it was clear that the least important descriptors were molecular weight and zero point energy, while for the C18 stationary phase the HOMO and LUMO parameters had low weights. The networks were trained again without those descriptors, and the results showed few differences with the model based on all the descriptors. For the CNT model the agreement with the validation set was good except for some compounds that were not well represented in their chemical classes, such as benzonitrile, benzaldehyde and ethyl benzoate. In fact, these compounds were the only representative molecules for their chemical class tested in our experiments. The training set was thus modified by adding these compounds. At the end, the trained network showed a good correlation with experimental data (see Fig. 3a), as indicated by the good correlation coefficient (R²). Moreover, the number of hidden neurons was reduced from 5 to 3, without affecting the correlation. The CNT stationary phase model resulting from neural networks showed how LUMO position and polarizability are the most important factors affecting the retention time.

Close

Figure 3

Calculated vs. experimental log k′ values for (a) CNT and (b) C18 ANNs (see text for details).

Export to PPT

The neural network model for the C18 stationary phase is acceptable (R² 0.6975, see Fig. 3b) with dipole and polarizability as the most important factors. In this case, the decrease of the network from 5 to 3 neurons reduced significantly the correlation with the validation set.

The models obtained by neural networks were compared with those aroused from multiple linear regression (with respect to neural network model, the multiple linear regression is comparable to the behavior of one neuron having linear response). The parameters (c_i) obtained from multiple linear fitting are gathered in Table 3. As it can been seen from Fig. 4, the R² resulted to be 0.78 for the CNT column and 0.77 for the C18 one, indicating a poor modeling (see Supplementary data for a complete list of parameters).

Close

Figure 4

Calculated vs. experimental log k′ values for (a) CNT and (b) C18 obtained by multiple linear regression.

Export to PPT

The linear and ANN models showed some similarities and differences. Starting from the model obtained for the CNT stationary phase, both models showed the importance of the polarizability, while the less important descriptor was molecular weight.

The principal difference for the ANN model is the importance given to the frontier orbitals, in particular LUMO. The importance of frontier orbitals for the modeling of interactions seems reasonable because CNT orbitals might interact with the analytes. On the contrary, linear models cannot take into account interactions having an energy range like those arousing from electron donor and acceptor properties. In the case of the C18 material, the ANN model did not show particular improvements with respect to the multiple linear model.

3.3 Chromatographic tests

As above anticipated, the CNT-based material turned out to be useful also for analytical applications due to its chromatographic properties. Indeed, the material was tested as stationary phase for separation of various aromatics among those above selected for the computational study. The column-to-column reproducibility was assessed on independently synthesized batches (Speltini et al., 2013a).

The MWCNT-column was able to selectively retain mono-substituted benzenes bearing different chemical groups on the aromatic ring, providing symmetrical peaks and baseline resolution in about 13 min run times (see chromatogram in Supplementary data).

The chromatographic profile in Fig. 5 refers to a sample containing eight aromatics; specifically, separation of toluene and p-xylene accounts for methyl group selectivity, while the methylene group is the discriminating factor in the separation of toluene and ethylbenzene. The overall interaction with the nanotubes for ethylbenzene and p-xylene, that are C₈H₁₀ isomers (see Supplementary data), prevented their chromatographic resolution. These two structural isomers resulted hard to be separated also on the C18 sorbent, as shown in Fig. 5 b with retention times up to 60 min under the same elution conditions applied to the CNT column. Notice that the C18 sorbent was unable to separate toluene and chlorobenzene; moreover, the elution order of the analytes was different (Fig. 5b). These facts well evidence the different retention process taking place on the two chromatographic materials.

Close

Figure 5

Chromatograms obtained for a mixture of benzene compounds (20–100 mg L⁻¹) on (a) the MWCNT-column (3.9 × 300 mm) and on (b) the C18 column (4.6 × 250 mm). Peak identification: phenol (1), nitrobenzene (2), benzene (3), ethylbenzoate (4), toluene (5), chlorobenzene (6), p-xilene (7), ethylbenzene (8). Mobile phase: ACN–water (35:65), flow rate 1 mL min⁻¹. Detection: 254 nm.

Export to PPT

As shown in Fig. 6, on CNTs chlorinated benzenes were separated in isocratic elution, although with a certain peak tailing for compound 5, as well the two positional isomers, that is o-dichlorobenzene and p-dichlorobenzene (resolution, R_s 0.7, selectivity factor, α 1.12).

Close

Figure 6

Chromatograms obtained for a mixture of (poly)chlorobenzenes (50 mg L⁻¹) on (a) the MWCNT-column (3.9 × 300 mm) and on (b) the C18 column (4.6 × 250 mm). Peak identification: benzene (1), chlorobenzene (2), o-dichlorobenzene (3), p-dichlorobenzene (4), 1,2,4-trichlorobenzene (5). Mobile phase: ACN–water (40:60), flow rate 1 mL min⁻¹. Detection: 210 nm.

Export to PPT