Heptane, [emim] - American Chemical Society

Apr 21, 2009 - separation of aromatic and aliphatic hydrocarbons,21 and so forth. ... imidazolium methylsulfate ILs in the working range (0 and 15 ...
9 downloads 0 Views 313KB Size
4998

Ind. Eng. Chem. Res. 2009, 48, 4998–5003

Determination of Toluene, n-Heptane, [emim][EtSO4], and [bmim][MeSO4] Ionic Liquids Concentrations in Quaternary Mixtures by UV-vis Spectroscopy Jose´ S. Torrecilla,* Ester Rojo, Julia´n Garcı´a, Mercedes Oliet, and Francisco Rodrı´guez Department of Chemical Engineering, Faculty of Chemistry, UniVersity Complutense of Madrid, 28040 Madrid, Spain

This article reports a new computerized approach to the simultaneous determination of low concentrations of toluene, heptane and 1-ethyl-3-methylimidazolium ethylsulfate ([emim][EtSO4]) and 1-butyl-3-methylimidazolium methylsulfate ([bmim][MeSO4]) ionic liquids (ILs) in acetone using a correlation between their concentrations and UV-vis absorbance values. The essential information (absorbance data) from UV-vis spectrometer of quaternary mixtures in acetone was selected by a principal component analysis (PCA) and transferred into linear regressions (LRs) or neural network (NN) trained computer to estimate the concentrations. Such an integrated PCA/NN/UV-vis spectroscopy system is capable of estimating the concentrations of chemicals in acetone, based on the created models and patterns, without any previous phenomenological knowledge. The mean difference between the estimated concentrations using the PCA/NN/UV-vis and the real concentrations of toluene, heptane, [emim][EtSO4], and [bmim][MeSO4] is less than 2.5%. The PCA/ NN/UV-vis is one of the first reliable approaches that can be used in the ILs mixtures field to determine the concentration of two ILs and hydrocarbons in quaternary mixtures while at the same time having simple applications. The short time required by PCA/NN/UV-vis to measure all four chemicals makes it especially useful in processes where long sample preparation times are required. 1. Introduction In recent years, ionic liquids (ILs), due to their unique properties, have attracted increasing attention as replacements for conventional organic solvents in catalysis, separation processes, electrochemistry, in the manufacture of space telescopes, and in many other scientific and industrial fields.1,2 From a chemical engineering point of view, one of the most attractive features of ILs is their ability to be tailor-made for a specific purpose by careful selection of the cation, anion, or both.3,4 Given that there are about 106 simple and 1018 ternary ionic liquids (mixture composed of ionic liquids with common anion or cation), the IL design would be greatly assisted by theoretical models for estimation of thermophysical properties of intended combinations.5,6 In the petrochemical field, the search for more environmentally friendly solvents is the goal of many researchers.1 Focusing on the separation of toluene from toluene/heptane mixtures, this aim is achieved using ILs. Moreover, the investment and energy costs are considerably lower using ILs as solvents. Ionic liquids based on sulfate anion are suitable for this separation.7,8 Given that most ILs have a negligible vapor pressure, ionic liquids and other more volatile compounds could be separated from their mixtures by evaporation.7 To evaluate and control the extraction process rate, the measurement of chemical concentration is necessary. Recently, these measurements have been carried out by the interpolation in the physicochemical properties correlation of the mixtures (density, viscosity, refractive index, etc.), proton nuclear magnetic resonance (1H NMR) spectroscopy,9 gas chromatograph (GC),10 and so forth. All of these techniques are not adequate to measure/control any extraction online process because these require a relatively long sample preparation time. Therefore, to control and evaluate these processes online, an * To whom correspondence should be addressed. Tel.: +34 91 394 42 40. Fax: +34 91 394 42 43. E-mail: [email protected].

analytical technique with a sample preparation time less than the sampling time of the process is necessary.11,12 In previous works, independent applications of PCA technique and neural network (NN) model to UV-vis spectroscopy data to determine concentrations of hydrocarbon/IL binary mixtures has been carried out.13,14 The selection of this technique is because the imidazolium ring is UV active. In addition, given that some hydrocarbons and most ILs based on imidazolium cation have a similar absorbance profile, the use of powerful chemometric tools to deconvolute overlapping signals is necessary. The application of PCA, linear regressions (LR), NN, and UV-vis spectroscopy to determine the concentrations of chemicals of binary mixtures composed of hydrocarbon and simple IL was demonstrated in literature.15–18 Recently, given the properties of the mixtures of two ionic liquids and ternary ILs, these mixtures are being explored. As a result of this, different binary mixtures of ionic liquids have been used in chemical applications viz. in the manufacture of solar cells,19 as green reaction media and catalysts,20 in the separation of aromatic and aliphatic hydrocarbons,21 and so forth. In addition, some physicochemical properties of these mixtures such as osmotic coefficients,22 heat capacity, phase transitions,23 and so forth have been measured and reported in literature. As can be expected, in the light of these applications and researches, a few reliable methods to determine qualitatively and quantitatively ILs in mixtures have started to be proposed.24 In light of the successful results that were met in binary mixtures composed of 1-ethyl-3-methylimidazolium ethylsulfate and toluene using the NN model and PCA technique and continuing with the research on the mathematical techniques to facilitate the quantifications of chemical mixtures,13,14 both algorithms have been used in series to determine the concentrations of chemicals in systems with two ILs and two hydrocarbons. In particular, the aim of this work is to verify PCA/LR and PCA/NN as mathematical tools of UV-vis data treatment to establish the concentration in acetone of toluene, heptane, 1-ethyl-3-methylimidazolium ethylsulfate, and 1-butyl-3-methyl-

10.1021/ie8014044 CCC: $40.75  2009 American Chemical Society Published on Web 04/21/2009

Ind. Eng. Chem. Res., Vol. 48, No. 10, 2009

imidazolium methylsulfate ILs in the working range (0 and 15 ppm) in room experimental conditions with a sample preparation time of less than the sampling time of the process. 2. Material and Methods 2.1. Reagents, Solutions, and Instrumentation. In this work, 1-ethyl-3-methylimidazolium ethylsulfate ([emim][EtSO4] g95% purity, from Sigma-Aldrich Chemie GmbH), 1-butyl-3methylimidazolium methylsulfate ([bmim][MeSO4] g95% purity, from Sigma-Aldrich Chemie GmbH) ionic liquids, n-heptane (heptane g99% purity, water 0.110 and MPE < 30%). This is because the signals of chemicals are overlapped and these cannot be deconvoluted by linear models. Because of this, to separate and quantify the chemical concentrations, a powerful nonlinear model (NN) has been tested. 3.4. Neural Network Model. 3.4.1. Learning, Verification, and Validation Samples. The learning sample was used to optimize the NN parameters. The learning, verification (internal validation sample), and validation (external validation sample) samples are composed of data that characterize the measurement process. These have as many rows as variables necessary to model process (seven, one for each PC) and the same number of columns as the number of vectors to describe the system to be measured. Given that 7 PCs are necessary, the input layer consists of 7 nodes. Then the NN model is made up of 7 input nodes and four output neurons. Learning, verification, and validation samples present the same format. These were composed of 11 rows (7 principal components and 4 chemical concentrations). The learning sample consists of 1843 columns (80% of whole data). The verification sample has 461 columns (remaining 20%). These data were divided randomly in learning and verification samples.13 The validation sample was made up of 25 new mixtures composed of toluene, heptane, [emim][EtSO4], and [bmim][MeSO4] between 0 and 15 ppm (vide infra), as seen in Figure 1. 3.4.2. Optimization of NN Parameters. The calculation process followed to optimize the NN is described in the previous “Optimization process of the NN” section. To reach the highest value of mean correlation coefficient and the lowest values of both MPE and iterations number, the parameters hidden neuron number, Lc, Lcd, and Lci were fixed to 20, 0.5, 0.018, and 51, respectively.13 Obviously, the parameter number of the NN model increases with the hidden neuron number, and by using more parameters the model could reproduce with better statistical results the real samples. However, as mentioned previously, to prevent the overfitting and overtraining effects, 20 hidden neurons were selected.29,30

3.4.3. Evaluation Predictive Capacity of Optimized NN Model. Internal Validation of Optimized NN. To test the optimized NN performance, an internal validation process was carried out (vide supra).32,33 Using the verification sample, the estimations were calculated and with these data, a mean correlation coefficient of 0.95 and a mean MPE less than 1.5% were reached. In addition, a residual analysis was carried out by graphic and numerical methods. These residuals were calculated as the percentage of the difference between the real and estimated values. These values versus every data set that form the verification sample are shown in Figure 4. In all cases, the residuals are less than 4%. These estimations are slightly better when [emim][EtSO4] concentrations are calculated. The residuals were also analyzed by two numerical methods. First, this was carried out using the Durbin-Watson test based on the assumption that successive errors are correlated, as shown in eq 4.36 Given that the Durbin-Watson coefficient (DW) values are about 2 (2, 1.97, 1.87, and 2.05 for toluene, heptane, [emim][EtSO4], and [bmim][MeSO4], respectively), the residuals and estimated concentrations are not correlated. Finally, the correlation coefficients of four estimated concentrations of chemicals and data set order were equal to 1 × 10-5, 3 × 10-3, 9 × 10-4, and 2 × 10-4 for toluene, heptane, [emim][EtSO4], and [bmim][MeSO4], respectively. Therefore, there is no relation between them. This is in agreement with the Durbin-Watson test results (vide supra). From the residuals point of view, the NN model is adequate to estimate the toluene, heptane, [emim][EtSO4], and [bmim][MeSO4] concentrations at the same time, that is, to solve the overlapping effect between the four chemicals. M

∑ (e - e t

DW )

t)2

(4)

M



2 t-1)

e2t

t)2

In eq 4, e is the residual value at a given order (tth) with respect to the previous time, and M is the total number of verification sets (461 data sets). External Validation of Optimized NN. Finally, 25 new solutions of toluene, heptane, [emim][EtSO4], and [bmim][MeSO4] in acetone inside the working range were prepared and monitored only once to make an external validation sample. The chemical concentration values of these are different from those used in the learning and verification processes, as shown in Figure 1. In this process, the mean MPE is less than 2.5%. The differences between real and predicted concentrations versus data sets are shown in Figure 5. In most cases, these differences are less than 6% to toluene and [bmim][MeSO4] and even lower when the NN model estimates heptane and [emim][EtSO4]

5002

Ind. Eng. Chem. Res., Vol. 48, No. 10, 2009

Figure 4. NN model performance (internal validation process). Estimated concentrations of a) toluene (R2 > 0.97), b) heptane (R2 > 0.90), c) [emim][EtSO4] (R2 > 0.97), and d) [bmim][MeSO4] (R2 > 0.95).

Figure 5. NN model performance (external validation process). Estimated concentrations of a) toluene (R2 > 0.92), b) heptane (R2 > 0.90), c) [emim][EtSO4] (R2 > 0.97), and d) [bmim][MeSO4] (R2 > 0.95).

4. Conclusions concentrations. All R2s between real and estimated values are slightly worse than those calculated in the internal validation (0.92, 0.90, 0.97, and 0.95 for toluene, heptane, [emim][EtSO4], and [bmim][MeSO4], respectively). From the residuals point of view, DW values are about 2 (1.7, 2.16, 2.02, and 1.61 for toluene, heptane, [emim][EtSO4], and [bmim][MeSO4], respectively) and the correlation coefficients of estimated and data set order were equal to 4 × 10-2, 7 × 10-3, 4 × 10-5, and 3 × 10-3 for toluene, heptane, [emim][EtSO4], and [bmim][MeSO4], respectively. Although the statistical results are worse in the external in relation to internal validation, PCA/NN/UV-vis is still adequate to estimate the four chemical concentrations in the working range studied.

A PCA/NN/UV-vis approach has been optimized and validated. Almost 800 different mixtures of toluene, heptane, [emim][EtSO4], and [bmim][MeSO4] in acetone have been made and monitored by UV-vis spectroscopy. Then to form the learning, verification, and validation samples to be used in linear and nonlinear models, the dimensionality of data was reduced from 710 absorbance values to 7 principal components by the PCA technique. A neural network based on an MLP algorithm (7 input, 20 hidden, and 4 outputs neurons) made up of the sigmoid transfer function and Bayesian regularization training function (Lc, Lcd, and Lci equal to 0.5, 0.018, and 51, respectively) has been used. To verify how adequate the PCA/NN/UV-vis system is to estimate the concentrations of chemicals, external and internal

Ind. Eng. Chem. Res., Vol. 48, No. 10, 2009

validation processes were carried out. In both cases, the NN is able to estimate these values with mean MPE values less than 1.5 and 2.5% and with mean correlation coefficients higher than 0.97 and 0.95, respectively. The PCA/NN/UV-vis can be adapted to estimate toluene, heptane, [emim][EtSO4], and [bmim][MeSO4] concentrations in quaternary mixtures simultaneously, that is it is able to deconvolute the contribution of each chemical of the mixture, without using previous phenomenological knowledge or mechanistic premises and takes less time than other classical methods to determine their concentrations. Therefore, the PCA/NN/UV-vis is very interesting for further applications to digital control or measurement devices. Acknowledgment The authors are grateful to the Spanish “Ministerio de Ciencia e Innovacio´n” for financial support for project CTQ2008-01591 and a Ramo´n y Cajal research contract of Jose´ S. Torrecilla. Literature Cited (1) Plechkova, N. V.; Seddon, K. R. Applications of ionic liquids in the chemical industry. Chem. Soc. ReV. 2008, 37, 123. (2) Nelson, W. Green solVents for chemistry: PerspectiVes and practice; Oxford University Press: New York, 2003. (3) Freemantle, M. Designer solvents - Ionic liquids may boost clean technology development. Chem. Eng. News 1998, 76, 32. (4) Borra, E. F.; Seddiki, O.; Angel, R.; Eisenstein, D.; Hickson, P.; Seddon, K. R.; Worden, S. P. Deposition of metal films on an ionic liquid as a basis for a lunar telescope. Nature (London) 2007, 447, 979. (5) Palomar, J.; Ferro, V. R.; Torrecilla, J. S.; Rodrı´guez, F. Density and molar volume predictions using COSMO-RS for ionic liquids. An approach to solvents design. Ind. Eng. Chem. Res. 2007, 46, 6041. (6) Palomar, J.; Torrecilla, J. S.; Ferro, V. R.; Rodrı´guez, F. Development an a priori ionic liquid design tool. 1. Integration of a novel COSMO-RS molecular descriptor on neural network. Ind. Eng. Chem. Res. 2008, 47, 4523. (7) Earle, M. J.; Esperanca, J. M. S. S.; Gilea, M. A.; Lopes, J. N. C.; Rebelo, L. P. N.; Magee, J. W.; Seddon, K. R.; Widegren, J. A. The distillation and volatility of ionic liquids. Nature (London) 2006, 439, 831. (8) Meindersma, G. W.; Podt J. G.; Meseguer M. G.; De Haan A. B. S. In Ionic Liquids IIIB, Fundamentals, Progress, Challenges, and Opportunities; Rogers, R. D., Seddon, K. Eds.; American Chemical Society, WA, 2005; Chapter 5, p 57-71. (9) Arce, A.; Rodriguez, O.; Soto, A. A comparative study on solvents for separation of tert-amyl ethyl ether and ethanol mixtures. New experimental data for 1-ethyl-3-methyl imidazolium ethyl sulfate ionic liquid. Chem. Eng. Sci. 2006, 61, 6929. (10) Meindersma, G. W.; Podt, A. J. G.; De Haan, A. B. S. Ternary liquidsliquid equilibria for mixtures of toluene + n-heptane + an ionic liquid. Fluid Phase Equilib. 2006, 247, 158. (11) Palancar, M. C.; Aragon, J. M.; Miguens, J. A.; Torrecilla, J. S. Application of a model reference adaptive control system to pH control. Effects of lag and delay time. Ind. Eng. Chem. Res. 1996, 35, 4100. (12) Palancar, M. C.; Arago´n, J. M.; Torrecilla, J. S. pH-Control system based on artificial neural networks. Ind. Eng. Chem. Res. 1998, 37, 2729. (13) Torrecilla, J. S.; Ferna´ndez, A.; Garcı´a, J.; Rodrı´guez, F. Determination of 1-ethyl-3-methylimidazolium ethylsulfate ionic liquid and toluene concentration in aqueous solutions by artificial neural Nnetwork/ UV spectroscopy. Ind. Eng. Chem. Res. 2007, 46, 3787. (14) Torrecilla, J. S.; Rojo, E.; Garcı´a, J.; Rodrı´guez, F. Principal component analysis/UV spectroscopy for the determination of 1-ethyl-3methylimidazolium ethylsulfate ionic liquid and toluene concentrations in aqueous solutions. Ind. Eng. Chem. Res. 2008, 47, 4025. (15) Ni, Y.; Zhang, G.; Kokot, S. Simultaneous spectrophotometric determination of maltol, ethyl maltol, vanillin and ethyl vanillin in foods by multivariate calibration and artificial neural networks. Food Chem. 2005, 89, 465.

5003

(16) McCleskey, S. C.; Floriano, P. N.; Wiskur, S. L.; Anslyn, E. V.; McDevitt, J. T. Citrate and calcium determination in flavored vodkas using artificial neural networks. Tetrahedron 2003, 59, 10089. (17) Sun, L. X.; Reddy, A. M.; Matsuda, N.; Takatsu, A.; Kato, K.; Okada, T. Simultaneous determination of methylene blue and new methylene blue by slab optical waveguide spectroscopy and artificial neural networks. Anal. Chim. Acta 2003, 487, 109. (18) Yin, C.; Shen, Y.; Liu, S.; Yin, Q.; Guo, W.; Pan, Z. Simultaneous quantitative UV spectrophotometric determination of multicomponents of amino acids using linear neural network. Comput. Chem. 2001, 25, 239. (19) Mohmeyer, N.; Kuang, D.; Wang, P.; Schmidt, H. W.; Zakeeruddin, S. M.; Gratzel, M. An efficient organogelator for ionic liquids to prepare stable quasi-solidstate dye-sensitized solar cells. J. Mater. Chem. 2006, 16, 2978. (20) Shi, F.; Gu, Y.; Zhang, Q.; Deng, Y. Development of ionic liquids as green reaction media and catalysts. Catal. SurV. Asia 2004, 8, 179. (21) Arce, A.; Earle, M. J.; Katdare, S. P.; Rodrıguez, H.; Seddon, K. R. Application of mutually immiscible ionic liquids to the separation of aromatic and aliphatic hydrocarbons by liquid extraction: A preliminary approach. Phys. Chem. Chem. Phys. 2008, 10, 2538. (22) Calvar, N.; Gonza´lez, B.; Domı´nguez, A.; Macedo, E. A. Osmotic coefficients of binary mixtures of four ionic liquids with ethanol or water at T ) (313.15 and 333.15) K. J. Chem. Thermodyn. DOI: 10.1016/ j.jct.2008.08.011. (23) Pavlechko, Y. U.; Kabo, G. Y. Heat capacity and phase transitions of a mixture of ionic liquids [C2 mim]NTf2 + [C6mim]NTf2. Russ. J. Phys. Chem. A 2008, 82, 1412. (24) Stepnowski, P.; Markowska, A. How to analyze ionic liquids anions? Investigation into the application of suppressed and non-suppressed ion chromatography. Aust. J. Chem. 2008, 61, 409. (25) Torrecilla, J. S.; Garcia, J.; Rojo, E.; Rodriguez, F. Estimation of toxicity of ionic liquids in leukemia rat cell line and acetylcholinesterase enzyme by principal component analysis, neural networks and multiple lineal regressions. J. Hazard. Mater. DOI: 10.1016/j.jhazmat.2008.08.022. 2008. (26) Heberger, K.; Csomos, E.; Simon-Sarkadi, L. J. Principal component and linear discriminant analyses of free amino acids and biogenic amines in Hungarian wines. Agric. Food Chem. 2003, 51, 8055. ´ lvarez, P. J.; Montilla, A.; Corzo, (27) Montan˜e´s, F.; Fornari, T.; Martı´n-A N.; Olano, A.; Iba´n˜ez, E. Selective fractionation of disaccharide mixtures by supercritical CO2 with ethanol as co-solvent. J. Supercrit. Fluids 2007, 41, 61. (28) NIST/SEMATECH e-Handbook of Statistical Methods, http:// www.itl.nist.gov/div898/handbook/, 2008. (29) Tetko, V.; Livingstone, D. J.; Luik, A. I. Neural network studies. 1. Comparison of overfitting and overtraining. J. Chem. Inf. Comput. Sci. 1995, 35, 826. (30) Demuth, H.; Beale, M.; Hagan, M. Neural network toolbox for use with MATLAB User’s guide, Version 4.0.6; Ninth printing revised for version 4.0.6 (Release 14SP3), 2005. (31) Vacic, V. Summary of the training functions in Matlab’s NNtoolbox. http://www.cs.ucr.edu/_vladimir/cs171/nn_summary.pdf, (accessed 2005). (32) Gramatica, P. Principles of QSAR models validation: Internal and external. QSAR Comb. Sci. 2007, 26, 694. (33) Guidance document on the Validation of (quantitatiVe) structure actiVity relationship [(Q)SAR] models; No. 69; OECD Series on testing and assessment; Organisation of Economic Cooperation and Development: Paris, France, 2007. (34) Kaiser, H. F. The application of electronic computers to factor analysis. Educ. Psychol. Meas. 1960, 20, 141. (35) Schott, J. R. A high-dimensional test for the equality of the smallest eigenvalues of a covariance matrix. J. MultiVar. Anal. 2006, 97, 827. (36) Chatterjee, S.; Hadi, A. S. Regression Analysis by Example, 4th edition, Wiley Series in Probability and Statistics; Wiley-Interscience: N.J., 2006.

ReceiVed for reView September 17, 2008 ReVised manuscript receiVed March 17, 2009 Accepted March 24, 2009 IE8014044