Modelagem multivariada para determinação de propriedades físico-químicas de petróleo e para quantificação de edulcorantes artificiais em adoçantes de mesa

Abstract

The development of analytical methods that are fast, robust and low cost is one of the major focuses of modern analytical chemistry in order to perform quantitative inferences. The application of spectroscopic techniques associated with multivariate modeling has shown significant growth over the last years. This work aims at the development of analytical methodologies, by application of different spectroscopic techniques combined with regression by Partial Least Squares (PLS), for two different applications. The former involves the determination of some physicochemical properties of crude oil and the latter the determination of concentration of four artificial sweeteners in solid tabletop sweeteners samples. Petroleum is basically formed by different hydrocarbons families with a huge range of compounds. Due to the complex composition of the matrix a large number of high costs and time-consuming physicochemical analyses are required. In this first part, PLS models were built from 1H Nuclear Magnetic Resonance data (1H-RMN), in order to determine the following properties: API gravity, Carbon Residue (CR), Wax Appearance Temperature (WAT) and Simulated Distillation (SimDis). The curves obtained by this methodology were called Boiling Points Estimated by Chemometrics (BPEC). In the second part of this work, multivariate models were built by the application of PLS algorithm to Raman spectroscopy data and Near Infrared (NIR) spectroscopy to quantify Aspartame (ASP), Cyclamate (CIC), Saccharin (SAC) and Acesulfame-K (ACSK) in tabletop sweetener powder forms. Therefore, the reference method used to provide information regarding the concentration of analytes was the Liquid Chromatography coupled with Mass Spectrometry (LC-MS). The NIR-PLS models were more effective for the prediction of ASP, CIC and SAC, while Raman-PLS model presented better results to ACSK. All models, for both studies, were evaluated based on the following modelling parameters: number of Latent Variables (LVs), correlation coefficient for Cross Validation and Prediction (R2vc and R2p, respectively), Root Mean Square Error for Cross Validation and Prediction (RMSECV and RMSEP, respectively). Finally, the residuals of each model built were submitted to statistical tests for evaluation of systematic errors and tendency errors.