ABSTRACT
This paper reports the use of improved numerical approaches to modelling extraction profiles, and shows that the approach substantially reduces statistical prediction uncertainties compared to those obtained on the basis of a three-point extrapolation from the later part of the extraction curve. Numerical fitting of manually obtained polycyclic aromatic hydrocarbon extraction data to a spherical particle diffusion model showed uncertainties typically reduced by a factor of three (with extremes at 1.02 and 770). Application to pressurised fluid extraction study of pelletised poly(vinylchloride) containing 30 mass% di(2-ethylhexyl)phthalate also showed good improvements. However, this high precision data showed small but significant lack of fit resulting in residual correlation and visibly biased prediction (more so than simple extrapolation). Re-fitting and uncertainty estimation using a first-order autoregression approximation to the covariance matrix produced more realistic uncertainty estimates and closer parameter estimates and is accordingly recommended for treating residual correlation from other causes, but did not entirely alleviate the problem. Different shape models (spherical, plane sheet and cylindrical) were applied without accounting fully for fitting error, and particle size effects were eliminated by modelling a simple size distribution. However, an approximate model based on linearly concentration-dependent diffusion coefficient showed excellent fit, confirming concentration-dependence as the most likely cause. This semiempirical model led to an uncertainty in total extractable material, at 0.2% of the total extractable value (with allowance for correlation). This is potentially good enough for recovery estimation and correction in certification of reference materials for validation purposes.
ABSTRACT
The paper describes experiments for the evaluation of uncertainties associated with a number of chromatographic parameters. Studies of the analysis of vitamins by HPLC illustrate the estimation of the uncertainties associated with experimental "input" parameters such as the detector wavelength, column temperature and mobile phase flow-rate. Experimental design techniques, which allow the efficient study a number of parameters simultaneously, are described. Multiple linear regression was used to fit response surfaces to the data. The resulting equations were used in the estimation of the uncertainties. Three approaches to uncertainty calculation were compared--Kragten's spreadsheet, symmetric spreadsheet and algebraic differentiation. In cases where non-linearity in the model was significant, agreement between the uncertainty estimates was poor as the spreadsheet approaches do not include second-order uncertainty terms.
