ISSN 0023-1584, Kinetics and Catalysis, 2006, Vol. 47, No. 4, pp. 537–548. © MAIK “Nauka /Interperiodica” (Russia), 2006.
Original Russian Text © A.L. Pomerantsev, O.Ye. Rodionova, 2006, published in Kinetika i Kataliz, 2006, Vol. 47, No. 4, pp. 553–565.
537
Data analysis and mechanistic studies for complex
chemical processes are the most important areas of
chemical kinetics. Estimating kinetic parameters from
experimental data or, in other words, solving the inverse
problem of chemical kinetics was formed as an individ-
ual area in the 1970s–1980s. Mathematicians [1, 2], as
well as kineticians [3, 4], have taken part in the devel-
opment of this area. Two main, fundamentally different
approaches to the problem of kinetic data analysis can
be distinguished. In the so-called soft approach [5–13],
experimental data are described in terms of a linear
multivariate model valid in a limited range of condi-
tions. In this case, it is not necessary to know the mech-
anism of the process. However, this approach does not
always provide the desired accuracy. The other
approach uses so-called hard physicochemical model-
ing [4, 14], which is based on fundamental kinetic prin-
ciples and allows parameters to be estimated with a
high degree of accuracy. However this method is appli-
cable only when a model of the process is known a pri-
ori. Both approaches have strong and weak points, and
both have advocates and opponents. Traditionally, Rus-
sian researchers develop the hard approach [15], while
Western researchers prefer soft methods [16].
The problem of data interpretation, model construc-
tion, and prediction of unknown values (called calibra-
tion for brevity) is among the oldest but still challeng-
ing scientific problems. Since Gauss (1794), this prob-
lem has been attacked using regression analysis. The
basic principle of this method is minimizing the devia-
tion of the model from experimental data (least-squares
method) [17]. The development of this approach,
including principal component analysis (1901) [18],
the maximum likelihood method (1912) [19], ridge
regression (1963) [20], and projection on latent struc-
tures (PLS,1975) [21], has made it applicable to com-
plex, ill-posed problems. However, all of these methods
provide predictions as point estimates, whereas interval
estimates taking into account the uncertainty of predic-
tion are often required in practice. Constructing confi-
dence intervals using conventional statistical methods
is impossible because of the complexity of the problem
[22], and employing simulation methods is impossible
because of the long computational time required [23].
In 1962, Kantorovich [24] suggested another
approach to the problem of linear calibration, which is
to replace the objection function by a set of inequalities
solvable by linear programming methods. In this case,
the result of prediction appears immediately as an inter-
val estimate. For this reason, this method was named
simple interval calculation (SIC). Previously, this idea
did not gain wide acceptance and was not developed
because of inadequate computer performance. In the
1970s–1990s, this approach was taken in a series of
interesting applied studies [25–32] but was not devel-
oped into a standard method. The results of those inves-
tigations were summed up in a monograph [33], where
the main problem solved by the authors of the above-
mentioned works is considered in detail. This problem
includes the interval estimation of the model parameters
and the immersion of the admitted region of these param-
eters into a hypercube, parallelepiped, ellipsoid, etc.
This problem formulation seems to be unprofitable
and not very promising. This view is proved by the fact
that no new works using this approach have been car-
ried out in the last decade. At the same time, we believe
that Kantorovich’s idea can provide interesting results
when applied to the interval prediction of response. In
Two Approaches to Kinetic Analysis
Applied to the Prediction of Antioxidant Activity
A. L. Pomerantsev and O. Ye. Rodionova
Semenov Institute of Chemical Physics, Russian Academy of Sciences, Moscow, 117977 Russia
e-mail: forecast@chph.ras.ru
Received December 28, 2004
Abstract—Differential scanning calorimetry (DSC) followed by mathematical data processing can be used
instead of the conventional method of long term thermal aging in predicting the activity of antioxidants in poly-
olefins. In this method, a regression relationship is established between the oxidation initial temperatures mea-
sured by DSC (X data) and the oxidation induction period values determined by thermal aging (Y data). Two
approaches, called hard and soft, are employed in the construction of models. In the first case, nonlinear regres-
sion analysis is used in combination with successive Bayesian estimation. The second approach combines par-
tial least squares regression and simple interval calculation. Use of a common data set makes it possible to com-
pare these approaches and to draw inferences as to the cases in which one or the other is preferable.
DOI: 10.1134/S0023158406040094
This article is protected by the copyright law. You may copy and distribute this article for your personal use only. Other uses are only allowed with written permission by the copyright holder.