Optimal constants in physicochemical models describing decomposition kinetics are searched for from thermogravimetric information by the method of nonlinear regression. The method requires an iterative numerical solution of differential equations for non-isothermal kinetics. The technique of conjugate functions is used to provide a fast quadratic convergence of the iterations. The method is described in detail for the cases where input information represents either derivatives of mass loss or perturbations of these derivatives caused by sinusoidal modulation of linear temperature-time relationships. In both cases two-stage decomposition with two parallel (independent) reactions is considered as a numerical example. The discussion is mainly focused on the analysis of the perturbations (modulations). In the case of modulated thermogravimetry, a relative error in the calculated activation energy is approximately equal to a relative error in the approximation of the perturbations. That is why activation energy can be calculated with high accuracy. Moreover, the conclusion about the adequacy of evaluated models can be made immediately after a visual verification of the approximation error. Two-stage decomposition of cotton is used as an experimental example. The only curve of the perturbations measured at one heating rate is used for calculating the kinetic constants. This curve contains information about mass loss in the hidden form. Nevertheless, one can ensure that the used mathematical models with the obtained kinetic constants are able to approximate not only the curve of mass loss subjected to the handling but also a curve measured at another heating rate. The high predictive force of modulated thermal analysis promises to be an obvious alternative to the classical technique. The weak sensitivity of predictions made by modulated thermogravimetry to a form of models used is a very important distinction for the wide application of this analysis in different fields of chemical engineering. (C) 2004 Elsevier Ltd. All rights reserved.