Evaluation of thermogravimetric measurements using neural networks

The method for evaluation of the results of thermal analysis (TG–DSC) measurements under non-isothermal conditions using neural networks has been described. A requirement of obtaining high-quality neural model was adopted as the criterion. Decomposition of ammonium metavanadate (NH4VO3) in dry air was studied. Empirically determined sets of the results of TG, DTG and DSC measurements and a subset (TG) 3 determined for the third stage of the process were examined. The independent (input) variables were the sample heating rate and the temperature, and the dependent (output) variable was one of the above functions. For the considered data sets, generalized neural networks have been selected. These networks were evaluated based on the criterions taken in the theory of neural networks. High statistical evaluation of the network was received, i.e.,: average error for the set of TG was 0.0397 %, the correlation 0.999957; average error for a subset of (TG) 3 0.118 %, the correlation 0.99995; average error for a set of DTG 0.183 %, the correlation 0.97709; and average error for a set of DSC 0.005 %, the correlation 0.99746. This means, that the sample heating rate and temperature are sufficient to describe the tested functions, the independent and dependent variables were determined with good accuracy, and the sets were numerous sufficiently.