Journal of Thermal Analysis, Vol. 49 (1997) 1553-1564 NEURAL NETWORKS, FOR KINETIC PARAMETERS DETERMINATION, SIGNAL FILTERING AND DECONVOLUTION IN THERMAL ANALYSIS N. SbirrazzuolL D. Brunel and L. Eldgant Laboratory of Experimental Thermodynamics U.M.R.-C. N. R. S- 139, University of Nice-Sophia Antipolis, 06108 Nice Cedex 2, France Abstract Feedforward neural networks have been used for kinetic parameters determination and signal filtering in differential scanning calorimetry. The proper learning function was chosen and the network topology was optimized, using an empiric procedure. The learning process was achieved using simulated thermoanalytical curves. The resilient-propagation algorithm have led to the best minimization of the error computed over all the patterns. Relative errors on the thermodynamic and kinetic parameters were evaluated and compared to those obtained with the usual thermal analysis methods (single scan methods). The errors are much lower, especially in presence of noisy signals. Then, our program was adapted to simulate thermal effects with known thermody- namic and kinetic parameters, generated electrically, using a PC computer and an electronic in- terface on the serial port. These thermal effects have been generated by using an inconel thread. Keywords: deconvolution, differential scanning calorimetry, feedforward neural networks, ki- netics, signal filtering, simulations, thermal analysis Introduction Artificial neural networks is a computing approach based on an analogy with the working of the nervous system of the brain, in which connections organize the units (neurons) into networks. One of the main advantages of the artificial neural net- works is that they are massively parallel and can improve their performances through examples by a dynamic learning process. Therefore, the use of artificial neural networks allows a modelling of complex systems by way of interconnection weights, without requiring the explicit formulation of the relationships that may ex- ist between variables. These weights are generated by the training that starts from random values of the coefficients (bias). It has been established that a standard lay- ered feedforward network architecture can approximate any function of interest, provided that a sufficient number of hidden neurons are used [1, 2]. Furthermore, artificial neural networks are highly tolerant to noisy or missing data in the training or test set samples, and suitable for pattern recognition, reconstruction applica- tions, filtering or time-series predictions. One of the greatest developments of these last years is an attempt to extracting kinetic information of a transformation by thermal analysis [3]. An application of 0368-4466197/$ 5.00 9 1997 Akad~miai Kiad6, Budapest John Wiley & Sons Limited Chichester