Modelling of Liquidus Temperature and Electrical Conductivity of Manganese Smelting Slags by the Use of Neural Nets

"Liquidus temperature and electrical conductivity data measured on synthetic slags were modelled by the use of neural nets (NN). In this work the applied multilayer feedforward NNs were trained by a conjugate-gradient optimisation for which a three layer formulation was used. Very good fits were obtained for both the liquidus temperature and the conductivity data. The synthetic slags were prepared from pure oxides to represent a wide range of compositions likely to be encountered in ferromanganese and silicomanganese smelting. The slag constituents were in the following range: Mn0;5-30%, Ca0;20-35%, Mg0;5-15%, Si02;27-58%, Al20 3;5%. Liquidus temperatures varied from 1300° to 1380° and increased with increasing basicity ratio. The electrical resistivity of slags decreased with the increase of basicity ratio from 0.55 to 1.1 but above 1.1 basicity ratio the resistivity tended to increase depending upon the MnO content.Neural NetsVarious neural net (NN) topologies such as Hopfied nets, Hamming nets, Carpenter/Grossberg classifiers have been discussed in some detail by Lipmann (1). In this work, the applied multilayer feedforward NNs are not trained by backpropagation (BP) using the generalized delta rule (2) (GDR), but by a conjugate-gradient optimisation [3,4] - a procedure similar to that proposed by Barnard and Cole [5]. A neural net is capable ·of learning the non-linear functional relationship existing between a set of inputs and outputs [ 6]. In this capacity a neural net is used to determine the functional relationship that exists . between N dependent and independent variables, which characterize the system under consideration For a three layer neural net [1] the following function may be formulated, which relates X; inputs to the output Y 0 from node o:"