Browsing by Author "Mongalo, Lehlohonolo"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemOpen AccessMolecular dynamics simulations of the electrical conductivities of high temperature metallurgical slags(2018) Mongalo, Lehlohonolo; Venter, Gerhard A; Lopis, Anton SThe structural properties and electrical conductivities of nine CaO-MgO-Al₂O₃-SiO₂ slags with compositions in the peralkaline region have been calculated using molecular dynamics simulations. Where applicable, results were compared to those estimated from composition data alone. The proportion of nonbridging oxygen (NBO) and bridging oxygen (BO) atoms were determined from simulation and shown to be in reasonable agreement with theoretical prediction, with the number of NBOs increasing as the number of network modifying cations increase. Bridging oxygen atoms were classified into Si-O-Si, Si-O-Al or Al-O-Al linkages and the results used to establish whether the Al avoidance principle is applicable. Consistent with experimental and simulation results reported elsewhere for aluminosilicates, a surprisingly large number of fivefold–coordinated Al atoms were found, even though the simulated compositions contain CaO and MgO far in excess of the tectosilicate join. The number of bridging oxygens coordinated to tetrahedral [SiO₄]⁴⁻ and [AlO₄]⁵⁻ units, namely the Qⁿ distribution, was determined. Although a good comparison to the theoretical average Q was found at low basicity, at higher basicity greater deviation was seen. Finally, electrical conductivities calculated using the Einstein relation, taking cross–correlations into account, were in excellent agreement with experimentally measured values, although Nernst–Einstein conductivities, estimated from self–diffusion coefficients alone, showed large deviations. In doing these calculations, it is implicitly assumed that the total electrical conductivity of the slags results from motion of the ions alone and that there is no electronic component to the conductivity. Results therefore show that molecular dynamics simulations are able to reliably predict conductivity, but values calculated indirectly, i.e. using the Nernst–Einstein relationship, should be used with care. At low basicity theoretical predictions of structural disorder, based on Zachariasen's Random Network Model, are in reasonable agreement with simulation, but this agreement worsens as the proportion of network modifying cations increases. Artificial neural network (ANN) models for predicting electrical conductivities of slags, based on structural properties, were also successfully developed. Two layer MLP feedforward ANN models, using the resilient back propagation algorithm for training, were used to predict conductivities. The input and output parameters were calculated using MD simulations and different combinations of input parameters, as well as number of hidden neurons, were used to find the best model. The best models were identified based on having low MSE errors, when applied on a test data set for which experimental results are known. Using a subset of structural parameters (average Q⁰, Q¹ distributions and the number of NBO atoms) yielded the best model with an MSE of 6.8. More general models using a greater set of structural parameters had MSEs in the range of 33.0 to 35.2. The artificial neural network models have demonstrated a reasonable agreement in predicting the MD calculated electrical conductivities of slags and hence, proved to be effective methods for the prediction of electrical conductivities of slags using structural properties as input parameters.