Optimization of artificial neural networks for prediction of the unit cell parameters in orthorhombic perovskites. Comparison with multiple linear regression

The unit cell parameters (a, b, c) of orthorhombic perovskites (of A2+B4+O3 and A3+B3+O3 type) were predicted both using multiple linear regression analysis (MLR) and two types of artificial neural networks (ANN). In these analyses, 70 compounds of above perovskite type were included: 47 in calibration set and 23 in test set, which were randomly chosen. In multiple linear regression, the unit cell parameters of 47 perovskites were expressed as bilinear function of the effective ionic radii of A and B cations, and then, using the obtained regression equation, the unit cell parameters of 23 perovskites were calculated and compared with the experimental data. Predictions using the same sets and the same dependent and independent variables were also done by feed-forward and cascade-forward ANN. The two different ANN models were compared to MLR model by F-test using their root mean square error (RMSEP). Although the two models give excellent results, it could be concluded that ANN have significantly better prediction abilities compared to MLR.