Figure 8 Regression plots for the estimated and experimental pKa values for 71 data points
The developed ANN model did an excellent job in the prediction of dissociation constants, however, the method id limited by the large number of properties to be used as input parameters limitation to be found in the literature or measured experimentally. The ANN model can be more useful and flexible if it was possible to use fewer input data. In this study, the ANN model was also tested by reducing the number of input data. The molecular weight, critical temperature and pressure were maintained as inputs in the new model as they identified the studied compound, in addition to the temperature as the dissociation constants values were temperature-dependent. As for the four physical parameters left, the ANN model was tested by removing them one by one, then two and three at the time. The performance of the ANN models based on Roverall, MSEtrain, MSEval and MSEtest were summarized in Table 13. It was found that the performance of the model deteriorated slightly by removing some input information, compared to the original model. However, the ANN model while leaving only the surface tension and refractive index data had a very good performance and is now called the new ANN model. Although the new ANN model had a slightly lower performance than the full model, the estimated pKa values were still in good agreement with the experimental values but the model allowed for more flexibility in its application.
As observed in Figure 9, the best validation performance was reached at epoch 76 and became unchanged with the increase in the number of epochs. Figure 9 shows the regression between the estimated and experimental values of the new ANN model. Regression plots for the estimated and experimental pKa values for 71 data points for the new ANN model are presented in Figure 10.
Figure 9 The performance improvement of the new ANN model