3.5 | ANN results
The data set consisting of 96 data were used for artificial neural network modeling. Of this set, 68 were allocated for training, 14 for testing and 14 for control data. In order for the network not to memorize, learning should take place properly, the data set should be blended well, and the training test and control data should be selected homogeneously. For this purpose, a series of experiments were performed with different initial condition, and the situation where all of the R2 were high was investigated. There are 6 input variables and 1 output variable in the general structure of the network. The value of hidden layer size was tried to be determined by preliminary experiments. Conversion was selected as output variable. Network structure with 2 outputs (conversion and selectivity) and single output (only selectivity) were also studied, but positive results could not be obtained from here. It has already been mentioned in the previous chapters that a significant correlation between the selectivity value and the feeding modes could not be seen. Here 6 input variables are; phenol/acetone ratio, co-catalyst usage state (0/1), feed time, waiting time, acid concentration and reaction time. The regression graphs obtained by plotting the conversion values in experimental data against the results of the network are given in Figure S4.
As can be seen, the network performed well for both training, test and control data, 0.99 R2 values were obtained for all. High training R2 is important because low means the network is not well trained. This value indicates the educational performance of the network, but only if this value is high this does not make sense. Higher R2 of test and control data are also desirable, otherwise, the network have memorized and cannot predict when new situations occur. After this point, 3-D graphics were drawn in MATLAB using the obtained neural network model. Conversion against time /HCl (%), conversion against time /mol ratio graphs are shown in Figure 5.