The error figures seen in the validation test are reasonable considering that the original simulated production data set used to train the neural network contained a component of random noise. The maximum value of the noise component was equivalent to 4% of the span of yield values in the training set. 
Conducting sensitivity tests 
The trained neural network model was then used to see which process variables have the greatest impact through sensitivity tests. These tests were done by submitting specifically designed data sets to the model. In each test, all the process variables were held constant using the values which corresponded to the highest yield in the training data set. Only one variable at a time was allowed to vary over its range of values seen within the training data. 
The results of these sensitivity tests for each process parameter can be seen in the graphs which follow. 

copyright 2019 Powell Simulation LLC