The results of the sensitivity tests are shown in graphs using the same vertical axis units so that the relative influence of each process parameter can be more easily seen. The results show that pH and Temperature variables have the largest influence on process yield, followed by conductivity and redox potential, respectively. The Neural network model shows that dissolved O2 and Agitator Power have no influence upon yield. 
All of these results are expected from the mathematical relationships originally used to create the simulated process data. These results validate that the Neural network model was able to assess the impact of each process variable correctly. 
The conclusion from these sensitivity tests is that the pH and Temperature parameters are the two process variables which would provide the biggest benefit from tighter process control. 
Determining the window of opportunity 
A test file was created in which the pH and Temperature variables were systematically varied in very small increments over the complete ranges seen in the training data. This file was then processed through the trained Neural network model to create a Yield prediction matrix. 
This matrix can then be analyzed using NumPy tools to see which combinations of process variables lead to a predicted Yield greater than a specified value. The opportunity windows, which lead to greater than 93% and 94% yield respectively, are shown below.

The area of the blue dots indicates the process parameters which would have to be achieved to obtain the specified yield. As expected, the opportunity window for greater yields is narrower than for lower yield levels.  
The question for the process engineer becomes: What is the cost of having a process control system which could consistently hit the window of opportunity, and what is the economic benefit of consistently achieving the higher process yield? 

copyright 2019 Powell Simulation LLC