Optimize both Yield and Cost with Minitab Design of Experiments
Design of Experiments is a powerful technique to optimize process outputs in complex processes where multiple factors affect the outputs. When we want to optimize more than one output the problem becomes more complex.
With this example we want to follow the steps, using Minitab, to optimize both process yield and cost by acting on 3 critical factors which have been identified by previous experiments.
These are the factors and levels we want to experiment with:
We will start with a full factorial experiment with 2 central points:
The resulting design is:
We are going to use a simulator in Excel to perform the experiments which you can download: Yield & Cost Simulator
We will just copy columns C5 - C7 from Minitab to the corresponding factor columns in the simulator:
We now copy the results of the experiments (columns Yield and Cost) back to the Minitab sheet in C8 - C9:
Now we start to analyze these results of our factorial experiment with Minitab:
These are the results for Yield:
This is telling us the factorial model is not valid: R-sq(adj) is very low (9,55%). This is the proportion of variability explained by this factorial mathematical model.
If we look at the p-values none of the factors or their interactions are significant.
We suspect curvature in spite of its p-value being slightly above 0,05.
To confirm curvature we can look at the graphs of main effects both for Yield and Cost:
This, clearly, shows curvature both for Yield and Cost because the Center points are totally out of line with the Corner points.
To further confirm curvature we repeat the factorial analysis including central points:
By including the central points in the model R-sq(adj) has improved a lot (97,97%) so this confirms we need quadratic terms in the formula but these can not be calculated with a factorial experiment. We need a Response Surface experiment to model this process.
Response Surface Experiment
We design a Response Surface experiment with the default values for 3 factors
But be aware that we must define our factor values as Axial points to avoid physically impossible values such as a negative pH.
The resulting response surface design is:
To run the experiments we copy columns C5-C7 to our simulator:
And back the results of the experiments (Yield and Cost) to the new Minitab worksheet:
We now perform the analysis of the Response Surface Design:
With the following results for Yield:
R-sq(adj) value of 96,85% confirms the validity of this model. Also all the p-values below 0.05. Only the Temperature*Time interaction and Temperature^2 are not significant. Before we remove these from the model we must make sure they do not significantly affect the other output variable we try to optimize:
Now we confirm that both Temperature^2 and Temperature*Time do not,significantly, affect Yield or Cost.
Therefore we will remove both from the model:
The resulting reduced model for Yield is then:
This is the formula to estimate Yield as a function of Temperature, Time and pH.
The reduced model for Cost is:
We can now see graphically the impact of the 3 factors in Yield with Contour Plots:
And also for Cost:
We can, graphically, see the optimal areas to maximize Yield and to minimize Cost. Unfortunately these areas do not coincide.
Now we will let Minitab calculate the optimal factor values to achieve our overall optimization:
We have decided to assign double weight to Yield Vs Cost. The result is:
This gives us the optimal factor settings to obtain a predicted Yield of 156 and a Cost of 97.
The following interactive graph lets us touch up the factors and see the results of both outputs:
The optimal factor values have been obtained by maximizing the Composite Desirability curves which are based on our condition that Yield is twice as important as Cost.
If we wanted to maximize Yield at any Cost we would obtain:
This improves Yield but it also increases cost.
With this example you can experience the power of Minitab to design and analyze experiments in order to optimize several outputs by calculating the optimal factor settings