# Response Surface (pt 2)¶

## Part 2/3 – Optimization¶

## Introduction¶

This tutorial shows how to use Design-Expert software for optimization experiments. It’s based on the data from Multifactor RSM Tutorial Part 1. You should go back to that tutorial if you’ve not completed it.

For details on optimization, see our online program help. Also, Stat-Ease provides in-depth training in our Response Surface Methods for Process Optimization workshop. Call or visit our web site for information on content and schedules.

In this section, you will work with predictive models for two responses, conversion and activity, as a function of three factors: time, temperature, and catalyst. These models are based on results from a central composite design (CCD) on a chemical reaction.

Use the **Help, Tutorial Data** menu and select **Chemical Conversion
(Analyzed)** from the list.

To see an overview of the design, click the **Summary** node under the
Design branch at the left of your screen. Within the design status screen you can
see we modeled conversion with a quadratic model and activity with a linear model,
as shown below.

Click on the **Coefficients Table** node at the bottom of the **Post Analysis**
branch.

This table provides a convenient comparison of the coefficients for all of the responses.

Note

Because the Coefficients Table is laid in terms of coded factors you can make inferences about the relative effects. For instance, notice that the coefficient for AC (11.375) in the conversion equation is much higher than the coefficients for Factor B (4.04057). This shows, for the region studied, that the AC interaction influences conversion more than Factor B. The coefficients in the table are color-coded by p-value, making it easy to see each term’s significance at a glance. In our example, we chose to use the full quadratic model. Therefore, some less significant terms (shown in black) are retained, even though they are not significant at the 0.10 level.

P.S. Right-click any cell to export this report to PowerPoint or Word for your presentation or report. Check it out: This is very handy!

## Numerical Optimization¶

Design-Expert’s numerical optimization will maximize, minimize, or target:

- A single response
- A single response, subject to upper and/or lower boundaries on other responses
- Combinations of two or more responses.

Under the Optimization branch to the left of the screen, click the **Numerical**
node to start.

### Setting the Optimization Criteria¶

Design-Expert allows you to set criteria for all variables, including factors and propagation of error (POE). (We will detail POE later.) The program restricts factor ranges to factorial levels (plus one to minus one in coded values) — the region for which this experimental design provides the most precise predictions. Response limits default to observed extremes. In this case, you should leave the settings for time, temperature, and catalyst factors alone, but you will need to make some changes to the response criteria.

Now you get to the crucial phase of numerical optimization: assigning “Optimization Parameters.” The program uses five possibilities as a “Goal” to construct desirability indices (di):

- Maximize,
- Minimize,
- Target->,
- In range,
- Equal to -> (factors only).

Desirabilities range from zero to one for any given response. The program
combines individual desirabilities into a single number and then searches for
the greatest overall desirability. A value of one represents the ideal case. A
zero indicates that one or more responses fall outside desirable limits.
Design-Expert uses an optimization method developed by Derringer and Suich,
described by Myers, Montgomery and Anderson-Cook in *Response Surface
Methodology*, 3rd edition, John Wiley and Sons, New York, 2009.

For this tutorial case study, assume you need to increase conversion. Click
**Conversion** and set its **Goal** to **maximize**. As shown below, set **Lower
Limit** to **80** (the lowest acceptable value), and **Upper Limit** to **100**,
the theoretical high.

You must provide both these thresholds so the desirability equation works properly. By default, thresholds will be set at the observed response range, in this case 51 to 97. By increasing the upper end for desirability to 100, we put in a ‘stretch’ for the maximization goal. Otherwise we may come up short of the potential optimum.

Now click the second response, **Activity**. Set its **Goal** to a **target->**
of **63**. Enter **Lower Limits** and **Upper Limits** of **60** and **66**,
respectively. These limits indicate that it is most desirable to achieve the
targeted value of 63, but values in the range of 60-66 are acceptable. Values
outside that range are not acceptable.

The above settings create the following desirability functions:

- Conversion:

- if less than 80%, desirability (di) equals zero
- from 80 to 100%, di ramps up from zero to one
- if over 100%, di equals one

- Activity:

- if less than 60, di equals zero
- from 60 to 63, di ramps up from zero to one
- from 63 to 66, di ramps back down to zero
- if greater than 66, di equals zero

Note

Recall that at your fingertips you’ll find advice for using sophisticated Design-Expert features by pressing the button to see Screen Tips on Numerical Optimization.

### Changing Desirability Weights and the (Relative) Importance of Variables¶

You can select additional parameters called “weights” for each response. Weights give added emphasis to upper or lower bounds or emphasize target values. With a weight of 1, di varies from 0 to 1 in linear fashion. Weights greater than 1 (maximum weight is 10) give more emphasis to goals. Weights less than 1 (minimum weight is 0.1) give less emphasis to goals.

Note

Weights can be quickly changed by clicking and dragging the handles (squares ▫) on desirability ramps. Try pulling the square on the left down and the square on the right up as shown below.

This might reflect a situation where your customer says they want the targeted
value (63), but if it must be missed due to a trade-off necessary for other
specifications, it would be better to err to the high side. Before moving on
from here, reenter the **Lower** and **Upper Weights** at their default values
of **1** and **1**; respectively. This straightens them to their original ‘tent’
shape.

“Importance” is a tool for changing relative priorities to achieve goals you
establish for some or all variables. If you want to emphasize one over the rest,
set its importance higher. Design-Expert offers five levels of importance ranging
from 1 plus (+) to 5 plus (+++++). For this study, leave the **Importance** field
at **+++**, a medium setting. By leaving all importance criteria at their defaults,
no goals are favored over others.

Now click the **Options** button to see what you can control for the numerical
optimization.

Note

Press **Help** to get details on the optimization options.
One that you should experiment with is the Duplicate Solution Filter, which
establishes the “epsilon” (minimum difference) for eliminating essentially
identical solutions. After doing your first search for the optimum, go back to
this Option and slide it one way and the other. Observe what happens to the
solutions presented by Design-Expert. If you move the Filter bar to the right,
you decrease the number. Conversely, moving the bar to the left increases the
solutions.

Click **OK** to close Optimization Options.

### Running the optimization¶

Start the optimization by clicking the **Solutions** tab. It defaults to the Ramps
view so you get a good visual on the best factor settings and the desirability of
the predicted responses.

The program randomly picks a set of conditions from which to start its search for desirable results – your results may differ. Multiple cycles improve the odds of finding multiple local optimums, some of which are higher in desirability than others. Design-Expert then sorts the results from most desirable to least. Due to random starting conditions, your results are likely to be slightly different from those in the report above.

Note

The ramp display combines
individual graphs for easier interpretation. The colored dot on each ramp
reflects the factor setting or response prediction for that solution. The height
of the dot shows how desirable it is. View different solutions from the
**Solutions** drop-down menu on the **Factors tools**; cycle through some of
them and watch the dots. They may move only very slightly from one solution to
the next. However, if you look closely at temperature, you should find two
distinct optimums, the first few near 90 degrees; further down the solution list,
others near 80 degrees.

The Solutions Toolbar provides three views of the same optimization. Click
**Report**.

Now select the **Bar Graph** view from the Solutions Toolbar.

The bar graph shows how well each variable satisfies the criteria: values near one are good.

### Optimization Graphs¶

Select the **Graphs** tab to view a contour graph of the overall desirability and
all of your responses. On the **Factors Tool** palette, right-click **C:Catalyst**.
Make it the **X2 axis**. Temperature then becomes a constant factor at 90 degrees
(this level is picked automatically by the selected solution #1).

You can also use the droplist on the Graphs toolbar to look at bigger versions of each plot, such as the desirability plot below.

The screen shot above is a graph displaying graduated colors – cool blue for lower desirability and warm yellow for higher.

Design-Expert sets a flag at the optimal point (as selected by the
button bar at the top , number 1 is selected in the screenshot above). To view a
response associated with the desirability, select the desired **Response** from
its droplist. Take a look at the **Conversion** plot.

Conversion contour plot (with optimum flagged)

Note

Right-click over this graph and choose **Graph Preferences**.
Then go to **Surface Graphs** and click **Show 2D grid lines**.

Grid lines help locate the optimum, but for a more precise locator right-click
the flag and **Flag Size->Full** to see the coordinates plus many more predicted
outcome details. To get just what you want on the flag, right-click it again and
select **Edit Info**.

To look at the desirability surface in three dimensions, again select
**Desirability** from the Graphs toolbar drop-down. Then, press **3D Surface**.
Next, right-click the graph and select **Set rotation** and change horizontal
control to **170**. Press your Tab key or click the graph. What a
spectacular view!

Now you can see there’s a ridge where desirability can be maintained at a high level over a range of catalyst levels. In other words, the solution is relatively robust to factor C.

Note

Right-click over your graph to bring up Graph preferences. Via the
**Surface Graphs** tab change the **3D graph shading** option to **Wire Frame**.

One way or another, please show your colleagues what Design-Expert does
for pointing out the most desirable process factor combinations. We’d like that!
The best way to show what you’ve accomplished is not on paper, but rather by
demonstrating it on your computer screen or by projecting your output to larger
audiences. In this case, you’d best shift back to the default colors and other
display schemes. Do this by pressing the **Default** button on Surface Graphs
and any other Graph Preference screens you experimented on.

P.S. Design-Expert offers a very high Graph resolution option. Try this if you like, but you may find that the processing time taken to render this, particularly while rotating the 3D graph, can be a bit bothersome. This, of course, depends on the speed of your computer and the graphics card capability.

Now move on to graphical optimization—this may the best way to convey the outcome of an RSM experiment by displaying the “sweet spot” for process optimization.

## Graphical Optimization¶

When you generated numerical optimization, you found an area of satisfactory
solutions at a temperature of 90 degrees. To see a broader operating window,
click the **Graphical** node. The requirements are essentially the same as in
numerical optimization:

- 80 < Conversion
- 60 < Activity < 66

For the first response – **Conversion** (if not already entered), type in **80**
for the **Lower Limit**. You need not enter a high limit for graphical
optimization to function properly.

Click **Activity** response. If not already entered, type in **60** for the
**Lower Limit** and **66** for the **Upper Limit**.

Now click the **Graphs** tab to produce the “overlay” plot. Notice that regions
not meeting your specifications are shaded out, leaving (hopefully!) an operating
window or “sweet spot.” Now go to the **Factors Tool** and right-click
**C:Catalyst**. Make it the **X2 axis**. Temperature then becomes a constant
factor at 90 degrees as before for Solution 1.

Notice the flag remains planted at the optimum. That’s handy! This Design-Expert display may not look as fancy as 3D desirability, but it can be very useful to show windows of operability where requirements simultaneously meet critical properties. Shaded areas on the graphical optimization plot do not meet the selection criteria. The yellow “window” shows where you can set factors that satisfy requirements for both responses.

Note

Go back to the **Criteria** and click **Use Interval
(one-sided)** for both **Conversion** and **Activity**. This provides a measure
of uncertainty on the boundaries predicted by the models — a buffer of sorts.

After looking at this, go back and turn off the intervals to re-set the graph to the default settings.

P.S. If you are subject to FDA regulation and participate in their quality by design (QBD) initiative, the CI-bounded window can be considered to be a functional design space; that is, a safe operating region for any particular unit operations. However, to establish a manufacturing design space one must impose tolerance intervals. This tutorial experiment provided too few runs to support imposition of TIs. To size designs properly for manufacturing QBD requires advanced know-how taught by Stat-Ease in the Designed Experiments for Pharma workshop.

Let’s say someone wonders whether the 80 minimum for Conversion can be increased. What will this do to the operating window? Find out by dragging the 80 conversion contour until it reaches a value near 90. Then right-click it and Set contour value to 90 on the nose.

It appears that the more ambitious goal of 90 percent conversion is feasible. This requirement change would make the lower activity specification superfluous as evidenced by it no longer being a limiting level; that is, not a boundary condition on the operating window.

Graphical optimization works great for two factors, but as factors increase, optimization becomes more and more tedious. You will find solutions much more quickly by using the numerical optimization feature. Then return to the graphical optimization and produce outputs at the various optimization solutions for exploration of the processing windows.

## Response Prediction and Confirmation¶

This feature in Design-Expert allows you to generate predicted
response(s) for any set of factors. To see how this works, click on the **Point
Prediction** node (lower left on your screen). Notice it now defaults to the
first solution. (Be thankful Design-Expert programmers thought of this, because
it saves you the trouble of dialing it up on the Factors Tool.)

### Confirmation¶

After finding the optimum settings based on your RSM models, the next step is to
confirm that they actually work. To do this, click the **Confirmation** node (left
side of your screen).

Look at the 95% prediction interval (“PI low” to “PI high”) for the Activity response. This tells you what to expect for an individual (n = 1) confirmation test on this product attribute. You might be surprised at the level of variability, but it will help you manage expectations. (Note: block effects, in this case day-by-day, cannot be accounted for in the prediction.)

Of course you would not convince many people by doing only one confirmation run.
Doing several would be better. For example, let’s say that the experimenters do
three confirmatory tests. Click the **Confirmation Location #1** tab and next to
**Runs** enter **3** then press tab. You will now have 3 rows under **Response
Data**, for **Activity** enter **62**, **63** and **64**.

Notice that the prediction interval (PI) narrows as n increases. Does the Data Mean (63) fall within this range? If so, the model is confirmed. If not, it will turn red. Isn’t DX helpful?

Note

Keep increasing the value for n. Observe the diminishing returns in terms of the precision, that is, the PI approaches a limit — the confidence interval (CI) that you saw in Point Prediction. The CI is a function of the number of experimental runs from which the model is derived. That means one can only go so far with the number of confirmation runs. 4 to 5 is a good recommendation.

### Save the Data to a File¶

Now that you’ve invested all this time into setting up the optimization for this
design, it would be prudent to save your work. Click the **File** menu item and
select **Save As**. You can now specify the **File name** (we suggest tut-RSM-opt)
for Save as type “*.dxpx” in the Data folder for Design-Expert (or wherever you
want to Save in).

If you are not worn out yet, you will need this file in Part 3 of this series of tutorials.

## Summary¶

Numerical optimization becomes essential when you investigate many factors with many responses. It provides powerful insights when combined with graphical analysis. However, subject-matter knowledge is essential to success. For example, a naive user may define impossible optimization criteria that results in zero desirability everywhere! To avoid this, try setting broad acceptable ranges. Narrow them down as you gain knowledge about how changing factor levels affect the responses. Often, you will need to make more than one pass to find the “best” factor levels that satisfy constraints on several responses simultaneously.

This tutorial completes the basic introduction to doing RSM with Design-Expert. Move on to the next tutorial on advanced topics for more detailing of what the software can do. If you want to learn more about response surface methods (not the software per se), attend our Stat-Ease workshop Response Surface Methods for Process Optimization.