Hello JMP community,
I am a new JMP user and had a question that I thought I would post to see if there are any tips for making my experiment more effective.
Basically I am running a DoE experiment involving a mixture of colors (red yellow and blue). I have a combination of colors that I start with and measure using a spectrophotometer. (For example: I start with a solution that is 1ml. It consists of 30% blue dye, 25% red dye, and 35% yellow dye. I then take a reading of this color in a spectrophotometer and get a wavelength that is unique to this mixture of colors)
I then use custom DoE design in JMP to create a table of combinations using mixture factors red, yellow, and blue (all with values of 0 to 1) with the goal set to minimize. After, I create those combinations and run them all through the spectrophotometer and see which one is closest to the mixture I am trying to find. (i.e take the reading of the resulting mixture and compare it to the reading of the mixture I am trying to re-create. If the reading of the original mixture reads 0.854, and a reading from a created mixture reads 0.754, the data I enter would be 0.1, with the lowest difference being the most desirable.)
I would like to repeat this process getting closer to the original color, seeing how many runs it may take me to get as close as possible. At the moment I will enter in the data I get from each run, look at the 'Fit Model' function and look at the 'maximum desirability' in the Prediction Profiler. I then will use Augment design and change the values of each factor to reflect the maximum desirability output. (i.e if the maximum desirability prediction indicates the combination with the lowest difference as 0.3 red, 0.3 blue, and 0.4 yellow, I would change the values of red and blue to be between 0.2 and 0.5, and yellow to be between 0.3 and 0.5) I then re run the experiment, repeating until I get a combination that is ~ equal to the desired combination.
I am just wondering if there might be a more efficient way of doing this, if there is a way I can use JMP to automatically take the existing data I have found and use that to automatically pick a range of combinations that are more likely to find the right combination, or if the way I am doing it currently is the most efficient.
thanks for any advice/tips
You said, "I would like to repeat this process getting closer to the original color, seeing how many runs it may take me to get as close as possible" and "if there is a way I can use JMP to automatically take the existing data I have found and use that to automatically pick a range of combinations that are more likely to find the right combination." These statements concern me. It reveals a 'testing mindset.' You are thinking of the solution to be found by 'trial and error.' The 'experimenting mindset' is different. It is about collecting the best (effective and efficient) data to fit a model that can predict the response over a wide range of factor levels. You don't want to use narrow factor ranges even if they contain the ultimate best settings. Your parameter estimates and test statistics will exhibit high variance and result in unstable predictions.
How did you select the wavelength to measure the absorbance? So any combination (likely more than one!) that produces the desired absorbance is acceptable?
You can use the Match Target goal and define an acceptance window with the Low and High values. Then supply the absorbance for Y, not the difference.
Do you expect non-linear changes in the absorbance as you proceed from a pure blend (only one color) to a binary or tertiary blend? If so, the model should include quadratic effects at least.
I attached this design as a JMP data table:
One more thing to add to @markbailey's comments: the testing approach that you are currently using does not seem to take into account variability. You can test the same blend multiple times and not get the same result. That does not seem to be taken into account, whereas the model-building approach that Mark describes has that built into the results. You will get a prediction along with an error to indicate the probable values with a certain confidence.
Thanks for the swift response.
Yes I agree this is more of a testing mindset, but I do not know any other way to effectively test 3 combinations in order to try to find a combination that matches a target goal.
I use 3 wavelengths (one for blue, one for yellow, and one for red). I take these readings of the desired mixture. Then I take these readings for each of the created mixtures. I find the difference between each of the 3 readings, taking the absolute value if it is negative. I then add the 3 differences between the 3 wavelengths together and that is the final number.
For example if my desired mixture is: Wavelength 1: 0.4. Wavelength 2: 0.35. Wavelength 3: 0.5. The created mixture gives: Wavelength 1: 0.3. Wavelength 2: 0.3. Wavelength 3: 0.4. The result would be (0.4 - 0.3) + (0.35- 0.3) + (0.5 - 0.4) and the final number would be 0.25. Because of this I
Right now I am using the Scheffe Cubic term to create the model.
You can have three responses each with the Match Target goal but unique values. Then the prediction profiler will use the desirability functions to find the mixture that simultaneously most closely matches those targets.
There are no labels assigned to this post.