Planning and analyzing a 2-level full factorial design in Python

[@Imran_et_al]

Great Explanation. Thanks for the video and code...!

[@MarcelButschle]

Your welcome!

[@PierriDellaBruna]

Great video! Really valuable material. Congratulations :)

[@MarcelButschle]

Thank you!

[@fightchris7359]

Nice, explained very well. I like the approach to let ChatGPT write most of the code. Keep going!

[@MarcelButschle]

Thank you fightchris! Highly appreciated. Yes, ChatGPT makes it extremely easy to write code. Saves a lot of time and it makes it very easy to get started also when you are less experienced.

[@sebastianoverri9482]

Love your work. Thanks, your teaching is very clear even for a newby inthe subject like me.

[@MarcelButschle]

Thank you! Highly appreciated feedback :)

[@christophersaid8985]

This is great, thanks for posting. I work at a startup that's extremely cheap, so having DOE capabilities in python is very cool. Would you be able to share the actual Jupyter Notebook? I couldn't find it on your site.

[@MarcelButschle]

Hey, sure! If you send me an email I can send you the actual Jupyter Notebook file. [e-mail]

[@victorrossing6484]

I have a problem when doing DOE in python. Lets say we have a 2^2 experiment single replicate with 5 centerpoints, then we have 4 degrees of freedom. One each of x1, x2, x1x2, whereby we only have one combined for x1^2 and x2^2, hence we need a pure quadratic term for these two combined, as they cant be separated. However OLS cant give me the pure quadratic term (just gives me an estimate of x1^2, which is equal to the estimate of x2^2), do you have a solution or another package you know that works? Code attached (chatGPT cant find a solution): test=pd.DataFrame({}) test["x1"]=[-1,-1,1,1,0,0,0,0,0] test["x2"]=[-1,1, -1,1, 0, 0, 0, 0,0] test["response"]=[39.3,40,40.9,41.5,40.3,40.5,40.7,40.2,40.6] model_poly = ols('Q("response") ~ x1 + x2 + I(x1 * x2) + I(x1**2 +x2**2)', data=test).fit() print(model_poly.summary())

[@MarcelButschle]

Hey, if you perform a 2^2 design with 5 centerpoints then you have a total of 4+5=9 runs which results in 9-1=8 degrees of freedom. Hence, you should be fine with estimating the squared terms individually, or am I completely off?

[@BarisDuygu-q2u]

Hey Marcel, Ive got the issue, that the results I have had some slip region included. Now of course when I remove these points, my cubic won't be complete. I tried to make a model based on the remaining runs, Used anova for this.

[@MarcelButschle]

Hey Baris, then you are dealing with an unbalanced design. That is not necessarily a problem and ANOVA might still work.

[@BarisDuygu-q2u]

Should I remove these points? As my model should actually be able to have an error less then 1 mm. If I include this points, the model seem to have high error.

[@MarcelButschle]

Well that depends… how many points do you have in total? And how many replicates did you include into your design? Would it make sense to measure these points that you want to exclude again? Is your prediction still valid when you exclude the points? What happens if you try to predict the results for new parameters that the model was not trained on?

[@rusty-y8r]

pydoe2 is not not importing well in anaconda python.. it says imp is needed but i cannot install imp in newer version of python. Can someone share what python version works for pydoe2?

[@MarcelButschle]

Hey, I am using python version 3.11.7 and it works fine.

[@danread1217]

I had this recently. Try rolling back your Python version to 3.9.19. Seems pyDOE2 hasn't been updated for some time.

[@danread1217]

Also, try pyDOE3. It is a recent fork of pyDOE2 and is currently being maintained.

[@mohammedbenali7546]

Thank you for this video. Please, if possible how we can do doehlert design in python, Video the scientific community please.

[@MarcelButschle]

Hi Mohammed, my approach is to keep it simple. I am aware that there are many different designs that can be very useful to some people but 90% of people will be fine with the combination of fractional design, full factorial design and central composite design. Therefore I am not planning to look into these „specialty designs“.