Characterizing Bio-processes With Augmented Full Quadratic Models and Fractionally Weighted Bootstrapping

[#1]

In this session, Phil Ramsey (Senior Data Scientist and Statistical Consultant at Predictum Inc. and Owner of the North Haven Group) uses bio-process development case studies to demonstrate that the augmented FQM models substantially outperform the traditional FQM in characterizing the full design space. After explaining why we need Design of Experiments (DoE) and predictive models for quality by design (QbD), Dr. Ramsey uses JMP Pro to demonstrate the application of augmented FQMs, FWB, and the Simulators in the JMP Profilers to fully characterize design spaces as required by QbD.

[ahmedmohamed]

The journal, please

[PatrickGiuliano]

Hello @philramsey!  Thank you for this wonderful and well thought-out presentation.  Would it be possible for you to provide the accompanying slides (or journal)? As always in your "style" they are in-and-of-themselves, an excellent reference.  

• You mentioned the underappreciated paper reference by Cornell an Montgomery (1996).  Can you please confirm the publication reference based on my internet search? 
  • John A. Cornell & Douglas C. Montgomery (1996) Interaction Models as Alternatives to Low-Order Polynomials, Journal of Quality Technology, 28:2, 163-176, DOI: 10.1080/00224065.1996.11979657
• What does it mean for the Autovalidation method (aka S-VEM) to be an Omnibus procedure? 
  • In the context of your discussion around 11-min where you describe how in this method you are assigning gamma weights to the training and validation set (copy of training set) such that the weights are essentially anticorrelated. 

For the JMP User Community at large, I link to the BrightTalk that first got me started down this road - given by you, @chris_gotwalt1 and @wjlevin  on Oct 21, 2020. 

@PatrickGiuliano 

[philramsey]

Hello Patrick,

You have the correct reference for the Cornell and Montgomery paper. It is truly underappreciated and should have had a better reception when it was first published.  The full quadratic model is indeed often not sufficient to describe the behavior or complex response surface and the reason lack of fit is so common in analyses of response surface designs.

By omnibus I mean a very general algorithm that can be used with almost any predictive modeling approach. SVEM is really agnostic in terms of the original data source and for the most part the predictive modeling approach.

Hope this  helps,

Phil