This presentation is adapted from a JMP Discovery Europe talk given by Silvio Miccio in 2017. The video begins with a description of the steps I will follow in the demo. Following is some background to the video:
Some Quotes
- Deming – “Statistics is to important to be left to the statisticians” (Analytics)
- Box – “All models are wrong, but how wrong do they have to be to not be useful?”
- Author Unknown – “Don’t fall in love with your models. They will NOT love you back” – Things change!
- Carly Fiorina – “The goal is to transform data into information and information into insight.”
- John Sall (the creator of JMP) - "Data rarely confesses all its information easily. You have to explore it. You have to use multiple tools. The process should be as easy as possible and the tools as powerful as possible.”
Introduction
When testing options of different raw materials or formulation ingredients, a common practice is to vary them as multilevel categorical variable e.g. A, B, C….etc. In this scenario an experiment for identifying the best candidate material all options have to be tested. A consequence of this is that time consuming physical testing is required and the resulting model is only applicable to predict the tested options but cannot predict options with different physical/chemical properties.
A much more efficient approach is to design the experiment based on the physical/chemical properties of each new material option.This can significantly decrease the number of required experimental conditions and results in a sustainable empirical model that can predict options not tested before.
Example Data Set Used in this Presentation
- 45 raw material options, which are a potential improvement of one ingredient of the product formulation
- Material options can be described by 19 physical and chemical descriptors (one categorical two level factor was coded as -1/1)
- Two responses to be optimized (Y1 max, Y2 min)
Typical Test Procedure
- Test (each) raw material option in base formulation
Objective
- Predict optimum material (even options that may not be physically available yet)
- Identify available material options closest to overall optimum
- Establish a sustainable empirical model that is also capable of predicting new raw material options
- This approach is based on a chemometric application known as QSAR (Quantitative Structure-Activity Relationship)
Physical and Chemical Properties Examples
- Mw – Molecular Weight
- C – the number of carbon atoms on the hydrophobic end of the surfactant
- Griffin HLB – the hydrophylic-lypophilic balance (HLB) according to Griffin
- CMC – critical micellar concentration
- Number of chains……
Summary
- Compress the available information regarding physical and chemical properties of different options via principal components of ExperimentsSelect the “corners of the box” for testing representative raw materials based on Design
- Select the “corners of the box” for testing representative raw materials based on Design of Experiments
- Model the data via PLS (or other appropriate modeling methods)
- Find the overall optimum solution
- Identify physical available options closest to the calculated optimum solution
→ Highly efficient experimentation
→ Sustainable empirical model based on physical/chemical properties
→ Predicts optimum raw material properties although material identified as optimum
Table with ChemicalBiological Properties (non P&G Example)
