Factor combination analysis

[#1]

Dear all, 

I gathered several sets of historical data containing formulation composition. This includes multiple formulation phases and represents 150 runs with 35 components. Formulation steps were carried out "step by step". From one run to the next, the % of raw materials changes and also the raw materials themselves. In total, this represents 150 runs with 35 different components. FI, each formulation contains approximately 10 components. 

I would like to investigate which combinations of components make the fluid hazy. I created a column appearance (OK/NOK) and I ran predictor screening but this only applies or provides information about a single component, not combinations. When I examine the component(s) that appear to be responsible, I cannot find any clear pattern, and I am sure that it is actually a combination effect.

Is there a tool that can be helpful for me to identify bad combinations? 

Thank you all in advance

[#2]

I'm not 100% certain of how your predictors are combining and the structure of the runs...but from what I gather it sounds like you have a situation where there is a high degree of multicollinearity among your predictor variables. When that's the case perhaps using a modeling method like partial least squares could help? With a categorical response you'll need JMP Pro to run the PLS-Discriminant Analysis modeling platform. The advantage of PLS in general is as part of the math it creates what are called latent structures or variables that can accommodate multicollinearity among the x's...and then the scoring plots and other visualizations in the platform can help you identify the individual x's having the highest impact on the response.

[#3]

Thank you for the response. There is no structure since the DoE methodology was not used.

I tried to run PCA analysis and predictor screening but the results are not fully coherent, according to me. In addition, I can't run PLS because my Y is not a continue variable. 

I attached a anonymized file 

[#4]

You could try to run a logistic regression analysis, since the response is binary.  With high collinearity in the factors, JMP Pro's Generalized regression would probably work better.