Martin Here. recently I developed one adhesive with JMP DOE. It is a 3 factors, 2 responses, 12 run, custom design. See details in below Excel table. Adhesion in water is very difficult to pass, it needs to be above 1 MPa. I am quite excited to find #12 gives adhesion in water 1.09 MPa. FactorsResponse OPCTernalanhydrite6 hrs adhesionadhesion in water#10.740.2600.551.00#20.730.210.060.121.00#30.790.210.000.5860.80#40.710.230.060.201.00#50.760.240.000.601.03#60.650.230.120.180.98#70.630.260.110.311.02#80.630.260.110.241.01#90.670.210.120.130.85#100.710.230.060.170.67#110.790.210.000.370.93#120.690.260.050.301.09 When I use profiler to optimize, I am surprised to find, JMP profiler recommend anhydrite to be set at 0%. I believe anhydrite at 0.05% is critical for #12 to reach such a good adhesion. Is JMP profiler could miss important information? Could anyone help explain? Thanks, Martin

nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

JMP profiler

MartinY

Martin Here. recently I developed one adhesive with JMP DOE. It is a 3 factors, 2 responses, 12 run, custom design.

See details in below Excel table.

Adhesion in water is very difficult to pass, it needs to be above 1 MPa. I am quite excited to find #12 gives adhesion in water 1.09 MPa.

	Factors	Response
	OPC	Ternal	anhydrite	6 hrs adhesion	adhesion in water
#1	0.74	0.26	0	0.55	1.00
#2	0.73	0.21	0.06	0.12	1.00
#3	0.79	0.21	0.00	0.586	0.80
#4	0.71	0.23	0.06	0.20	1.00
#5	0.76	0.24	0.00	0.60	1.03
#6	0.65	0.23	0.12	0.18	0.98
#7	0.63	0.26	0.11	0.31	1.02
#8	0.63	0.26	0.11	0.24	1.01
#9	0.67	0.21	0.12	0.13	0.85
#10	0.71	0.23	0.06	0.17	0.67
#11	0.79	0.21	0.00	0.37	0.93
#12	0.69	0.26	0.05	0.30	1.09

When I use profiler to optimize, I am surprised to find, JMP profiler recommend anhydrite to be set at

0%. I believe anhydrite at 0.05% is critical for #12 to reach such a good adhesion.

Is JMP profiler could miss important information? Could anyone help explain?

Thanks, Martin

JMP profiler.PNG

Accepted answers

louv

Hi Martin,

Yes that is correct. Looking at the parameter estimates for mixture design models is confusing due to the nature of the co-dependance of the factors. If you read about Cox Mixture parameterization you will find that those p values are calculated by seeing what the effect of going from the low setting to the high setting of a factor while keeping the ratio of the the other components equal to a constant. For me the most helpful output is the Profiler since visualizing the effect of changing the levels of the components is interactive. If you notice the Profiler slopes and curvature agree with the Cox Mixture parameterization and not the conventional parameter estimates.

It would be best if you read up on Mixture Designs (Cornell) since they are a unique type of DOE and present a uniqueness due to that co-dependance of the components. I normally close down all windows except the Summary of Fit, ANOVA and Cox Mixture Parameter estimates.

The bottom line is that due to the correlation between the components in mixture designs the regular t-test is not useful to test significance of each component.

All comments

Bill_Worley

Martin,

Most if not all of my colleagues will tell you that there is no replacement for domain knowledge. What you know about the chemistry and interactions of your system takes precedent over any models you build. JMP only works with the data it is given, it has no intuition as to what is right or wrong or what is needed like you do.

That said the solution that JMP gave is the optimal solution for the inputs and adhesion results you have provided. If you want to make sure 0.05% anhydrite is part of the optimal solution you can use the Profiler to make that happen.

If you Control >left Click anywhere in the anhydrite block you will get a dialogue box that allows you to lock the setting. See the image below. Set your Current Value to 0.05 and then check the Lock Factor Setting at the bottom. You will need to rerun the optimization because your other two factors might shift as well.

Profiler Lock.JPG

HTH

louv

Just an observation. The model Rsq is suspect and if you examine your replicate treatment combination rows, there are a pair, namely row 7 and row 8 which seem quite reproducible (1.02 and 1.01). However if you examine row 4 and row 10 the reproducibility is suspect (1.0 and 0.67). This concerns me as this may indicate there needs to be more understanding built into the test perhaps? The profiler is only as good as your model and in this case perhaps further investigation as to why the treatment combinations represented in rows 4 and 10 adhesions are so different even though they are the same formulation.

louv

Martin,

I had a quick look and simply evaluating the design with row 10 excluded (assuming assignable cause), improves your model and your Profiler agrees with your domain expectation. I would rerun that treatment combination and of course verify the optimal settings.

Screen Shot 2018-03-02 at 10.24.52 PM.png

Screen Shot 2018-03-02 at 9.57.49 PM.png

MartinY

Thanks for taking time help me. The factor fixing is very useful.

Martin

MartinY

Hi Lou:

Thanks for pointing out the huge variations between 4 and 10, I decide to repeat these 2 conditioins.

From my experience, the varations caused by raw materails, testing, etc, could not be this big.

Martin

MartinY

Hi Lou:

Thanks for taking time help me out, really appreciate it.

One question about Cox mixture model: how to interpret " Intercept" in Cox model?

Also in Cox model you attached, The P-value of OPC, Ternal is less than 0.05, so they are significant factors, anhydrite is not significant factor.

From Profiler Optimization, anhydrite is part of the formulation recommended.

So should we stick to the Cox model about significant factors or do we use Profiler Optimization? What is the general procedure for next round experiment?

Thanks.

Martin

louv

Hi Martin,

Since your model is reasonable, assuming the one run has an assignable cause for being so non-reproducible, the Profiler is the approach that I take since the Cox parameterization is consistent with the Profiler. Please note that the anhydrite has a P-value of 0.18 and the polynomial term of anhydrite has a P-value of 0.08 and thus I think it is important to include them. Where I would go from here is to repeat the suspect run to see how the formulation performs. I would also run the optimized setting as a verification of the model. As far as further experiments, I always challenge myself to ask what other process parameters might need to be investigated.

Happy experimenting!

Lou

MartinY

Hi Lou:

When you talk about P-value and polynomial term of anhydrite, I assume you are referring to Cox mixture model table? P-value=0.1953, anhydrite^2=0.1124.

So what is the use of Parameter Estimate table? Should I just ignore it. I am pretty weak in analyzing JMP output, there are so many tables,

Thanks again.

Martin

table .PNG