More details would help.

Years ago I was involved with scale up and prediction of large polymerization reactors. I understand the reactors are much bigger now.

We used 1 quart reactors to model some of the properties of 35,000 gallon reactors with 100,000 pound monomer charges. many good batches had to be made to pay for one bad batch as well as disposal costs.

The results were to test properties of individual components (suspension agents and surface active ingredients come to mind) to determine dose. Properties tested were cloud point, viscosity and solids concentration. If we got fancy we looked a MW and MW distribution, and degree of blocking of OH groups.

The method was collecting data of randomized batch results and "back engineering" a 2 or 3 factorial DOE. After some success we evolved to multiple regressions using as many data sources as we could find. (One of the most interesting results was related to a quality problem that was tracked to times of the year when the ambient temperature dropped below the chilled brine temperature. Although unexpected, further work found what was going on and a solution implemented.)

Younger engineers have switched to a neural net approach with better results than my DOE and regressions.

Back to your question: we also used adjustments for raw material variation using pH treatment, peroxides, oxygen scavenging, filtering, UV light treatment, salting, adding or reducing shear. As I said before, more details would help.

Please respond if this helps. I only check this link weekly so be patient if you want more discussion.

Thanks.This sounds helpful.

I am more interested in the methodology you used to solve the problem rather than which variables to investigate as I am working with some polymer chemists who can assist me with this.

However any recommendations on how to approach the problem from a modeling and process perspective is appreciated.

What is magnitude of variation of the raw material? What s distribution of variation? Is it fluctuation over time? or random over time?

This sounds like a tolerance stack up problem, only chemical tolerance, not dimensional.

It also sounds like a purchasing problem- your Purchasing dept doesn't want to pay for raw materials that meet your requirements.So you have to bugger up a workaround.

I would not be compensating for "expecting variation"; I would be testing then building recipe based on sample analysis results at each level of expected. As other poster has said, DOE, and Regression are powerful tools.

More powerful though is visit to supplier to improve the process capability of his process to reduce variation that you are being forced to deal with.

If you were toyota, what would you do?

My guess is it would be to send team to supplier... not change your welding process with each new roll of steel...

milo

The raw material which is a problem is a polyester unfortunately neither the manufacturer of the polyester or our own chemists can account for the variation in the polyester. Even if it meets specification or the measured properties are identical a variation in the final product is observed for different batches of the polyester. We are able to consistently reproduce batches of our product with the same batch of polyester.

Attempts have been made to analyze the differences in the polyester but to date nothing has been successful and the variation or cause there of is still unknown.

Therefore for my thesis I am looking at treating the polyester as an unknown.

I see.

But still with out knowing the cause of the variation, you are flying blind.

Your chemists have analyzed the different batches, they tested for chemical attributes.

Have you looked at physical property differences or particle size?

Any other additives?

HAve you created a matrix of the batches that work well vs the batches that 'varied' to tie them to a supplier, a supplier's processing line, etc? With sufficient data you may be able to determine that batches produced on line a at supplier a always conform UNLESS they were run after a batch of "who knows what" in which case those batches appear to vary more...

I respect your efforts to improve process control, but respectfully submit that ultimately determining its cause is the better approach.

I have found that when I put my data in rows and columns it speaks to me.

In this case two columns batches that perform as expected, and batches that dont.

Then I would work the hell out of both lists to see what consistencies exist between each column's residents. And what differences exist between the batches in different columns.

Then you can run DOE and adjust your spec.

milo

I must agree. Finding the underlying cause would be first prize.

The problem is that at the university we have world class equipment for classifying the polyester however this comes at a cost. Upgrading the suppliers equipment is unfortunitally not possible as it is not worth their effort as we are a relatively small customer. Secondly having it analyzed by an external laboratory would be to expensive and therefore the product would become uneconomical.

Here are a few more thoughts on my experience.

In the past we experienced product degradation, contamination and changes with the product between when it was shipped and when it was received. We also saw the need to check shifting of subjective evaluations and test methods. This lead to a sample retention and review program. Our raw materials were made in a continuous process and held in a Day Tank. They were blended and shipped as one lot from a day tank after it was filled and quality checked. I fought many battles about how the day tank was tested and approved. (average of line data, actual day tank tests, could any line data be out of specification if the average and day tank test was in spec?) What was the resolution of the test compared to the spec? The new quality programs (six sigma and ISO) have made it easy to hide and cheat bad product and accountability. (I know I will get a lot of responses about that!)

We developed a data base with two programs. A "thumbprint" program was used to test a retained batch of feedstock with current materials. A "round robin" test program was set up to exchange samples with the supplier and out other plants. We found that a sample from the supplier was often different than the shipment. We returned blind shipment samples for repeat analysis to the manufacturers lab.

Sorry I can't answer past what I offered. My mind keeps gioing back to finding a measureable property, ratio, operating condition in the feedstock. If the Mfg will not help with data or details, and you can not get into his process details you may have to resort to a shotgun approach to testing and regressing data.

Good Luck.

Paddler

I must agree. Finding the underlying cause would be first prize.

The problem is that at the university we have world class equipment for classifying the polyester however this comes at a cost. Upgrading the suppliers equipment is unfortunitally not possible as it is not worth their effort as we are a relatively small customer. Secondly having it analyzed by an external laboratory would be to expensive and therefore the product would become uneconomical.

Hence my attempt at this approach. This is why I was thinking of testing the product on the lab scale and using this data as the input to the model for adjusting the parameters and scaling up.

A similar problem occurs with industries such as the recycling of paper and alloys, where the feed stock continuously varies.