As a global leader in the food and beverage industry, Tate & Lyle has long been recognized for its impressive achievements in “making food exceptional” by transforming corn and tapioca and other ingredients into ingredients that add flavor, texture and nutrients to food. One of America’s most famous products from this company is SPLENDA® Sucralose sweetener.
When a corn sugar refining process ran into problems, Tate & Lyle turned to Minitab software to help.
CHALLENGE: Equalize grain size
When Adam Russell started working as a Global Operations Master Black Belt at Tate & Lyle, he was challenged to keep the grain size of the corn sugar the company produces.
“One of the important characteristics for the quality of the single crystallization process was the constant grain size,” Russell said. “Why is this so important? When these products were developed for consumer use 20 or 30 years ago, consumers wanted corn sugar that had the same taste and texture as regular sugar (granulated sugar) or sucrose. I shouldn’t have.”
Tate & Lyle were struggling with an out-of-tolerance issue with these particles, but could not figure out the cause. The company believed in the past that the following factors affected the difference in particle size:
- flow rate
There were many other factors.
Examples of using Minitab
Tate & Lyle started by using Companion by Minitab (now Minitab Engage™) to create a process map that outlines the crystallization process ( Learn more about process maps ). Since the company did not consistently obtain a narrow particle size distribution, they wanted to know what caused the difference and how to control it.
“Everything is measured in a chemical plant,” Russell said. “At every possible point there is a transmitter that feeds the information back to the data historian. So that’s very useful, but the problem is that the information gets so much that you don’t know what to do with it.”
However, many of the relationships between variables are non-linear, making it difficult to determine the effect of one variable on another. Also, during the drying phase, the sweetener becomes a gel-like form called a “slurry” that is intermediate between a liquid and a solid, so the particle size is not known until it is placed in a bag to be served to the consumer.
There are over 1,000 pieces of information that can be entered into these models. Multiple regression models alone could not provide an answer.
With numerous predictors interacting with each other in infinitely complex ways, a systematic approach was needed to identify the predictors that had the greatest impact on particle size. In other words , I needed TreeNet from Salford Predictive Modeler (SPM) .
“It was difficult because we only used traditional modeling techniques,” Russell said. “It was very difficult to understand the relationship between variables and outcomes. Fortunately, with SPM’s TreeNet, it was very easy to focus our attention on key predictors and devise strategies to effectively handle those variables. I worked with Minitab and SPM’s TreeNet I believe the algorithms can work together very effectively. It’s clear that SPM is not a substitute for Minitab or any other statistical program, but I think that together, we get to the answers we want faster.”
Russell used TreeNet’s default settings and adjusted the number of trees. When he started filtering out the predictors, Russell began to understand the effect of each factor by comparing it to a test R-squared value.
To discover the true meaning of these important variables, Russell used SPM’s partial dependency plot. Certain variables were marked on the steep part of the partial dependence plot, indicating their importance. Without the SPM partial dependence curve, we would not have known the importance of these variables.
Next, Russell took a simple step-by-step approach. He took out the variables one at a time and watched what the R-squared would be. There was no significant change until he excluded the fourth most important variable. Russell took this variable to the manufacturing team and asked for more information about the variable.
Russell used SPM’s variable importance rankings to simply reduce over 1,000 predictors to just eight. These eight predictors alone accounted for nearly half of the variation in the test sample.
Using SPM’s “shaving from the top” feature, Russell was able to quickly determine that one variable has a far greater effect on R-squared than any other variable. Although this variable was found to be related to the flow to the crystallization system, the effect of this variable on the final product was not clearly known until Russell created the SPM model.
Russell then used SPM’s partial dependence plot to see why this variable is so important for particle size instability. SPM’s partial dependence plots showed how this variable might change in response to a change if “moved on a distribution curve”.
“We were moving on the steep part of this distribution curve,” Russell said. “On lucky days the coefficient of variation is low, but on the unlucky days the coefficient of variation is high. Without SPM, we would never have known this.”
Believing that the goal had been satisfactorily achieved, Russell found several ways to reduce variations in final corn sugar crystal size and help food manufacturers use these ingredients to improve their products for consumers.
*This case study was written using Companion by Minitab prior to the introduction of Minitab Engage in 2021.