Umetrics Suite Blog

How to optimize chocolate - and profits - with multivariate data analytics

November 6, 2018

Making the perfect bar of chocolate is not just about mixing the right amount of sugar and cocoa, or adjusting the process for product quality.  Another factor that must be taken into account to optimize both taste and profits is the grinding time of the cocoa beans. Let’s take a look at how data analytics can be used to elevate both, and to find the right combination of ingredients and process to support the business goals — an important factor in the food and beverage industry.

how MVDA can be used to improve chocolate making

Creating better chocolate is about the process as much as the ingredients. MVDA lends insights into how to optimize production for taste, price and quality.

Vintage Plantations is an exclusive producer of handmade, bean-to-bar chocolate in Umeå in Northern Sweden. The founder of Vintage Plantations, Jenny Berg or “Chocolate Jenny”, has provided Sartorius Stedim Data Analytics with different kinds of chocolate for a series of consumer experiments. By using multivariate data analysis, we obtained some interesting results that surprised even the chocolate makers.

Different choices depending on business model and branding

Whereas multivariate data analysis is mostly used by the food and beverage industry for process optimization, chocolate is a great example of a food where data analysis can also be used to optimize both taste and profit margins. Apart from the long process of fermenting and roasting the cocoa beans, the exact percentage of cocoa as well as the exact time for grinding the beans to a smooth paste affect both nutrient content and the amount of sugar that must be added to get a palatable taste.

These parameters, or variables, are important in order to optimize product quality. This applies whether the business model is based on offering budget chocolate or premium chocolate, where the branding is based on offering exquisite taste and a high nutrient content.

Optimizing cheaper brands is not only about process optimization

Dark chocolate consists mainly of two ingredients, cocoa and sugar, and a small amount of lecithin for texture (exclusive chocolates often do not contain any lecithin). Optimizing chocolate, however, is not only about mixing a certain amount of cocoa with sugar and then putting all focus on process optimization, but also about grinding. Although the perceived bitterness in chocolate relates more to the cocoa content than sugar content, the grinding time has some effect as well: The sugar hides the cocoa flavors to some extent, but the more you grind, the weaker the impact of sugar will be on the flavors. As sugar is cheaper than cocoa, this would be the direction for budget chocolates – lots of grinding and lots of sugar. While not directly related to process optimization, data analysis could be used in this way to optimize product quality by finding the right balance between grinding, sugar, budget and taste.

chocolate production

Mapping customer preferences can help determine which processes and combination of ingredients are optimal in food and beverage production, such as chocolate manufacturing.

Optimizing exclusive chocolate brands

Grinding the cocoa beans for a long time implies that more nutrients will be destroyed in the process. High-quality chocolate bars are thus normally based on cocoa beans that have been ground for a shorter time. Less grinding also means that the taste of sugar will be noticed more clearly and the chocolate bar will be tasty and sweet on the tongue even with a higher percentage of cocoa. Consumers that are attracted to quality chocolate often want a high percentage of cocoa but still tend to prefer chocolate that is not too bitter. A quality chocolate bar is typically not bitter. This would thus be the direction for producers aiming for high product quality and branding based on offering exclusive chocolates.

Experiments under strict protocols

Vintage Plantations provided Sartorius Stedim with chocolate according to an experimental design with five different combinations of cocoa content and grinding. Three different groups of consumers tried the combinations under strict protocols. The results are interesting from a business perspective as well as in the quest for the perfect bar of chocolate.

MVDA chocolate experiments with useful implications

While there are probably thousands of chocolate brands in the global market, there are, in fact, only a few large chocolate manufacturers. These big companies produce chocolate pellets that other chocolate brands use to make their own chocolate bars. In addition to the big manufacturers, there are also a few hundred small-scale producers, often with a dedicated focus on sustainability and preservation of native cocoa varieties.

Vintage Plantations is one of only four chocolate manufacturers in Sweden that produce chocolate from “bean to bar”, that is, from the cultivation of cocoa beans to fermenting, roasting, and grinding. Their generous participation has given Sartorius a unique opportunity to carry out the experiments. 

vintage plantations chocolate

Vintage Plantations produces handmade, bean-to-bar chocolate in Umeå in Northern Sweden. (Image from Vintage Plantations)

Experimenting with different combinations of grinding and cocoa content

Five different combinations of grinding and cocoa content have been used in a number of experiments where participants at fairs, seminars, and conferences have tasted and ranked the chocolate bars under strict protocols. Vintage Plantations’ cocoa beans are normally ground for 18 hours (many commercial chocolates are based on beans that have been ground for 36 to 72 hours). For the chocolate bars that were produced for the experiments, the grinding time varied between 12 and 36 hours, whereas the cocoa content varied between 50 and 85%.

MODDE, part of the Umetrics® Suite of Data Analytics Solutions, was used to set up a design of experiments (DOE) based on the two parameters: cocoa content and grinding time. As experiments are usually expensive to set up and carry out (and sometimes might take a long time, for example in agriculture and forestry), the purpose of a DOE is to conduct as few experiments as possible while getting as much information as possible.

Out of the five combinations of grinding and cocoa content, two were in fact identical in order to verify the reproducibility of the experiment. All test subjects were asked to fill in a form with details about metadata such as gender, age, nationality, and other parameters that may affect which chocolate bar they prefer. Since the tests were anonymous, each participant was instead assigned an ID number.

Multivariate data analysis in SIMCA

The participants tested the chocolate bars in random order and were asked to rank a number of parameters – taste, sweetness, texture, and bitterness. While taste is subjective, the other parameters can be controlled objectively. SIMCA, part of the Umetrics Suite of Data Analytics Solutions, was then used for the analysis. The interesting parameter is, of course, taste.

The main results from a data analytics model zooming in on the taste attribute are provided in the two scatter graphs seen below. In the left-hand plot, each subject tasting the chocolates is represented by one dot. Dots that are close together correspond to individuals that have similar taste preferences. Dots that are far apart correspond to individuals having very different taste preferences. And, finally, dots that are near the origin of the plot have average taste preferences within this set of chocolate consumers.

The plot is colored according to gender, with a green color for females and a blue color for males. Since both green and blue dots spread uniformly one main conclusion is that there are no differences in in chocolate liking scores related to gender.

In the right-hand plot, we get information on how the five different chocolates (five combinations of cocoa content and grinding time, regulated in the DOE) combine in characterizing the taste preferences of the respondents. The two boxed taste variables in the right part correspond to chocolates with high cocoa content. The two boxed taste variables in the top left part represent chocolates with low cocoa content.

main outputsdata analytics plots

Main outputs from the data analytics model. In the left plot, each dot corresponds to one respondent. In the right plot, each dot corresponds to the taste profile for each of the five chocolates.

A simultaneous inspection of the two scatter graphs will then aid in figuring out what it is that the different respondents like or dislike. As an example, the interpretation can be done on the individual level. The two red arrows assist here: One red arrow points to a person who likes all chocolates (#76) and another red arrow points to a person who dislikes all chocolates (#31).

Alternatively, the interpretation can be done on a non-individual level by trying to group together consumers with similar preference profiles. Individuals in the right-hand part fancy the high-cocoa content chocolate whereas individuals in the top, left-hand part fancy low-cocoa content chocolate.

Implications for the food and beverage industry

Multivariate data analysis has already made inroads into the food and beverage industry as it brings great advantages for process optimization. Most probably, it will also be used to an increasing extent for product development and product optimization. An experiment like this one could, for example, be used to find the optimum level of grinding time and cocoa content, depending on the business model and branding: budget or premium chocolate.

The same process can be applied in other industries looking to maximize quality while reducing the use of expensive ingredients or processing steps without impacting customer preferences.

Want to know more?

Read more about consumer preference mapping in this blog.

Watch a recorded webinar showing how SIMCA software can be used to do preference mapping. (Registration required).
Watch webinar



Topics: Multivariate Data Analysis, Manufacturing Processes

Jonas Elfving

Written by Jonas Elfving

Product Manager at Sartorius Stedim Data Analytics