Big and Small AI Aids Product Development

Using NotCo’s process, primarily the AI engine it calls Giuseppe, the Kraft mac & cheese project took eight to 10 months instead of what they claim would have been 24 months.

The process: First you score your product, giving quantitative measures to gold-standard characteristics as well as goals for the new variation you’re trying to create. You feed that data into Giuseppe and the AI platform recommends formulas or recipes to the product developer.

It's a lot more complicated than that, but we’ll get into that in a minute.

Giuseppe dates back 10 years but largely was under wraps until the Kraft project, as well as one that developed an eggless custard for Shake Shack. Several other Kraft Heinz products also resulted, such as Not Mayo, Not Cheese Slices and Not hot dogs and sausages. While those projects were custom and proprietary applications of Guiseppe, NotCo this year started offering the Guiseppe engine to all food & beverage processors.

NotCo developed a cocoa-free brownie prototype for demonstration purposes. It solves two current priorities in product development: finding a replacement for spiraling cocoa costs and improving sustainability.

The first module within Giuseppe is Concept Quant. “First you create the concept from scratch,” says Bernardo Moltedo, culinary scientist manager in NotCo’s Chile headquarters. “The first goal was sustainable replacements for the cocoa. We also add the description of what makes a brownie.”

That’s enough to start, but those are pretty vague goals. “We also have the functionality to incorporate the brand identity and recommend concepts that would fit that brand identity, or target audiences,” adds Alisia Heath, NotCo’s vice president of research and development for business-to-business.

From there, the process moves into FomulateOS, another subset of Guiseppe, where the product developers add more specific needs of the product, including the traditional recipe and all the ingredients for a generic brownie. It also assumes a standard process for manufacturing the product – something the client should alter to match his processing abilities.

That module also starts recommending ingredients – some that Giuseppe has in its database, including ones geared toward the tastes and other demands of different countries, plus the ingredients the client has on hand or is familiar with, as well as novel ingredients that should be considered. “From there you get a good, skeletal formula,” says Moltedo.

The final module in Guiseppe is Synthesis, an optimizer. The product developer needs to further refine the variables and objectives, and Synthesis will search its database of 500 ingredients to find the optimal ones for your cocoa-free brownie. “Instead of 100 trials with different ingredients, this will narrow the field with the objectives and variables you input, also pricing, to the first five suggested formulations,” continues Moltedo.

The product developers will test those first five formulations and input their reactions and further suggestions and instructions – looking to fine tune texture, color, sweetness vs. bitterness, moisture. “The learnings come from doing the physical trials on the bench top and having the [client’s] team report on sensory,” says Heath. “Then input into Giuseppe how well this trial turned out, how close to that control product are we or how close we are to the product goals.”

“Synthesis will learn from those first five trials, give you another five. You may end up doing 15 trials, but not 100,” says Moltedo. “This could be a loop until you get the desired product. You can go back into Synthesis to rewrite sweetness, moisture, color and other attributes to get to the final product.”

Even this “final” round probably consists of three or four formulas, which should be put to the ultimate test: consumer panels. The best recipe might be chosen right away, or the consumer responses could provide more input for another round in the Guiseppe process. Eventually the outcome will be the final product, hopefully perfect.

“Giuseppe was built to address the pain points of product developers, not to replace them,” concludes Heath, “to accelerate their ability to deliver solutions on the bench top.”

A smaller approach

NotCo may be built for big companies. An approach for smaller companies comes from a newly formed company. “FlavorMind is built for small to mid-sized shops or beverage brands that already have popular drinks but need to hit a nutrition target, find ingredient swaps or lower costs without sacrificing flavor, et cetera,” says Ben Wolpert, one of the two founders.

Wolpert is a software engineer. The other founder, Alexia Ciarfella, is a former product developer who, in six years with Mondelez, built a machine learning model used to augment the Oreo cookie recipe development process, which reduced prototype iteration time by 33%. Their collaboration is symbolic of AI’s melding of data science with the art of the food scientist.

“What FlavorMind needs is a copy of your ingredient inventory, current recipe and cost sheet. That’s usually a spreadsheet and a quick call,” says Wolpert. “Our trials are a handful of recipe trials over a few weeks, using the ingredients they already buy or have easy access to.”

They warn that an initial project for a company will require a fair amount of up-front work, primarily to “digitize” and quantify that company’s products, ingredients, corporate goals and other critical inputs.

“You need to digitize a robust set of characteristics, things that you can grade somehow in a quantitative way,” says Wolpert. “You need to have them in a database rather than just written down somewhere, having them in a way that the model can work with.

“This is a necessary first step, and a lot of companies will need to get there at some point anyway,” he continues. “One of the challenges is collecting this data or formulating a coherent way for you to capture this data, the characteristics that describe the product.”

“Sweetness – how do you talk about that in numbers?” asks Ciarfella. “You need to quantify the total amount you want, the calories, the particle size, the desired sweetness level, its sensory attributes. You can get creative with how you describe things, but you need to get it all into numbers.”

“You can replace sugar with something that gives the same amount of sweetness, but it may bake differently, affect texture or act differently as the product ages,” Wolpert continues. An AI program should predict those negative outcomes before you bake the first batch of cookies.

There are going to be attributes that don’t seem to lend themselves to numbers – like packaging, environmental impact, corporate goals. “But you can create scales and rate them on those scales and incorporate those numbers into the formula just as you would carbohydrates,” says Ciarfella.

As an example, FlavorMind reformulated a tropical smoothie for a retailer. While there was not a lot of added sugar to begin with, they still wanted to reduce sugar. “You could replace the mango puree with something that seems as sweet but has less sugar, but you don’t want to lose the overall profile of the original recipe,” says Wolpert. “We were going to replace some mango puree with coconut water, but the taste lacked a certain brightness. So the model added back some acidity via lime juice or lemon juice.”

Ultimately, samples were tested with consumers, who also provided quantifiable feedback that led to further tweaks in the recipe. The end result was a hit with customers.

The same smoothie client also had a problem with wasted spinach. They had to buy it in bulk for one of their green drinks, but always ended up throwing some away as it spoiled. So the program found ways to use it in some of the existing products and also suggested new recipes that used spinach.

“I think the approach a lot of companies will take is that [they would like to minimize] the expensive processes that take humans a lot of time,” concludes Wolpert. “You would rather find out that a formula doesn’t work in a program run than wasting vats of a product on the plant floor.”

Speaking of the plant floor

Much has been written about artificial intelligence’s use in the plant – for machine learning, predictive maintenance, quality control, supply chain optimization and overall intelligent automation. How can an AI system start in the lab but continue to assist when the product leaves the lab — for a pilot plant and eventually to a production facility?

BCD iLabs (bcdilabs.com) created its AI program, Innov8 OS, for the product development lab but it also has a reach into production processes. "Innov8 OS offers end-to-end functionality to complete the entire product development R&D lifecycle and extend seamlessly to managing production processes at scale,” says Vinay Indraganti, founder and CEO of BCD iLabs.

But its foundation remains in the lab. “Innov8 OS optimizes R&D workflows, tailors formulations for specific attributes, enhances taste and texture profiles and enables hyper-personalized product development based on cost and regulatory restriction,” he says. “We look at what customers have for sweeteners, colorings, flavors, etc. We look at their goals, what they’re trying to achieve. We can provide them a limited set of experiments with a high degree of certainty that one of them will solve their formulation challenge.

“All of their targets can be addressed in a single set of experiments,” he continues. “Instead of doing 50 experiments, you’ll be done in 20 – and that’s the first time. Eventually, with the system’s familiarity with your company and products, the number of experiments will be down to 10 and maybe even five. After using the platform for a year, one of our customers successfully developed a stable formula in just a single experiment.

“But you cannot take the human out of the equation,” Indraganti concludes. “The food technologist remains key, but we make his or her research more efficient.”