INFORM February 2025 Volume 36 (2)

12 • inform February 2025, Vol. 36 (2)

NO CHATGPT FOR FOOD CHEMISTRY IFF has been actively hiring data scientists to help analyze the mountains of experimental results it is collecting. A lot of young scientists now begin learning programming in their first semester at university, Møller said, which is “a fantastic advan tage.” Beyond programming, though, data scientists in the food field need to communicate across different business and research roles on a project, from a scientist focused on fats to a margarine producer. “It is very much about collaborating with different inter nal and external people, trying to interpret what it is that they want and how they can get it,” he said. Of course, not every scientist wants to program, and as yet, there is no ChatGPT-like interface that plugs people into the wide world of triglycerides data. “Tools like ChatGPT and Microsoft Copilot are nice, but they are all trained on text,” Møller said. “For us to use AI on our data, we have to build the models ourselves.” At the same time, the data sets on fats in food chemistry remain either proprietary or limited or both. Marangoni advo cates for putting more resources into hosting central reposito ries and creating useable interfaces for them. “Just having the numbers is not enough,” said Marangoni. “You have to give people the ability to use those numbers. Then we will see the real value of data science coming around.” Christina Nunez is a writer and editor based near Washington, DC. She writes about science, technology, and innovation for a variety of organizations, including National Geographic and the US Department of Energy.

confined to simple and/or theoretical studies of cutting-edge fats that may or may not be able to be scaled. Industry research is on an accelerated path to a viable product, which lends particular value to assistance from AI. “Consumers expect meat-free and meat-analogue prod ucts to maintain the same texture, handling properties, and eating experience, regardless of the fat source used, which adds to the complexity of food processing,” Bhattacharya said. To that end, Møller said IFF is using its analytical data to predict not only microstructure and mechanics but sensory qualities such as color, softness, or flavor. This can cut down on the amount of sensory judging panels needed during the development process. “I often say that sensory panels are probably the most expensive instrument we have in a company like ours,” Møller said. “It is very time-consuming to calibrate and run them.” A panel might have 10 to 15 judges, all of whom have been trained specifically for that panel. This training, known as calibration, ensures that everyone understands the prod uct’s attributes and is consistent in evaluating them. But even supertasters get tired after sampling many items in succes sion. Imagine differentiating between 10 marinades that you tasted consecutively. To avoid tasting fatigue, samples are repeated and randomized. Then it may be time to go back to the drawing board—but by the time another panel needs to be conducted, the judges might have forgotten the calibration. “Humans have a short memory, so that is another reason why we need to develop these methods,” he said, referring to their data analysis.