The task of giving food a novel but delicious taste can be an arduous task largely based on one’s intuition. There are simply too many ingredients to choose from, and finding the perfect match is like finding a needle in a haystack. To offer scientific reasoning as to which certain ingredients work together and to provide a new method for finding such matches, Korea University’s (KU) Professor Kang Jae-woo (Department of Computer Science and Engineering) and his team partnered with Sony AI to develop an ingredient-pairing artificial intelligence (AI) model. For this process, they introduced a mapping tool called FlavorGraph in a research paper titled FlavorGraph: a large-scale food-chemical graph for generating food representations and recommending food pairings in January 2021.

The foundation of FlavorGraph stems from the early works of Professor Kang Jae-woo’s research team. Back in May 2019, Professor Kang’s team introduced KitcheNette, an AI model that could predict food ingredient pairing scores. This AI model was built upon data collected from existing recipes, which allowed the AI to learn which flavors are compatible. However, the team realized that they needed to amass more information on not only food recipes but also food molecules in order to create a more detailed database for the AI model to learn from.

The team’s next step was partnering with Sony AI to further their research. While working with Sony AI, the team attempted to answer two major questions. According to the company, the team was curious as to how they could extract meaningful information from existing recipe data as well as data on food molecules. Assuming the extraction would be successful, they also wondered how they could use such data to train the AI to find new food pairings. After contemplating the answers to these questions, the joint team produced a creative mapping tool called FlavorGraph to aid and complement KitcheNette.

According to the research paper released by the joint team, FlavorGraph’s purpose is to recommend complimentary food pairings by detecting the level of compatibility between ingredients and food compounds. In addition to the data used to train KitcheNette, this time around, the team also incorporated data on chemical compounds in food to enter into FlavorGraph.

Thus, the relations found by FlavorGraph have been able to take on a more chemical approach to food pairing than before, allowing the team to uncover many complementary matches. According to KU research team’s Professor Park Dong-hyeon (Research Institute of Computer Information and Communication), some pairings suggested by FlavorGraph were already well known, such as ice-cream and caramel syrup, but some pairings such as ice-cream and rhubarb or ice-cream, chocolate, and the liqueur Baileys Irish Cream were quite surprising.

The development of FlavorGraph proves that food pairing is indeed becoming more scientific, offering a new method to finding the perfect food combinations. As more improvements are made, it is expected that this AI model will offer more novel opportunities to the culinary world.