Imagine a world where AI-generated recipes grace our plates, personalised nutritional recommendations guide our dining choices, and sustainability becomes the norm when deciding what to eat.
While calorific information is already available with a Google search, the information is segmented and not always standardised. The problem is similar for information of different types of food and their benefits, the genes they affect, the diseases they are linked with, or even the carbon footprint they leave behind.
Enter artificial intelligence (AI). If you can clean up and organise the data that underlies knowledge, an AI can be trained to learn from it. To that end, since 2017, Professor Ganesh Bagler and his lab at the Indraprastha Institute of Information Technology Delhi (IIIT-D) has been turning food into data, building databases on everything from nutrient values to tastes and food-disease relationships.
With RecipeDB, his lab has compiled over 118,000 recipes, ingredients, nutrition, and flavour profiles as well as health associations. In FlavorDB, his lab has catalogued 25,595 flavour molecules associated with taste and odour. DietRX has catalogued data on 21,207 positive/negative food-disease associations for 1,781 food entities in 24 categories, mined from over 36,000 research articles published over 90 years.
While the hard work of data collection is ongoing, the true potential is realised when AI starts learning from this data — and creates its own recipes. This is the vision of computational gastronomy that Professor Bagler is interested in – what he calls “data science that blends food and computing”.
Finding common chords in our food
Says Bagler, “My journey in computational gastronomy started serendipitously when I was teaching a ‘Complex Networks’ course at IIT Jodhpur in 2014. When discussing a research article that investigated food pairing phenomenon in recipes using the flavour network of ingredients, I wondered whether the recipes from Indian cuisine would have a generic pattern of ingredient combinations.”
Following a year-long investigation, his team found that Indian recipes tend to blend ingredients with contrasting flavour profiles. “Importantly, we realised that spices form the molecular fulcrum of Indian recipes, an observation that struck a chord with anyone familiar with the Indian way of cooking,” Bagler adds.
Bagler says there is already a demand for the use of his data to create applications and models including large language models (LLMs). But his lab does not want to stop at this. They also want to create their own AI solutions while making the data available through a commercial model.
This is where his own AI, a recipe-generator called Ratatouille, comes in.
“Building on the query of Alan Turing, who wondered whether the machine could think, we asked questions deeply rooted in the heart of artificial intelligence. Can machines think like a chef? Can they create novel recipes? Can such AI-generated recipe instructions fool a chef into thinking they are real?” Bagler says, sharing some of the questions his team looked at when developing the AI.
Most importantly, an AI-generated recipe must be tasty, he acknowledges. To evaluate Ratataouille, his team built a “Turing Test for Chef” — a platform to evaluate AI recipes by engaging personal chefs and other experts. If the chef cannot tell the difference between an AI-generated recipe and a real one, the AI’s version passes, much like the equivalent Turing Test for artificial intelligence.
With a proof-of-concept in hand, Bagler’s team is working on strategies to make the AI recipes more nuanced — capable of following constraints like culinary style, nutritional profile, dietary preferences, cost and even their carbon footprint (another database, SustainableFoodDB, lists the estimated carbon footprint of different recipes).
When food becomes data
Bagler sees parallels with computational gastronomy and the digitisation of photography — which democratised the medium and created a vibrant ecosystem by accelerating what was once a slow, cumbersome process. “Similarly, computational gastronomy presents a quantifiable and data-driven approach to food,” he adds.
“With the revolutionary computer and data science tools unleashing disruptive technologies on the world, food stands at the tipping point of a data revolution. In this new world, we should be able to create recipes that are not only palatable but tasty, nutritionally enriching, targeted for specific diet-linked disorders, affordable, and ideally, have minimal environmental impact,” he says.
Democratising AI for food recommendations could help ordinary users more easily find recipes that suit their palette, and it can even play a role in the growing field of personalised nutrition.
“By factoring in inter-personal variations in the genome sequences, gut microbiome, dietary intake, and such, it is eminently possible to arrive at personalized food recommendations for improving nutrition and health. Beyond DietRx, which provides food-disease associations, we are building databases and tools to act as AI Co-Pilot and personalised nutrition coaches,” Bagler shares.
Bagler hopes that these AI tools will help to shape a sustainable food system. “Such a deep-tech-enabled food ecosystem is crucial for addressing the challenge of feeding an anticipated global population of 10 billion by 2050,” he concludes.