Prof, scholar develop AI tool "Ratatouille" that thinks like a chef, generates 1.18L recipes

New Delhi: Can an AI-generated recipe help someone win a cook off ?

"Ratatouille", an artificial intelligence (AI)-powered tool, developed by a professor and a research scholar of the Indraprastha Institute of Information Technology (IIIT), Delhi, could make this possible.

Designed to "think like a chef", the tool can generate more than 1.18 lakh "unique" traditional recipes from 74 countries, claimed its developers -- Prof. Ganesh Bagler and PhD scholar Mansi Goel of the institute's Infosys Centre for Artificial Intelligence.

Moreover, "Ratatouille" -- a name inspired from an animated movie in which a rat helps an aspiring chef become one -- has passed a rigorous test conducted by a 24-member panel of faculty chefs from the Institute of Hotel Management Catering and Nutrition (IHM) in Delhi's Pusa, according to them.

Bagler -- a professor of network science, network biology and computational gastronomy -- and Goel -- a student at the centre -- developed the tool in a lab at the IIIT over nine years using large language models, a technology that drives ChatGPT, a widely used AI-powered platform for writing and learning.

"We began around nine years ago, exploring recipes and ingredient patterns. We have done three main things here. First, we compiled data and created structured databases of recipes, including over 1.18 lakh recipes from 74 countries, along with data on flavour molecules, nutrition profiles and carbon footprints," Bagler told PTI.

He said then algorithms were applied to this data to generate "novel recipes" that follow patterns found in existing recipes.

Talking about the test "Ratatouille" underwent, he said, "Finally, we conducted a 'turing test for chefs'. We presented chefs with recipes from our database and AI-generated recipes in a format that made it impossible to distinguish between them."

"We also asked the chefs to determine if the recipes were genuine or computer-generated. We achieved a 70 per cent success rate, meaning the chefs couldn't reliably tell the difference," Bagler said.

The IHM panel of chefs evaluated 1,472 recipes on a scale of zero-to-five, where zero meant the recipe was fake and five meant it was genuine, he said.

Next, the developers planned to evaluate the recipes by actually cooking them and using sensory panels to assess taste, flavour, aroma, and aftertaste.

This helped make the AI-generated recipes "not only novel but also palatable", Bagler said.

The project, which aimed at enhancing culinary creativity by leveraging AI and making food computable, has been published in the Nature Partner Journal, 'Systems Biology and Applications'.

The idea to come up with such an AI tool struck Bagler while he was taking a class at the IIT-Jodhpur in 2015.

"It started with my curiosity about food. I realized over time the potential (of the project) for industrial and societal applications. The field of computational gastronomy, which originated in India, offers significant opportunities for innovation and improvement," he said.

"While AI can't replace the creative aspect of cooking, it can assist chefs by generating recipes and offering new ideas. We aim to support chefs in exploring new culinary possibilities and making personalised recommendations," Bagler said.

"Ratatouille" can assist in making meal kits and in nutrition coaching by offering customised recipes, the professor said and added that he plans to offer this as a B2B service and is open to making data available for non-profit purposes.

The one challenge, he feels that this project may face is collecting data as per individual needs. "We've gathered data at the cultural level but need more individual-level data for personalised recommendations. Ensuring the quality of this data is critical," he said.

Inputs from PTI.