Source-linked AI summary
Generative Artificial Intelligence creates delicious, sustainable, and nutritious burgers
Vahidullah Tac, Christopher Gardner, Ellen Kuhl
TL;DR
Designing foods that are simultaneously delicious, nutritious, and sustainable remains difficult. This study trains generative AI on human-designed burger recipes to learn palate structure and optimize generated recipes across these objectives. The model rediscovered the Big Mac and produced burgers with improved sustainability or nutrition, while some novel designs matched or exceeded its sensory ratings.
Problem
Designing foods that satisfy environmental and nutritional objectives while meeting sensory expectations requires a quantitative understanding of the human palate.
Method
The model learns higher-order burger-design structure from human-generated recipes and uses measurable novelty, environmental, and nutritional criteria to explore generated designs.
Results
The model rediscovered the Big Mac and generated novel burgers with comparable or better palatability, substantially lower environmental impact, or nearly twice the nutritional score.
Takeaways & Limitations
Generative AI provides a quantitative framework for navigating trade-offs among deliciousness, sustainability, and nutrition in food design.
Takeaways & Limitations
The results are limited by biases in source recipes, omission of processing and cooking transformations, aggregated databases, and sensory validation involving limited burgers and participants.
Abstract
from arXiv · showhide
Food choices shape both human and planetary health; yet, designing foods that are delicious, nutritious, and sustainable remains challenging. Here we show that generative artificial intelligence can learn the structure of the human palate directly from large-scale, human-generated recipe data to create novel foods within a structured design space. Using burgers as a model system, the generative AI rediscovers the classic Big Mac without explicit supervision and generates novel burgers optimized for deliciousness, sustainability, or nutrition. Compared to the Big Mac, its delicious burgers score the same or better in overall liking, flavor, and texture in a blinded sensory evaluation conducted in a restaurant setting with 101 participants; its mushroom burger achieves an environmental impact score more than an order of magnitude lower; and its bean burger attains nearly twice the nutritional score. Together, these results establish generative AI as a quantitative framework for learning human taste and navigating complex trade-offs in principled food design.
Results
Generative AI learns higher-order structure in human-designed burger recipes, rediscovering the Big Mac and generating novel burgers that navigate trade-offs among palatability, sustainability, and nutrition. Sensory evaluation and objective scores show that different generated designs can optimize deliciousness, environmental impact, or nutritional quality.
- Classic rediscovery and novelty: SDS = 0 identifies a match with a reference recipe, while SDS > 0 quantifies novelty; the model rediscovered the Big Mac and generated increasingly novel delicious burgers.Delicious Burger 1 had SDS = 3 and Delicious Burger 2 had SDS = 6, while retaining familiar burger structure.
- Classic rediscovery and novelty: Delicious Burger 1 exceeded the Big Mac in flavor, and Delicious Burger 2 exceeded it in overall liking and flavor, while both had statistically similar texture ratings.The blinded sensory evaluation included 101 participants.
- Nutrition and personalization: The Nutritious Burger achieved a healthy eating index of 63.12 versus 33.71 for the Big Mac while reducing environmental impact by a factor of six.Its improved profile included more vegetables, whole grains, and plant protein, alongside less refined grains, sodium, and saturated fat.
- Model validation: The model captures higher-order burger-design regularities, including ingredient co-occurrence, recipe length, and quantity distributions.It learns latent structure directly from large-scale human-generated recipes rather than reproducing only superficial statistics.
- Implications and limitations: The study defines food formulation as a quantitative, data-driven design process that uses recipes to navigate measurable taste, nutrition, and sustainability objectives.The authors frame recipes as a human-centered interface between culinary tradition and artificial intelligence.
- Implications and limitations: The model inherits biases from existing recipes, omits processing and cooking transformations, relies on aggregated environmental and nutritional data, and was tested on limited burgers and participants.The authors state that broader studies are needed to establish generalizability across populations and contexts.
Supplementary File 1
The paper is titled “Generative Artificial Intelligence creates delicious, sustainable, and nutritious burgers” and is authored by Vahidullah Tac, Christopher Gardner, and Ellen Kuhl.
- The study presents generative artificial intelligence for creating delicious, sustainable, and nutritious burgers.
- Vahidullah Tac, Christopher Gardner, and Ellen Kuhl are listed as the authors.
Data preparation
The authors construct a standardized burger-recipe dataset from open-source food databases by identifying burger recipes, parsing heterogeneous ingredient descriptions, normalizing measurements, and reducing ingredient categories.
- Burger recipes are identified using both recipe names containing “burger” and a large language model’s classification, requiring both signals for inclusion.
- Ingredient descriptions are parsed into ingredient name, quantity, and unit, then normalized into numeric quantities and standardized measurements.
- 522,517 recipes from the combined food.com database yielded 3,882 identified burger recipes for training-data construction.
- Manual review corrects parsed inconsistencies, producing a curated dataset of 3,882 recipes with 3,683 unique ingredients and 171 measurement units.
- Ingredient names are consolidated by removing descriptors, merging related products, correcting spellings, and filtering rare ingredients.
- Measurements are converted to grams, ingredient volumes to weights, and recipes are normalized to 500 kcal before storage as a 2,216 × 146 ingredient-weight array.
Modeling
The model represents burgers through ingredient selection and quantities, using a two-stage diffusion framework to generate recipes from 146 ingredients and learn their statistical structure.
- A burger recipe is represented as a selection of ingredients and their quantities, with 146 unique ingredients available.
- The two-stage approach first identifies ingredients with a mask model and then predicts their quantities with a value model.
- Multinomial diffusion models the discrete ingredient mask, while a score-based generative model models continuous ingredient weights.
- The training data contains 2,216 recipes, and ingredient-frequency and weight distributions summarize its most-used ingredients.
- The multinomial diffusion process repeatedly resamples categorical ingredient states, with β_t controlling the chance of uniform resampling.
- The score-based model learns a conditional distribution of ingredient weights given the recipe mask using forward noising and reverse denoising processes.
Training and validation
The mask and value models use separate neural-network architectures and optimization settings, with the value-model data split into training and validation subsets.
- The mask model uses an embedding layer and three fully connected layers of 512 neurons each, trained for 100,000 epochs.
- The value model uses four hidden layers of 256 neurons each and is trained for 20,000 epochs.
- The value-model dataset is split into 80% training and 20% validation data.
Rediscovering the Big Mac®
The study uses the proprietary Big Mac as a benchmark for recipe rediscovery and develops environmental and nutritional criteria for selecting alternative burger recipes. It also presents repeated recipes across sustainability and nutrition score ranges, including personalized formulations.
- Rediscovering the Big Mac®: The Big Mac® benchmark is approximated as a 500 kcal recipe containing seven ingredients with specified quantities.The estimated formulation includes 81 g ground beef, 74 g bun, 32 g mayonnaise, 27 g pickle, 27 g onion, 21 g lettuce, and 18 g cheese.
- Rediscovering the Big Mac®: The Substantial Difference Score compares recipes using ingredient presence and quantity ratios, with SDS = 0 defining a match and SDS > 0 indicating novelty.The score is the sum of binary ingredient-level distances across the database.
- Sustainability: Environmental impact scores combine land use, greenhouse gas emissions, eutrophication potential, and scarcity-weighted water use into ingredient- and recipe-level sustainability measures.Ingredient impacts are mapped from a 146-ingredient model onto environmental database categories and aggregated using ingredient quantities.
- Nutrition: Nutrition-based selection identifies repeated recipes in the 90th and 95th percentiles of the Healthy Eating Index and balanced hybrid nutrient density score distributions.The supplementary figures include bean-corn blends and alternative formulations selected by HEI and bHNDS, as well as recipes personalized to individual demographic and activity profiles.
- Nutrition: A personalized nutrient profiling model generates burger recipes tailored to differences in age, sex, body composition, physical activity, and physiological status.The examples cover sedentary adults, a very active adolescent, and a moderately active older adult.
Supplementary File 2
The supplementary file documents the study team, blinded restaurant-based sensory evaluation, participant survey design, and the six-burger preparation and labeling procedure.
- Sensory evaluation: The sensory evaluation enrolled voluntary participants from the general population to assess all six burgers in a restaurant setting.Participants evaluated burgers after answering background questions.
- Sensory evaluation: Participants rated overall liking, flavor, and texture on 7-point Likert scales and reported selected flavor and texture attributes.The survey also collected demographic, dietary, purchasing, and food-priority information.
- Survey design: The supplementary materials classify respondents by age, gender, education, dietary preference, burger consumption frequency, grocery purchasing habits, and food priorities.Food priorities included health, price, taste, convenience, familiarity, environment, and animal welfare.
- Blinding and preparation: The six burgers were prepared and photographed before sampling, with Sustainable Burger 1, Delicious Burger 1, and the Nutritious Burger assigned blinded labels 413, 674, and 491.The sensory evaluation took place at an active restaurant.
Sensory feedback
The sensory questionnaire measured burger acceptance through separate ratings of liking, flavor, and texture, supplemented by selected descriptive attributes.
- Liking: Overall liking was measured on a 7-point scale ranging from “Like very much” to “Dislike very much.”
- Flavor: Flavor feedback included attributes such as meaty, off-flavor, bland, earthy, smoky, sweet, salty, fatty, savory, and good aftertaste.
- Texture: Texture feedback included attributes such as chewy, crispy or crunchy, crumbly or grainy, firm, soft, moist, dry, tough, fibrous, gummy, and sticky.
Population demographics
The sensory study included 101 participants described as a broad cross section of the general population, with reported distributions across demographics and dietary preferences. Taste and price were the most frequently selected purchasing priorities.
- Demographics: The 101 participants were distributed across age groups, with 22% aged 18–25 and 16% older than 55.The remaining participants were distributed across the 26–35, 36–45, and 46–55 age groups.
- Demographics: Participants were 47.5% male, 47.5% female, 3% non-binary, and 2% who preferred not to say.
- Dietary preference: Dietary preferences comprised 65% omnivores and 35% flexitarians.
Guidelines for Recipe Development
The document provides five AI-generated burger ingredient lists and states that chefs should use the listed ingredients and quantities to develop the recipes.
- The study includes five burgers, with each burger’s ingredients and quantities provided for chef development.
- The first burger list specifies beef, bun, ketchup, lettuce, mayonnaise, onion, cheese, pickle, and tomato.
- The second burger list specifies beef, brown sugar, bun, garlic, onion, cheese, sauce, thyme, tomato, and Worcestershire sauce.
- The third and fourth lists specify mushroom-based recipes, respectively using arugula, mayonnaise, mustard, oil, portobello, rosemary, and garlic, or bacon, beef, cheddar, ketchup, mushroom, onion, and pepper.
- The fifth list specifies a bean-based recipe containing bun, cilantro, cornflour, cumin, egg, jalapeno, oats, oil, onion, oregano, and black pepper.
Guidelines
The guidelines constrain chefs to use the listed ingredients and quantities while allowing them to determine processing, ingredient varieties, and certain preparation choices.
- Chefs must develop recipes with the listed ingredients and amounts without adding or removing ingredients.
- Processing steps are left to the chefs, and singular ingredient labels are used for convenience.
- Chefs should adjust salt to taste, record the amount used, and may discard peels, stems, sautéing oil, or aromatic herbs after use.
- Overflowing burgers are acceptable, but chefs are asked to prepare presentable burgers with structural integrity.
- Chefs may choose ingredient varieties, interpret beef as intact meat or ground beef, and select suitable options for unspecified cheese or sauce categories.
Recipes with ingredients and preparation instructions
The chef document contains AI-generated ingredient lists, quantities, preparation steps, and photographs of assembled burgers, with the Big Mac ingredient list included for comparison.
- The document gives AI-generated ingredient lists, their quantities, preparation steps, and photographs of the chef-assembled burgers.
- It also includes the Big Mac ingredient list while using a purchased Big Mac for the survey’s authentic comparison experience.
Big Mac®
The recipes include a purchased Big Mac benchmark and five burgers spanning deliciousness, sustainability, and nutrition designs, each with specified ingredients and preparation instructions.
- Big Mac®: The Big Mac benchmark contains ground beef, bun, mayonnaise, pickle, onion, lettuce, and cheese, and was purchased rather than chef-prepared for sensory authenticity.
- Delicious burgers: Delicious Burger 1 uses ground beef, bun, ketchup, lettuce, mayonnaise, onion, cheese, pickle, and tomato.
- Delicious burgers: Delicious Burger 2 combines beef with brown sugar, garlic, onion, Gruyere cheese, remoulade, thyme, tomato, and Worcestershire sauce.
- Sustainable burgers: Sustainable Burger 1 is a portobello mushroom burger with bun, garlic, mayonnaise, mustard, oil, rosemary, and salt.
- Sustainable burgers: Sustainable Burger 2 combines bacon, ground beef, brioche bun, cheddar, ketchup, crimini mushroom, onion, pepper, and salt.
- Nutritious burger: The Nutritious Burger is bean-based and includes kidney beans, bun, cilantro, corn flour, cumin, egg, jalapeno, oats, oil, onion, oregano, pepper, and salt.