Can AI Predict Deliciousness?

Taste remains one of the biggest barriers to wider adoption of plant-based foods. Soham Patnaik, a machine learning engineer at the Food Intelligence Lab, thinks artificial intelligence (AI) may be able to help product developers tackle the problem earlier in the formulation process—before every promising idea has to be made and put in front of a sensory panel.

Speaking at IFT FIRST Annual Event and Expo, Patnaik cited Food Frontier’s 2024 Consumer Study, in which 46% of respondents who said they would not buy plant-based meat again identified poor taste as a barrier to repurchase. Data from Nectar, a Food System Innovations program that conducts large-scale sensory testing, points to a similar gap. Among more than 5,000 omnivores, 30% rated the plant-based meat products as “like” or “like very much,” compared with 68% for animal-product benchmark products. For plant-based dairy, the figures were 35% and 65%, respectively.

The Food Intelligence Lab, part of the nonprofit Food System Innovations, is approaching that taste gap as a design problem. Patnaik described two areas of work: predicting sensory performance and using algorithms to reduce the number of formulation iterations needed to reach a target.

For the first project, the team built a benchmark using more than 21,000 sensory evaluations of 215 plant-based products across 24 categories. Its prediction model can draw from four types of information—ingredients, nutrition, product images, and chemical compounds—to estimate how products might rank in a sensory panel.

The goal isn’t to replace sensory testing, but to use lower-cost information earlier in development to narrow the field to the most promising candidates.

For example, when evaluating a plant-based burger, the model can combine several types of information to estimate how consumers might respond to it in sensory testing. Ingredient and nutrition data can provide clues about factors such as flavor and saltiness, while product images can reveal characteristics such as color, surface texture, greasiness, and even the fibrous structure visible in a cross-section.

So far, the model has shown promise as a screening tool. Across the product categories used in the benchmark, the product that ranked best in human sensory testing was also the model’s top prediction 33% of the time. In 67% of the categories, the No. 1 product in sensory testing appeared among the model’s top three predictions. By comparison, a random top-three selection would have included the No. 1 ranked product in sensory testing about 36% of the time.

“That can help de-risk formulations,” Patnaik said, allowing developers to send the strongest candidates to sensory testing rather than testing every possibility.

A second project, conducted with AI-powered formulation company Proxy Foods, looked at whether algorithms could also help food scientists reach a desired formulation in fewer rounds of experimentation.

For a plant-based yogurt case study, the researchers targeted four sensory characteristics: creaminess, tanginess, consistency, and uniformity. An AI agent selected seven ingredients, and a Bayesian optimization algorithm—which uses results from previous experiments to recommend the most promising next formulations— proposed three initial recipes. The researchers made the products, collected sensory-panel ratings, fed those results back into the system, and asked it to propose the next three formulations.

They repeated the process for 10 rounds, producing 30 recipes in all. In a head-to-head comparison with a food scientist working with the same 40-hour time budget, the algorithm finished within 1% of the target, while the food scientist came within 15%.

Patnaik was careful not to draw broad conclusions from a single comparison. “We would like to run more trials before generalizing,” he said.

The group has tested the approach on another product as well. In a plant-based cottage cheese project, Patnaik said the final formulation received sensory ratings 26% higher than the baseline recipe.

For product developers, the more immediate opportunity may be less about handing formulation over to an algorithm than reducing the amount of trial and error required to reach a strong candidate. Sensory prediction can help narrow the field, while iterative modeling can guide the next experiment based on what has already been learned.

Both approaches still depend on human input. People make the products, evaluate them, and supply the data that allows the models to improve. And, Patnaik noted, better shared datasets and tools will be needed if these methods are to become more useful across the alternative protein sector.

For now, AI may not be able to declare what tastes good. But it may help food scientists decide what is worth tasting next.