Purdue researchers use AI to predict texture in food, transforming development process

A team of researchers at Purdue University’s College of Agriculture has developed an artificial intelligence model that accurately predicts how food will feel in the mouth, offering a major advance in food design. The innovation, described in a new paper published in Food Research International, comes from the Transport Phenomena Laboratory within Purdue’s Department of Food Science.

Led by Associate Professor Carlos Corvalan, the team created a sensory-based autoencoder – a type of neural network – that can forecast texture perception using a minimal dataset. This tool allows food scientists to link measurable lab data with the subjective eating experience of texture, known in the industry as “mouthfeel.”

“We’ve developed an AI tool that predicts how food feels in the mouth based on physical properties we can measure in the lab,” said Corvalan. “It allows us to pave the way for smarter food design.”

Predicting how a food will taste and feel typically relies on trial and error. Developers create and test formulations in cycles, guided by tasting panels composed of individuals with subjective preferences. While instruments can assess certain physical attributes like viscosity or elasticity, predicting how these translate to sensory experience has remained elusive – especially for complex, non-Newtonian foods like yogurts and sauces.

“There is no equation for predicting sensory feelings,” Corvalan said. “The link between quantitative properties and subjective feeling is very complex.”

That complexity was the challenge Corvalan presented to students in his graduate-level course Scientific Machine Learning. Paul Kraessig, then an undergraduate majoring in computer science and honors mathematics, led the effort as the paper’s first author. Collaborating with Corvalan and fellow students Shyamvanshikumar Singh and Jiakai Lu, Kraessig trained the model using data from a previous study on the perceived thickness of bouillon samples. The team’s autoencoder used cross-validation techniques to make the most of a small dataset, avoiding the typical requirement of vast amounts of training data.

“Machine learning can sometimes be a bit of a black box,” Corvalan noted. “This shows you can do real-world predictions with very, very few data points and careful validation.”

Lu, who earned his master’s and PhD at Purdue, is now an assistant professor at the University of Massachusetts Amherst. Together with Purdue, UMass Amherst is part of a collaborative research initiative called Scientific Machine Learning for Food Manufacturing, which focuses on enhancing food design and production using AI.

The model developed by the group is already proving valuable. Instead of conducting multiple rounds of taste tests and texture tweaks, food developers can now use the autoencoder to guide formulation decisions from the outset. This can reduce development costs, shorten product cycles, and lead to more consistent results.

“The main objective is to design more appealing foods,” Corvalan said. “With this technology, we can determine how to improve texture while managing the cost and quality of ingredients.”

The implications go beyond product appeal. Texture is not only a factor of enjoyment but also a critical element for individuals with specific dietary needs. For people who have difficulty swallowing – such as elderly individuals or stroke patients – the texture of food can make the difference between safe consumption and serious health risks.

“With this tool, we can reverse-engineer foods that are tailored to people with particular needs,” Corvalan said. “It’s very, very important to get the texture right in those cases.”

The group has published its findings as open research, inviting food scientists and developers to test, refine and build upon the model. Interest in the research hub is growing, with other academic institutions and industry partners expressing interest in joining the initiative. Future projects will include designing plant-based foods, including fish alternatives, that replicate the mouthfeel of conventional animal products.

With support from the US Department of Agriculture, the work coming out of Purdue’s Transport Phenomena Laboratory represents a significant step forward in the integration of machine learning into food science. By making food development faster, more cost-effective, and inclusive, this approach signals a shift in how new products are brought from lab to table – not only with better flavor but with textures designed to meet the needs and preferences of a diverse range of consumers.