Food Scientist Jiakai Lu Wins USDA Grant to Study Savory Textures in Alternative Protein Sources
Content
Jiakai Lu, associate professor in the College of Natural Sciences's Department of Food Science, has received a three-year, $611,000 USDA grant to develop a hybrid, computer-aided engineering framework combining multiphysics simulations and a machine-learning approach to better predict the texture of plant protein products that mimic meat, which persists as a major challenge in creating realistic and savory alternatives to conventional meat.
Ultimately, the research aims to revolutionize the way “high-moisture meat analogs” are produced by better understanding how the fibrous structures of plant-based proteins are formed. This in turn holds promise to improve food system sustainability and address the increasing global demand for plant-based protein options.
“Plant-based meat analog has been a hype for the last few years but one of the reasons it’s not becoming more and more popular in the market is because of the taste,” Lu says. “The texture, the fillings, are still a little bit short from real meat. That’s why we want to understand better not just the ingredients but the processing itself to improve the internal structure of the meat analogs.”
Extrusion is the process that turns raw plant material into fibrous structures that resemble meat textures. Plant protein powders are processed under high temperature and shear in the extruder, where proteins denature and align under flow. As the material enters the cooling die, controlled thermal and flow gradients drive phase separation and solidification, forming the fibrous structure of the final product.
“Unfortunately, it’s almost a black box process,” Lu says. “We don’t see what’s happening inside during the process, so a large part of our work is to do computer simulations.”
This, too, has been a challenge because of the complex material being processed. “It’s constantly changing inside the whole process, and it’s very, very difficult to model or characterize it using conventional multiphysics simulations,” Lu explains.
Enter an emerging AI technique called scientific machine learning. Rather than relying solely on data correlations, scientific machine learning integrates fundamental physical laws, governing equations and domain expertise directly into machine-learning architectures. This fusion of physics-based modeling and data-driven inference enables models that are both interpretable and predictive across complex scientific and engineering systems.
Earlier this year, Lu co-authored research in the journal Food Research International that used a type of scientific machine-learning technique to accurately predict the relationship between food properties and sensory perception. The lead author of the paper, Carlos Corvalan, an associate professor of food science at Purdue University, is also the co-investigator on Lu’s USDA grant. The collaboration between Purdue and UMass Amherst is part of a research hub called Scientific Machine Learning for Food Manufacturing. The initiative is focused on optimizing food design and production.
“We have been applying this scientific machine-learning strategy to many food-related areas in the last few years,” Lu says, including ongoing, groundbreaking research to predict the shelf life of bulk oils by tracking antioxidant changes, another USDA-funded project led by Eric Decker, research professor of food science. This research, which aims to both increase food safety and decrease food waste, has been published in the Journal of Agricultural and Food Chemistry; the journal Food Chemistry; and the journal Foods.
In his new project, Lu and his PhD student Carlos Parra Escudero will use scientific machine learning to embed physical‐laws constraints—such as force equilibrium or momentum conservation—within the model architecture.
“The machine thus learns not only from empirical data but from the governing physics of the system. By integrating constitutive modeling, multiphysics simulation and neural-differential equations, the model aims to generate a comprehensive, predictive framework that conventional data-only approaches cannot readily achieve,” Lu says.
The goal is to simulate what’s happening during the structure formation inside the cooling die and thereby guide the design of more efficient cooling strategies. After developing the simulation, Lu will collaborate with another co-investigator, Girish Ganjyal, a professor and extension food processing specialist at the University of Washington in Seattle. Ganjyal’s team has been working to develop novel cooling dies that can optimize the structure formation process.
Lu’s computation-aided method will offer a shortcut in the cooling die design. “The cooling process is the one of the most important processes in making the structures of the meat analog,” he says. “We will do the simulations to create the algorithms and then work with the University of Washington, where the cooling die will be built using our suggestions.”
This story was originally published by the UMass News Office.