January 21, 2025
2 minimum read
Sensitive electronic tongue can taste when juice starts to spoil
AI analysis and chemical sensors determine drink dilution, freshness, and type

The food and beverage industry has struggled for decades to find automated ways to “sample” products at the speed and scale of mass production. But in a new study, researchers used machine learning to overcome the limitations of a promising type of chemical sensor. That means robot tongues may soon appreciate milk and merlot before humans.
When ions in a liquid (such as a tasty drink) come into contact with the conductive sheet of an ion-sensitive field effect transistor (ISFET), the current flowing changes based on the exact composition of the liquid and the applied voltage. This allows scientists to use ISFETs to convert chemical changes into electrical signals. The chemical composition of any drink, and therefore its taste, is affected by contamination and freshness, which ISFET can identify.
“The food industry has a lot of problems in knowing if the food is adulterated or contains any toxic substances,” said Saptarshi Das, an engineer at Pennsylvania State University. says. Although the first ISFET was demonstrated over 50 years ago, this sensor is not widely used commercially. The advent of graphene, an ideal conductive material, has helped researchers create improved ISFET sensors that detect specific chemical ions. However, a major problem remained. Measurements vary from sensor to sensor and due to changing conditions such as temperature and humidity.
About supporting science journalism
If you enjoyed this article, please consider supporting our award-winning journalism. Currently subscribing. By subscribing, you help ensure future generations of influential stories about the discoveries and ideas that shape the world today.
in natureDas and his colleagues addressed this problem by combining ISFET and neural networks to train a machine learning algorithm that uses sensor readings to classify drinks. The resulting system can determine whether milk is diluted, distinguish between soda brands and coffee blends, and identify different fruit juices while determining freshness.
During development, the team attempted to train based on human-selected data points, but given the algorithm all of the device’s measurements and selected its own data features on which to base its decisions, the specified It turned out to be more accurate. Human-selected features are sensitive to device variation, whereas the algorithm analyzed all the data at once to find the least variable elements. “Machine learning can pick out more nuanced differences that are difficult for humans to define,” Das explains. The system managed more than 97% accuracy in real-world tasks.
“The data are very convincing,” says Kiana Allan, an engineer at the University of California, San Diego, who co-founded a company commercializing graphene-based biosensors. Unlike the human tongue, which detects specific molecules, this type of ISFET system only detects chemicals. change— is “limited to a specific, pre-defined chemical profile,” such as brand formulation and freshness range, she says.
Next, Das and his colleagues plan to test larger, more diverse training datasets and more complex algorithms while expanding the scope of the system. For example, “this technology can be used for healthcare applications such as blood sugar and sweat monitoring,” says Das. “That would be another area we would like to explore.”