The search for an automated way to “taste-test” products at mass-production speed and scale has stumped the food and beverage industry for decades. But in a new study, researchers used machine learning to overcome the limitations of a promising type of chemical sensor, meaning that a robotic tongue may soon assess your milk or merlot before you do.
When ions in a liquid—say, a delicious drink—touch the conductive sheet of an ion-sensitive field-effect transistor (ISFET), the electric current that flows through changes based on the liquid’s exact composition and the voltage applied. This lets scientists use ISFETs to convert chemical changes into electrical signals. The chemical makeup of any drink, and thus its taste, is influenced by contamination and freshness—which ISFETs can discern.
“The food industry has a lot of problems in terms of figuring out whether food is adulterated or has something toxic in it,” says Pennsylvania State University engineer Saptarshi Das. The first ISFETs were demonstrated more than 50 years ago, but the sensors aren’t used much commercially. The advent of graphene, an ideal conductive material, helped researchers create improved ISFET sensors that detect specific chemical ions. But a big problem remained: readings varied from sensor to sensor and with changes in conditions such as temperature or humidity.
On supporting science journalism
If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
In Nature, Das and his colleagues addressed this issue by marrying ISFETs with neural networks, training a machine-learning algorithm to classify drinks using the sensors’ readings. The resulting system could tell whether milk was diluted, distinguish among soda brands or coffee blends, and identify different fruit juices while judging their freshness.
During development the team tried training based on human-selected data points, but the scientists found that designations were more accurate if the algorithm was given all device measurements and chose its own data features to base decisions on. Human-chosen features were vulnerable to variations in the devices, whereas the algorithm analyzed all the data at once, finding elements that change less. “Machine learning is able to figure out more subtle differences” that humans would find hard to define, Das explains. The system managed more than 97 percent accuracy on practical tasks.
“The data are very convincing,” says University of California, San Diego, engineer Kiana Aran, who co-founded a company to commercialize graphene-based biosensors. Unlike the human tongue, which detects specific molecules, this type of ISFET system detects only chemical changes—“which limits it to specific, predefined chemical profiles” such as brand formulations or ranges of freshness, she says.
Next, Das and his colleagues will test larger, more diverse training datasets and more complex algorithms, as they expand the system’s reach. For example, “you can use this technology for health-care applications: blood glucose level or sweat monitoring,” Das says. “That’s going to be another area we want to explore.”