Chemosensors are essential in many applications. Since 1970, ion-sensitive field-effect transistors, or ISFETs, have been used. They work by converting changes in the composition of chemical solutions into electrical signals. The ions in a liquid touch the conductive layer of an ISFET, changing the electrical current flowing through it based on the exact composition of the liquid and the applied voltage. This allows for testing and monitoring the pH level of a solution, or even the taste and freshness of a beverage like wine or juice.
ISFETs can be thought of as electronic tongues. A team at Penn State University has linked this technology with machine learning to enable new applications. Their study, reviewed by independent experts, was published in the journal Nature.
Although ISFETs have been in use for almost 60 years and have improved over time with materials like graphene, they have some drawbacks. The reliability of ISFETs is affected by various factors, including cycle-to-cycle fluctuations, sensor-to-sensor differences, and chip-to-chip differences. These result from manufacturing processes, material properties, environmental conditions, and design decisions.
To address these issues, the team led by engineer and materials scientist Saptarshi Das combined modern graphene ISFETs with neural networks. They first trained a machine learning algorithm to classify various beverages based on sensor readings. These included coffee, milk, cola, fruit juices, and wine, each in different varieties or states. For juices, measurements were taken over four consecutive days and several hundred times throughout the day.
This approach allowed the algorithm to learn how the chemical composition of the beverage changed over time. Based on this data, the system could then determine if milk was diluted and distinguish between different types of cola, coffee, and fruit juices. It did so with impressive accuracy, achieving 97% accuracy in identifying the type of juice and how long it had been open.
“We found that the network pays attention to subtler features in the measurement data—things that are difficult for humans to define accurately,” says study leader Das. In other words, the ISFETs detected changes in the beverage composition that might not be noticeable to the human tongue. Since the trained model incorporates data from many different sensors into the analysis, it can also bridge the differences between individual sensors mentioned earlier.
The researchers believe that ion-sensitive field-effect transistors could become more appealing to the food industry. “Timely detection of harmful contaminants in food production is a constant challenge,” they write in the study. Using ISFETs in combination with neural networks could offer a cost-effective alternative to current testing methods. Next, Saptarshi Das and his colleagues plan to expand the training datasets to include more products and liquids.