The idea being that as for the human sense of taste, a sensing device could rely on the ability of few individual sensors to respond simultaneously to different chemicals (combinatorial sensing) to get a holistic sensing pattern or a global fingerprint of the liquid being tested (or tasted).
Dubbed Hypertaste, the small lime-slice shaped AI-assisted e-tongue packs an array of electropolymerized ion-sensitive films with a microcontroller for data acquisition. When the Hypertaste device is clipped on the side of a glass, the sensor’s electrodes dip into the liquid to be tasted and a series of differential voltages can be recorded, collectively yielding a unique signature – the liquid’s fingerprint – to be analyzed and classified on the cloud by a trained AI algorithm. That digital fingerprint can then be compared to a database of known liquids, and the algorithm figures out which liquids in the database are most chemically similar to the liquid being tasted.
The new approach was presented at the 2019 ISOCS/IEEE International Symposium on Olfaction and Electronic Nose (ISOEN) in a paper titled “A portable potentiometric electronic tongue leveraging smartphone and cloud platforms”. The researchers report that for trained systems, inferencing tasks such as the classification of liquids are realized within less than one minute including data acquisition at the edge and inference using a cloud-deployed machine learning model.
IBM Research believes the low-cost Hypertaste could serve a wide range of industrial and scientific users be it for on-the-fly water quality checks or enabling beverage producers to identify counterfeit products or check raw materials. The quick, in-situ identification and classification of liquids would also be relevant to the pharmaceutical and healthcare industries.
Here, one big advantage of having the machine learning models running on the cloud is that the sensors could be rapidly reconfigured from anywhere without changing the hardware. Only a few changes of parameters in the machine learning models would make them adjust to the new sensor set. Crowdsourcing sensing data through field-deployed connected sensors would further reinforce the learning.
As emphasized by the researchers, fooling a combinatorial sensing system such as Hypertaste is much harder than fooling individual analyte-specific lab tests, as there is no single substance on which the identification relies.