A group of Korean researchers from KAIST, UNIST And POSTECH has created a system of transistors stacked vertically which processes together the most recent signals and those received shortly before. The system could allow small wearable devices to analyze movements and physiological parameters locally, reducing the need for external computing resources. The study was published in the journal Advanced Materials.
Transistors use carbon nanotubes controlled by one Solid state ionogela material that contains mobile ions capable of influencing the passage of current. When a voltage is applied the ions move in the material, and when the voltage is removed they do not immediately return to the starting position: thus a temporary electrical memory of the previous signals. The researchers treated this delay as a resource for processing time-varying information, rather than a limitation of the device.
By adjusting the concentration of ions and it thickness of the ionogelthe team obtained devices with different response speeds, then stacked them on top of each other. The fast layer reacts mainly to the latest input, while the slow layer retains the accumulated information over a longer interval. By combining the two outputs, the system has both what just happened and what happened immediately before.
The ionic materials used so far in devices of this type have often been soft liquids or gels, poorly suited to precision manufacturing and multi-layer stacking. Bring the ionogel to solid thin film it allowed us to better control its properties during the construction of the different layers. The devices were produced both on 4 inch wafer be on flexible substratessuitable for electronics that must adhere to non-flat surfaces, and according to the researchers they have maintained stable electrical characteristics for 55 months from manufacturing.
In the hardware test, the system received four consecutive signals, each on or off, and distinguished correctly all 16 combinations possible. With the measurements collected on the devices, the researchers then built a simulation to classify videos of moving handwritten digits played at different speeds, achieving validation accuracy above 90%. The data on the video sequences therefore comes from a model based on the real devices, and not from a complete chip.
“These devices can be fabricated on large-area substrates with current thin-film semiconductor processes, and can also be stacked in multiple layers,” he explained Jimin Kwonprofessor at KAIST. According to Kwon, the technology could lead to chips that analyze movement and physiological signals in low-power devices like the smartwatches. However, the energy saving has not yet been demonstrated on a real smartwatch: further tests will be needed to understand how the architecture behaves, and how much energy it saves, once integrated into a wearable product.
