A single semiconductor packs in memory, light sensing and learning [Unboxing Lab]

[Financial News] A team led by Professor Jeong Mun-seok of Hanyang University has developed a semiconductor device that stores information, detects light and even uses light for learning. The team confirmed that the device’s information-storage range reaches a maximum of 58.5 V, and that it can also respond to light by sending and receiving signals like nerve cells. The team said the findings point to the device’s potential use in artificial intelligence (AI) semiconductors.

This research could help in developing semiconductors that reduce the power needed for AI computations. Generally, memory stores information, while processing units process it. The device developed by the team stores information while also changing its state in response to light and electrical signals. It could be used to develop semiconductors that combine information storage and processing in a single device.
Neuromorphic semiconductors are a field of research focused on developing this kind of semiconductor. They are designed to mimic how the human brain remembers and learns information. In the brain, synapses connecting nerve cells change the strength of their connections in response to stimuli. The team created a similar operating principle by using light and electrical signals to change the semiconductor device’s state in multiple steps.
The device’s light-responsive property could also be used to develop semiconductors that record information with light or detect light. However, this study did not produce an actual AI product or measure how much power was saved. It is a step toward implementing multiple functions in one device and confirming its potential for future use in other semiconductor technologies.

The team used amorphous tellurium oxide (a-TeO3) as the semiconductor material. Tellurium is a chemical element. Amorphous means that the atoms are not arranged in a regular pattern.
The team exposed thin tellurium to ultraviolet light and ozone, triggering a reaction with oxygen. They used the tellurium oxide produced through this process as the channel—the path through which current flows—to make a semiconductor device. The device is a p-type oxide semiconductor.
In semiconductors, current is carried by electrons or holes. A semiconductor in which electrons carry the current is called n-type, while one in which holes carry the current is called p-type. A hole is a concept that treats the vacant space left by an electron as if it were a particle.
Research on oxide semiconductors has so far focused on n-type materials. Research on p-type oxide semiconductors has been limited because they are relatively difficult to make. The team fabricated a p-type semiconductor device using material made by oxidizing tellurium, and confirmed that the device both stores information and responds to light.

The team first tested the device’s memory function by applying electrical signals. The results showed that its memory operating window reached a maximum of 58.5 V. The memory operating window indicates how distinctly the device exhibits different states when the voltage is changed.
When the device’s states are clearly distinguishable, stored information is easier to read. Dividing the states into multiple levels can also allow a single device to store information at several levels. The team confirmed that this kind of information storage was possible with the device.
The team also examined how long the device’s state lasted after information was recorded. In the basic experiment, different resistance states were maintained for at least 200 seconds. In an additional experiment, the recorded state was observed for up to about 10,000 seconds.
The team also tested repeatedly recording and erasing information. In the basic experiment, different states remained distinguishable over more than 250 repeated cycles. A separate additional experiment went up to 1,000 cycles. This means the device retained its memory function through repeated recording and erasing.
The team also measured how well the device detected light. Under 390 nm ultraviolet light, its responsivity was 411 A/W and its detectivity was 2.6×10¹¹ Jones. Responsivity indicates how strongly an electrical signal is produced when light hits the device. Detectivity indicates its ability to detect weak light.
The team also observed that current did not disappear immediately after the light was turned off, but continued for a while. This is because when light shines on the device, charge remains inside it and affects the state in which current flows. This property could be used to change the device’s state or record information with light.

The team also tested the device’s artificial synapse function using light. The device’s state changed when exposed to ultraviolet light or when an electrical signal was applied. Changing the conditions of the stimulus caused the level of current flow to increase or decrease in multiple steps. Conductance is a measure of how easily current flows.
These changes resemble the way synapses in the brain adjust the strength of their connections in response to stimuli. By changing the device’s state with light and electrical signals, the team confirmed that an artificial synapse could express changes needed for learning.
The team also examined whether the device could be used in AI. It built an AI model to classify handwritten digit images from 0 to 9. The model connected 484 input neurons, 100 hidden neurons and 10 output neurons. The input neurons receive digit images, and the hidden neurons process the input information. The output neurons indicate the result of classifying which digit an image shows.
The team created an AI model that reflected the operating characteristics of the actual device. It then compared the learning results with those of a digital model that assumed the device worked perfectly. The two models learned to distinguish handwritten digits in a similar way. The results showed that digit-classification learning was possible even in a model reflecting the actual device’s characteristics.
The team also combined the device with another semiconductor device to create a circuit that mimicked the brain’s principle of body temperature regulation. It used the p-type tellurium oxide device developed by the Hanyang University team together with an n-type molybdenum disulfide device. The circuit was configured so the two devices operated in different ways, with one producing a response corresponding to cooling and the other to heating.
The paper used conditions of 26°C and 20°C to simulate situations requiring cooling and heating, respectively. The higher-temperature condition produced a response corresponding to cooling, while the lower-temperature condition produced one corresponding to heating. This was not an experiment regulating an actual person’s body temperature, but one that used an electronic circuit to mimic how the brain responds to changes in temperature.
This research implemented information storage, light detection and artificial synapse functions in a single p-type oxide semiconductor device. The team also explored its potential applications through AI digit-classification experiments and a circuit experiment simulating the principles of heating and cooling.
Meanwhile, the findings were published in Volume 22, Issue 53 of Small, an international journal in the fields of nanoscience and materials, and were selected as a cover paper. 

[email protected] Kim Man-gi Reporter