Abstract:
A research team from the University of Würzburg in Germany has developed an oxide interface electronic device that can switch between three modes: transistor, memristor and memcapacitor according to the circuit connection method. The research, published in Nature Communications, demonstrates the possibility of integrating computing, memory and signal processing into the same material platform, providing new experimental solutions for brain-inspired computing hardware.
The device is composed of lanthanum aluminum oxide (LaAlO₃) and strontium titanate (SrTiO₃). Both materials are insulators individually, but the contact interface can form an extremely thin layer of quasi-two-dimensional electron gas, allowing electrons to move along the interface. The researchers controlled this conductive channel through lateral gates to create a nanowire structure; the conductive channel of the experimental device was about 100 nanometers wide, and the distance between the channel and the gate was about 400 nanometers.
Under different wiring and voltage conditions, the same device can behave as a transistor, memristor, or memcapacitor. Transistors regulate electrical current; memristors retain electrical signal history through changes in resistance; and memcapacitors allow capacitance states to change in response to previous inputs. Such a "memory" effect allows the circuit to respond differently to the sequence of consecutive pulses, and the research team built a reserve pool calculation circuit accordingly.
The researchers used a 5×4 pixel black and white image to demonstrate the number recognition from 0 to 9, converted each line of the image into pulse input, and then used the device output to form different state features, and the subsequent classification layer judged the number. The test used a reserve pool array of devices and trained the output layer; this is a hardware demonstration of a specific task and does not mean that a single component has general artificial intelligence capabilities. The paper also demonstrates a circuit composed of transistors and memristors that can perform AND and OR logic and store the results of the operation inside the circuit.
Under specific test conditions of the digital recognition circuit, the research team measured that each input pulse consumes approximately 2.71 nanojoules; using a control reserve circuit composed of transistors and memristors, it consumes approximately 16.57 nanojoules under similar working conditions. This result reflects the circuit configuration tested in the paper and cannot be directly equated to the energy consumption of a complete AI system or an actual product.
Researchers believe that such reconfigurable devices may help reduce the need to transmit data back and forth between the processor and memory in traditional computing in the future. They may also be used in wearable devices to locally process physiological signals such as heart rate, blood pressure, body temperature, or blood sugar. However, the current results are still laboratory materials and circuit prototypes; large-scale manufacturing, stability, integration with existing chip processes, as well as energy consumption and performance in real applications still need to be further verified.
Comments