Hanyang University, Sungkyunkwan University, and Gachon University's research team strengthened synaptic device performance through the laser process

 

Professor Park Hui-joon of Hanyang University's Department of Organic and Nano Engineering has developed a laser process that can maximize the performance of artificial intelligence semiconductor synapse devices with Professor Ko Jong-hwan of Sungkyunkwan University and Professor Lee Dae-ho of Gachon University, Hanyang University said on April 6.

 

With the acceleration of science and technology development, there is an increasing need for computing technology that can efficiently and quickly process large amounts of unstructured data. However, in the current computer structure (Von Neumann architecture), where memory and processor are physically separated, bottlenecks occur when processing large amounts of data, resulting in severe performance degradation and energy consumption.

 

The neuromorphic computing system can overcome the shortages of the Von Neumann architecture by minimizing processing time and power consumption of large amounts of unstructured data by simultaneously storing and calculating data in an artificial neural network without data transmission between memory and processors.

 

In general, the multi-level resistive memory cell has been widely used to implement synaptic devices that make up artificial neural networks, but it has been difficult to secure the reliability of synaptic devices due to uneven formation and rupture behavior of metal filaments that induce changes in resistance.

 

In this study, the joint research team introduced a new laser process to implement a metal oxide thin film with a uniform three-dimensional ion transfer path in the nanometer range, and when used as a switching medium, the formation and rupture behavior of conductive filaments are optimized to maximize the performance and reliability of synaptic devices.

 

In particular, it was possible to maximize the ability to precisely control the weight of synapses, which is a decisive factor that determines the accuracy of learning and reasoning in a neuromorphic computing system, and to maximize the repetition characteristics of devices, uniformity of performance between devices, and durability.

 

When it became possible to produce synaptic devices with high reliability, the joint research team was able to propose a way to easily learn neuromorphic hardware and verified its effectiveness by confirming that it can secure high accuracy of 97% or more when used in ultrasonic signal-based biological tissue classification.

 

The laser process developed by the joint research team is not only compatible with the existing CMOS process but can also be applied on a flexible board, so it can be used as a base technology for various flexible electronics.

 

The researchers expressed expectations, saying, "It can be used as a key technology for next-generation PIM (processing-in-memory) artificial intelligence semiconductor devices that can overcome the computational limitations of existing computers, and it will be possible to pioneer new industries through the implementation of a neuromorphic computing system that can store and process data at the same time."

 

 

This study, supported by the PIM Artificial Intelligence Semiconductor Core Technology Development (Device) Project and Basic Laboratory Support Project of the Korea Research Foundation of the Ministry of Science and ICT, was published in the “Advanced Functional Materials”, a world-renowned journal in material Science and was selected as the Front Cover Article in recognition of its excellence.

 

[Photo 1] A joint research team of Hanyang University, Sungkyunkwan University, and Gachon University. (from left) Professor Park Hui-joon of Hanyang University, Professor Ko Jong-hwan of Sungkyunkwan University, Professor Lee Dae-ho of Gachon University,  researher Kim Do-hyung of Hanyang University

 

[Photo 2] Schematic diagram of PIM artificial neural network based on 3D ion channel memristor crossbar array implemented by the joint research team

 

[Photo 3] Schematic diagram of PIM artificial neural network based on 3D ion channel memristor crossbar array implemented by the joint research team