KIST-Kyung Hee University Research Team, Next-Generation Semiconductor Nanomagnet
Development of a State-Defining Energy-Minimizing Variant Autoencoder
Presenting a solution to optimization problems using machine learning technology


An optimization problem refers to the problem of finding the most suitable solution to achieve a specific objective. While simple problems can be solved by examining all possible cases, complex problems involve an immeasurable number of possibilities and are therefore treated as representative difficult problems across various research fields.

Optimization problems also emerge in next-generation semiconductor research. A prime example is the field of spintronics, which develops low-power, high-performance semiconductors to overcome the integration limitations of silicon semiconductors. If the most stable state of the nanomagnet material is not identified and its properties are not understood in detail, the precise operating characteristics and range of the spin device cannot be designed.

A research team led by Dr. Heeyoung Kwon and Dr. Junwoo Choi of the Spin Fusion Research Group at the Korea Institute of Science and Technology (KIST) and Professor Changyeon Won of Kyung Hee University announced the development of the 'Energy-minimization Variational Autoencoder (E-VAE),' a generative machine learning model that estimates the spin structure appearing in the ground state, the most stable state of nanomagnets.

Generative machine learning techniques are used to learn from given data, extract its characteristics, and recombine them to generate new data. The research team confirmed that applying existing generative machine learning models to nanomagnets results in local noise and blurring effects, as well as the creation of conditions that violate the laws of physics.

In existing models, there was no process to lower the energy of the newly generated state compared to the input state, making it difficult to use for exploring the ground state of nanomagnets. The research team developed an E-VAE model that includes a process to minimize the energies of states generated in a Variational Autoencoder (VAE), an existing model, and succeeded in efficiently finding the optimal state that the spin structure of a nanomagnet can possess.

This demonstrated high efficiency and accuracy in finding the optimal state compared to the existing Simulated Annealing (SA) technique.

Dr. Hee-Young Kwon of KIST stated, “Optimization problems are important research topics not only in pure science and semiconductor research but also in the fields of mathematics and computer science,” adding, “The E-VAE model is expected to hold high academic value across various fields.” This research was conducted as part of a major KIST project supported by the Ministry of Science and ICT and a program to foster the next generation of scholars supported by the Ministry of Education. The research results were published in the June issue of the international journal Advanced Science.