Semiconductor technology is advancing toward faster, smaller, and more energy-efficient devices. Among the materials attracting growing attention, amorphous hafnium oxide (a-HfO_x) is a key insulating material in semiconductor devices. Under strong electric fields, its atomic and electronic behaviors can change significantly, directly affecting device speed, lifetime, and stability.

Conventionally, ab-initio molecular dynamics simulations have been used to study atomic-scale behavior with high accuracy. However, these methods are computationally expensive and time-consuming. Classical molecular dynamics or empirical potential-based methods offer faster calculations, but their predictive accuracy is limited.

To overcome this limitation, a research team led by Professor Donghwa LEE (MSE, AMS) has developed a new AI model that integrates charge equilibration method with a graph neural network technique. The model learns interatomic interactions and simultaneously predicts atomic charges, energies, and forces, enabling the implementation of a machine learning potential for a-HfO_x systems. Notably, even though the model was trained only on data obtained without an applied electric field, it successfully reproduced the changes in atomic motion and charge behavior that occur under electric-field conditions.

Using this technology, the team was able to predict oxygen-ion migration tendencies within a-HfO_x as well as the dielectric breakdown voltage of semiconductor devices. Prof. Lee stated, “In the past, predicting material responses under electric fields required costly additional datasets generated under field-applied conditions. Our approach achieves high accuracy without the need for separate electric-field data.”

The study was selected for the special collection in npj Computational Materials.