■ 제 목: Data-driven material research based on first-principles calculations
■ 연 사: 한승우 교수 (서울대학교)
■ 일 시: 4월 23일(화) 16:00
■ 장 소: 응용공학동 (W1) 1층 영상강의실
■ Host : 김경민 교수
■ Abstract :
Data-driven research is receiving attentions in every engineering field. In material engineering, however, it takes strenuous efforts to produce reliable data due to non-equilibrium process conditions. Recently, with the development in high-speed computation and first-principles methods, it became possible to produce data on the scale of 104-106. Here I would like to highlight two activities in my laboratory related to the data-driven material research. First, by carrying out high-throughput first-principles calculations, we screen promising p-type oxides from ~18,000 oxides that are essentially all the oxides identified to date by experiment. In particular, we use the hydrogen descriptor that can sense the valence band position efficiently. We find several promising oxides such as La2O2Te and CuLiO that are expected to outperform the well-known p-type oxides such as SnO and CuAlO3. From the pool of screened p-type oxides, we reveal the chemical principle underlying the high-performance p-type oxides, which are consistent with the empirical knowledge.
Second, I will introduce using the machine-learning (ML) approach to develop interatomic potentials from the massive first-principles calculations, which can be used in large-scale classical molecular dynamics (MD) simulations in various applications. As a learning tool, we employ the artificial neural network. We propose an efficient weighting method to increase the accuracy and reliability of the neural network potential by rectifying the sampling bias. Several simulations will be presented using the neural network potential such silicidation process and phase-change materials.