기계학습으로 낮은 열팽창계수를 가지는 초기능성 합금 발견
Direct composition-to-property prediction for materials design remains a challenge for material scientists.
thermal expansion coefficient (TEC)
two-stage ensemble regression model (TERM)
two-stage ensemble regression model (TERM)
generative alloy design (a generator for suggesting new alloy compositions)
TEC가 아주 낮은 합금 찾기 성공
인공지능 모델, 제일원리계산, 시뮬레이션, 데이터 등을 종합적으로 활용함.
제일원리 계산, 생성모델 + 회귀모델, 실험을 모두 활용함.
The authors first trained the learning algorithm using published data of ~700 alloys and then tasked the algorithm to generate a large number of candidate compositions with low thermal coefficients.
The top 1000 candidates among the generated compositions were selected by using another algorithm that takes into account atomic features such as the valence electron concentrations and atomic radius of the elements.
The selected 1000 compositions were further narrowed down to 20 to 30 by using another learning algorithm with additional inputs about physical properties from density functional theory calculations and thermodynamic databases.
Last, the top three candidates were physically created in the laboratory, then their properties were measured and fed back to the experimental dataset. In doing so, the incorporation of the physical properties as well as the feedback of experimental data improve prediction accuracy and efficiency.
By repeating this procedure six times, Rao et al. discovered 17 Invar HEAs out of millions of possible compositions (with each of the elements ranging from 0 to 100% with 1% resolution), and more Invar HEAs could be identified with subsequent cycles.
---------------------------------------------------------------------------------------------------------------------
CNN 결과를 fitness function으로 활용함.
covariance matrix vs fitness function
---------------------------------------------------------------------------------------------------------------------
덧글