2020 Fall Physics Colloquium
강연일자: 2020. 11. 13. 화. 오후 5시
장소: 비대면 온라인 Zoom
강연자: 한명준 (KAIST)
강연주제: Toward the Better First-principles Description of Correlated Materials:
Application of Modern Machine Learning Techniques
‘First-principles’ description of correlated electron materials has long
been a great challenge in the electronic structure calculation community.
Although there are several outstanding methods available thanks to the great
effort and progress in the past decades, it still remains as an open problem.
In the first part of my talk, I will try to give a brief overview of technical
and methodological issues which can hopefully serve also as a practical
guideline for users of these methods. In particular, I will mostly focus on so-called
LDA+U (local density approximation plus Hubbard U) and DFT+DMFT (density
functional theory plus dynamical mean-field theory) both of which are widely
used and share the same physical idea. For the case of latter, quite different
types of difficulty often arise, and it is the topic of the second part. In the
second part, I will report our recent efforts to deal with ‘analytic
continuation’. This fundamental problem in many-body physics corresponds to a
notorious inverse problem in numerics. By adopting modern machine learning
techniques such as CNN (convolutional neural network) and generative models, we
were successfully generating ‘spectral function (i.e., the electronic structure
of correlated materials)’ from imaginary Green’s function. Our machine
learning-based approaches show the better performance in terms of accuracy and
computation speed compared to the conventional maximum entropy and stochastic
method.