TY - JOUR
T1 - Optimization of Work Function via Bayesian Machine Learning Combined with First-Principles Calculation
AU - Hashimoto, Wataru
AU - Tsuji, Yuta
AU - Yoshizawa, Kazunari
N1 - Funding Information:
This work was supported by KAKENHI Grants (Numbers JP17K14440 and JP17H03117) from the Japan Society for the Promotion of Science (JSPS) and the Ministry of Education, Culture, Sports, Science and Technology of Japan (MEXT) through the MEXT projects Integrated Research Consortium on Chemical Sciences, Cooperative Research Program of Network Joint Research Center for Materials and Devices and Elements Strategy Initiative to Form Core Research Center, and by JST-CREST JPMJCR15P5 and JST-Mirai JPMJMI18A2. The computations in this work were primarily performed using the computer facilities at the Research Institute for Information Technology, Kyushu University. Y.T. is grateful for a JSPS Grant-in-Aid for Scientific Research on Innovative Areas (Discrete Geometric Analysis for Materials Design, Grant Number JP18H04488 and Mixed Anion, Grant Number JP19H04700).
Publisher Copyright:
© 2020 American Chemical Society.
PY - 2020/5/7
Y1 - 2020/5/7
N2 - Work function is one of the most fundamental and important physical quantities in surface science. Materials with either lower work function or higher work function would find various applications, such as electronic devices and high-performance catalysts. However, it would be challenging to find a material with the optimal work function exploiting the all-search or random approach, whether it is based on an experimental or theoretical method. In this paper, we use the Bayesian optimization (BO) approach, which is one of the most powerful machine-learning tools for optimization, in order to effectively explore a candidate material with a higher or lower work function value out of hundreds of thousands of materials registered in a material database. We introduce a quick measure of the work function based on the depth of the Fermi level calculated from the first-principles computation for the crystalline bulk structure of a material. We call this the approximate work function, treating it as the objective function of our BO scheme. Since we do not need any time-consuming surface calculation with the slab model for the evaluation of the approximate work function, a quick search of a material with the highest or the lowest work function is achieved. As input variables for our BO implementation, we employ some bulk-specific properties of materials, which can be fetched from the database. The demonstration of our BO-based exploration of the database shows that materials with both low and high limits of the approximate work function can be discovered more efficiently in BO than a random exploration. The top 10 lowest work function materials thus found are in line with our chemical intuition in that all of them include either alkali or alkaline earth metal. On the other hand, we found the top 10 highest work function materials with amazement because they also include either alkali or alkaline earth metal and a lanthanide element.
AB - Work function is one of the most fundamental and important physical quantities in surface science. Materials with either lower work function or higher work function would find various applications, such as electronic devices and high-performance catalysts. However, it would be challenging to find a material with the optimal work function exploiting the all-search or random approach, whether it is based on an experimental or theoretical method. In this paper, we use the Bayesian optimization (BO) approach, which is one of the most powerful machine-learning tools for optimization, in order to effectively explore a candidate material with a higher or lower work function value out of hundreds of thousands of materials registered in a material database. We introduce a quick measure of the work function based on the depth of the Fermi level calculated from the first-principles computation for the crystalline bulk structure of a material. We call this the approximate work function, treating it as the objective function of our BO scheme. Since we do not need any time-consuming surface calculation with the slab model for the evaluation of the approximate work function, a quick search of a material with the highest or the lowest work function is achieved. As input variables for our BO implementation, we employ some bulk-specific properties of materials, which can be fetched from the database. The demonstration of our BO-based exploration of the database shows that materials with both low and high limits of the approximate work function can be discovered more efficiently in BO than a random exploration. The top 10 lowest work function materials thus found are in line with our chemical intuition in that all of them include either alkali or alkaline earth metal. On the other hand, we found the top 10 highest work function materials with amazement because they also include either alkali or alkaline earth metal and a lanthanide element.
UR - http://www.scopus.com/inward/record.url?scp=85084940760&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85084940760&partnerID=8YFLogxK
U2 - 10.1021/acs.jpcc.0c01106
DO - 10.1021/acs.jpcc.0c01106
M3 - Article
AN - SCOPUS:85084940760
SN - 1932-7447
VL - 124
SP - 9958
EP - 9970
JO - Journal of Physical Chemistry C
JF - Journal of Physical Chemistry C
IS - 18
ER -