|
Type |
Symposium |
Area |
Data-Enabled Computational Chemistry |
Room No. |
Room 324A |
Time |
FRI 10:15-: |
Code |
PHYS2-4 |
Subject |
Design of Novel Catalysts: From First-Principles to Machine-Learning |
Authors |
Sang Soo Han Computational Science Research Center, Korea Institute of Science and Technology, Korea |
Abstract |
Electronic structures of materials such as electron densities of states (DOSs) are one of the most critical keys to determine material properties. Thus, prediction of the electronic structure using a first-principles density-functional theory (DFT) calculation has been regarded as a very useful tool for design of novel catalysts. Recently, a combination of the DFT calculation and the automation technique provides a high-throughput screening for a catalysis design. With this method, we have recently developed novel metallic catalysts for direct synthesis of hydrogen peroxide (H2O2), which shows a superior catalytic properties over the state of the art material, palladium (Pd). Moreover, the high-throughput screening technique is also efficient to build the DOS database, with which we can apply a machine-learning technique for design of novel catalysts. In this talk, we will also discuss the relevant recent advances.
On the other hand, although the DFT calculation provides an accurate DOS for a material, the calculation is very time-consuming. Herein, we will discuss a cost-effective method to predict the DOS of multi-component ally systems based on a machine-learning algorithm called a principal component analysis, with which the shape of DOS can be even predictable. Within this framework, we input only a crystal structure and a composition. In comparison with the DFT calculation (GGA level), our machine learning method can provide the DOS results (both of value and shape) with an accuracy of >95% and a 1,000 times faster speed than the DFT calculation. To our best knowledge, this work is the first machine learning approach to predict the complete DOS information (value and shape) of multi-component alloys. In addition, we will introduce a neural network model for predicting binding energies of adsorbates on catalyst surfaces from the DOS. And, we have been recently developed a crystal graph convolutional neural network model for description of surface structures, which provides a novel way to predict such binding energies with no first-principles calculation information.
|
E-mail |
sangsoo@kist.re.kr |
|