Materials simulation can produce various materials data difficult to obtain experimentally.
Many institutions and companies are currently working on developing unknown materials through machine learning techniques, based on their accumulated materials data.
However, to design new materials or solve problems of materials using machine learning and materials data, two conditions should be met.
First, The method of yielding the data should be consistent.
Secondly, The correlation between the data needs to be verified clearly.
In fact, there are not many institutions or companies worldwide that meet those both conditions.
Material simulation can be an alternative to meet these conditions.
A. If the first condition is satisfied but the second condition is not
Most of research institutions or companies with a long history and tradition reserve data set meeting the first condition.
But it is difficult for the data to have a clear correlation between the data, the second condition.
Material simulation can be an advantageous approach to overcome this hindrance.
In general, the easiest machine learning model is to predict creep lifetime from alloying elements.
However, since most machine learning techniques is an interpolation, creep lifetime can be mispredicted if data on unlearned alloy elements or compositions are entered.
As an alternative, this institution predicted the creep lifetime from microstructural information obtained using thermodynamic calculations.
This institution showed that the correlation between alloying elements and creep lifetime was not very good, but the correlation between phases and creep lifetime was good.
Therefore, it allows a very useful approach that predicting the creep lifetime from phase. Since the microstructural information on phases of unknown alloy elements and compositions can be obtained using thermodynamic calculations.
Particularly, If the correlation between the data is very clear, you do not need much data when developing a machine learning model.
Shin, D. et al. 2019. “Modern Data Analytics Approach to Predict Creep of High-Temperature Alloys.” Acta Materialia 168: 321–30.
B. If both conditions are not satisfied
Most materials data have been difficult to meet both criteria yet.
Therefore, as an alternative, appropriately using materials simulations is required.
For example, in the case of Pt catalyst, the density of state (DOS) which is the information for electronic band structure is associated with the properties of Pt catalysts.
Accordingly, one institution studied to find some materials that have similar DOS to DOS of Pt catalyst.
DOS can be obtained using the first-principles calculation, but this also takes time to get DOS data.
To save time, this institution developed a machine learning model that predicts the DOS from the information on the crystal structure. Here, the training data on the DOS were obtained using the first-principles calculation.
In practice, this institution found alternative materials to Pt catalyst using this machine learning model.
Kim, Myungjoon, Byung Chul Yeo, Sang Soo Han, and Donghun Kim. 2018. “Slab Graph Convolutional Neural Network for Discovery of N2 Electroreduction Catalysts.” cond-mat.mtrl-sci.
Yeo, Byung Chul et al. 2017. “Atomistic Simulation Protocol for Improved Design of Si–O–C Hybrid Nanostructures as Li-Ion Battery Anodes: ReaxFF Reactive Force Field.” The Journal of Physical Chemistry C 121(42): 23268–75.
Materials simulation is an essential technique in the materials data science.
Like the previous two examples, materials simulation is considerably utilized in the data science.
The key to the Fourth Industrial Revolution is materials. In order to utilize various technologies, relevant materials to realize technologies are required.
Materials simulation can be an essential technique to develop a variety of materials faster.
Young-Kwang Kim, Ph.d.
Sign up for free and get $10.
This post is licensed under a Creative Commons Attribution-Nonprofit-ShareAlike 4.0 International License.