Searching Lightweight and High-Stiffness Glasses by Machine Learning

Liang Qi's group has developed a new machine learning model that significantly boosts the search for lighter and stronger glasses with many promising engineering applications.
Searching Lightweight and High-Stiffness Glasses by Machine Learning

An efficient and reliable machine learning method has been developed by Assistant Professor Liang Qi's group that can predict the properties of silica-based glass materials. This method can output the densities and elastic stiffness for ~100,000 of glass compositions in several hours by just using a regular personal computer. Elastic stiffness, also known as elastic modulus, is a measure of the resistance offered by a material to fully recoverable mechanical deformation. This method significantly boosts the search for lighter and stronger glasses with many promising engineering applications. 

Glasses are among the most common materials that we see and touch every day, especially given the ubiquity of touch-screen devices in our culture. Besides the “visible” functions, glasses have many other “invisible” applications. For example, composite materials strengthened by high-stiffness glasses can make wind turbine blades generate more electricity and automobiles more lightweight for better fuel efficiency. Glasses are also ancient materials; people started to produce them about five thousand years ago. Despite these facts, people still have a limited ability to understand and predict their properties and performances.

The key difficulties originate from the atomistic structures of glasses. Glasses are amorphous materials, which means all atoms inside glasses are not located in ordered positions like those in metals and ceramics. Because of the disordered arrangements, it is usually difficult to describe their atomistic structures and the corresponding physical and chemical properties.

Computers simulations can be applied to speed up the study of glasses. However, in conventional computer simulation methods like molecular dynamics (MD),computers need to calculate and record the position, velocity, and force of each atom inside the glasses. It requires a considerable amount of time and computing resources to study the properties of just one glass composition. Meanwhile, a common commercial glass can be made of five to ten types of elementary compounds, such as silica, soda, lime, plus several additive oxides. Thus, it is impossible to apply these simulations to investigate each possible glass composition.

In the paper recently published in npj Computational Materials from Qi’s group, the research team presents a machine learning model to predict the densities and elastic stiffness of silica-based glasses. Running on a laptop, this model can immediately output these properties of any glass composition across a complex compositional space with multiple (>10) types of additive oxides.

To apply machine learning for predictions of material properties is different than other typical machine learning applications, such as facial recognition and autonomous driving. A “machine” (the computer algorithm) “learns” once it is “trained” using a large amount (maybe millions or more) of data related to the prediction target. However, data for materials research are usually not sufficient due to the expensive experimental or computational costs. Machine learning based on the limited data maybe not reliable for targeted materials that are quite different than those in training data, just like a political poll conducted in Ohio cannot predict the election in Michigan.

Qi’s group trained the machine learning model using a data set generated by high-throughput computer simulations. The densities and elastic stiffness of about 500 randomly chosen compositions of silica-based glasses were generated from MD simulations. They and their collaborators, Prof. Ge Zhao from Portland State University and Dr. Maarten de Jong from the University of California, Berkeley, developed a machine learning model to reduce the risks of overfitting, a common problem for machine learning model based on a small amount of data.

“Another critical step is that our machine learning model includes the physical mechanism related to glass structures and mechanical deformation. The machine learning inputs are constructed based on parameters that characterize the interaction strengths between atoms inside the glasses. Fortunately, all the parameters are already available in the numerical functions that describe the interatomic forces in MD simulations. These physical-based inputs make our machine learning prediction more reliable,” Qi explained.

“We were amazed to find that our machine learning can be reasonably accurate to predict the densities and elastic stiffness of glass compositions that are significantly different from those in the training data. More surprisingly, our predictions have been confirmed to be generally consistent with experimental results of ~1000 glass compositions. These experimental data were never included in our machine learning training process,” said Yong-Jie Hu, a postdoctoral researcher in materials science and engineering and first author of the paper.

“This discovery is a clear example that the combination of machine learning and human intuition on physical mechanisms is an effective strategy to explore complex physical systems like amorphous materials,” Qi noted. “The next step is to find a more systematic way to transform the human intuition into the computer algorithm for general glass problems.”

For this reason, Qi’s group has shared all of the related data and the machine learning model on Materials Commons, a public open-access repository located at the University of Michigan, and an open-access cloud computing platform (http://vglassdata.org). They emphasize that it is just one critical step to achieve the machine learning design of novel glasses. Models for other glass properties, such as liquidus temperatures and toughness, are also required based on more complex physical mechanisms. To openly share the data and methods can inspire the development of new models in the glass research community.

More details on this research can be found in the paper published in the journal npj Computational Materials, titled, “Predicting densities and elastic moduli of SiO2-based glasses by machine learning.