Page 71 - Read Online

Hu et al. J Mater Inf 2023;3:1 I http://dx.doi.org/10.20517/jmi.2022.28 Page 11 of 15
A

0.6

feature importance 0.4
0.5

0.3
0.2

0.1
0.0 2 3
B 1 2 f B 2 f B f B 1 f B 2 f B f B 1
0.0
0.1
feature importance 0.3
0.2

0.4
0.5

0.6
2 1 2 f B 2 f 2 B f B 1 f B 2 2 2 f B 3

Figure 7. Feature importance analysis of the optimized machine learning model. (A) Positive correlations. (B) Negative correlations.

testing data has never been known by the training model. In this way, we show the predicted against
the measured in Figure 6. It is quite encouraging that the RMSE is only around 0.212 and a direct linear
fitting gives ≈ 0.9. These demonstrate the reliability of our training model to predict new glass formers
2
in the world of computer simulations. The larger fluctuation for better glass formers (lower ) is because of
the greater difficulty in measuring the accurate . Some of these binary alloys can require a longer time to
crystallize beyond the current computational power. The machine learning model will be helpful in predicting
new materials with enhanced GFA for the study of glass physics. For instance, the Kob-Andersen model has
been the most popular model for glass study in the past three decades [35] . It was assumed as a very good
glass former. However, recent studies show that with a larger simulation box and GPU acceleration, the Kob-
Andersen model is vulnerable to crystallization [36,37] . It is actually a poor glass former. The non-additivity
and non-classical energy mixing make the model difficult to understand. A simpler yet better model is of great
interest for glass studies.

We note that there are a variety of machine learning algorithms available and many of them have been applied
in materials development [33,38,39] , such as linear models with regularization, tree-based models, and neural
networks. Forinstance,theiterativerandomforestmodelhasbeenstudiedwidelyinclassifyingglassformation
and regressing glass properties [33,39] . These complex models already showed prediction power, but suffered
from a higher risk of over-fitting. In addition, the interpretability will drastically decrease with increasing
model complexity. Therefore, in this study, we choose to start from the simple linear model with non-linear
combinations of basic features instead of complex, non-linear fancy models, such as artificial neural networks.
We hope to better extract the important couplings of these basic features in glass formation. Interestingly, we
found that the couplings of 1, 2, and are rather crucial in determining the GFA. In Figure 7, we plot the
most important features in the linear model. We split them into two graphs based on whether it is positively or