Page 63 - Read Online

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

icosahedral order, have been considered as the most important factors in glass formation. However, in recent
years, we performed systematical studies on the glass formation from model alloys [13–17] and found that the
local chemical ordering can be more important in glass formation than previously thought. It can outperform
local icosahedral orderings even when the size mismatch is considerable. Because of the multi-component na-
ture, the atomic interactions are more complex, and the local chemical ordering can be more significant. The
aging or cooling behavior is important in determining local chemical ordering by controlling atomic diffusion.
Generally, the atomic rearrangements during structural relaxation towards the local equilibrium control local
chemical ordering. Macroscopically, it will depend on the energetic parameters, atomic sizes, and composi-
tion. In addition, the competition among crystalline symmetry is much weaker than that between crystalline
symmetry and crystal-incompatible symmetry. This explains why icosahedral clusters are usually found in
metallic glass formers. These studies refresh the current understanding of the physical mechanisms of crystal-
lization and glass formation and provide guidelines for experimental glass design. Nevertheless, we have never
mined the data itself and built reliable models to predict new glasses. That falls into the efforts of the current
work.

In this paper, we are going to utilize the supervised machine learning method to dig into the simulation dataset
and try to build an optimized model to predict new binary glasses. Since the particle size ratio is helpful in
grouping our dataset, we use the “out-of-group” strategy to make predictions. That is, we leave out a subgroup
of samples with a specific particle size ratio and make predictions for them. Since these data are completely
independent of the others and have not been seen by the training model, we can treat them as “new”. More
importantly, we aim to unveil the key features (factors) that determine the GFA of binary alloys. We find
that non-linear coupling of the elemental features and alloy properties is critical in glass formation. In more
detail, the GFA does not depend on the basic elemental features individually and additively; instead, it depends
on the various non-linear couplings of them. The interactions of these basic elemental features to different
polynomial degrees are more important in making good predictions. These interaction terms have never been
identified previously and can serve as guidelines for future model development and experimental glass design.
Therefore, the results will provide new insights into unravelling the physical mechanism of glass formation and
help accelerate future material design.

2. METHODS
2.1. Molecular dynamics simulations
To generate a clean GFA dataset, we started from the simple binary models with Lennard-Jones potential:

[ ]
( ) 12 ( ) 6

( ) = 4 − , (1)

where , indicate which of the particles (A or B) are interacting and is the separation between parti-
cles and . All the simulations were performed with periodic boundary conditions in all directions. The
cubic simulation box contains = 2000 particles of equal mass . We studied both monodisperse system
( = = ) and additive bidisperse system ( = ( + )/2). Instead, we tuned the inter-
particle interaction strengths ( , and ) widely. We keep ≤ = 1.0 to differentiate the species.
The sampling library is exemplified in Figure 1 as a two-dimensional function of ( − )/( + ) and
2 /( + ). These two variables take both the same species and inter-species interactions into consid-
eration. We then simulated the systems with size mismatch within 5% so that they can mostly crystallize in
the computational time scale. We also sample different compositions (the fraction of B particles in the
total number) from 0.1 to 0.9 in an interval of 0.1. Because of the broad energetic preferences, we employed
(constant number, constant pressure, constant temperature) ensemble with = 10 to avoid cavitation in
any system. To map the experiments, we quenched the high-temperature liquid to very low temperatures at a
series of cooling rates . In this way, the critical cooling rate can be quantified after characterizing the crystal