Page 72 - Read Online

Page 12 of 15 Hu et al. J Mater Inf 2023;3:1 I http://dx.doi.org/10.20517/jmi.2022.28

negatively correlated with . It is surprising to see that the energetic parameter ( 1 2 ) is the most critical
one, without the particle size information. This may provide a plausible explanation for why local chemical
ordering is so important in glass formation [13,15,16] . This top feature is then followed by the couplings of the
particle size and the energetics. These results clarified the critical features in glass formation, at least for binary
systems. They may provide further insights for future experimental glass design. Note that having a binary
alloy with exceptional GFA will be ideal for glass study and applications.

We note that bridging the important model features in the current study to those determined in experiments
is interesting and important. From the work by Liu et al. [40] , it was surprisingly found that the ‘random’ fea-
ture generation from elemental features without enough physical insights is insufficient in machine learning.
Overfittingisreadilytherebyfeedingthosehigh-dimensionalfeaturestoanon-linearrandomforestalgorithm.
Instead, a model with only three features sophisticatedly derived from both elemental and alloy features pro-
vides some predictability. Similarly, in the work by You et al. [41] , an artificial neural network model can classify
crystalline versus amorphous phases in high-throughput fabrications by using a small number of elemental
and alloy features, especially from excess electrical resistivity. The significance of these alloy properties un-
ambiguously demonstrates the non-trivial couplings of elemental properties in metallic alloys. These studies
convey critical messages. On the one hand, physics-driven features from elemental and alloy properties are sig-
nificant. Ontheotherhand, howthecouplingofelementalpropertiestodeterminethealloypropertyiscrucial
in feature engineering. The current study is in line with these spirits: we first identified the four fundamental
physics-driven features. They consider the energetic interactions, atomic sizes, and compositions, which are
consistent with theexperimental inputs. Furthermore, weidentified their critical couplings[Figure 7] thatmay
correlate with some alloy properties. How to directly map these fundamental model parameters in Figure 7 to
experimentally measurable quantities, such as electrical resistivity and liquidus temperature, is interesting for
future study.

4. SUMMARY
The glass-forming ability has been one of the central mysteries for MGs, unlike other families of glass. The
critical cooling rates of MGs can differ by more than 10 orders of magnitude. This huge time gap has fascinated
glass researchers to explore the underlying physics and to design new materials with desired properties. To
accelerateMGdevelopment,weneedtounderstandthephysicalmechanismsofglassformationandlearnfrom
the existing big data accumulated so far. In this study, without relying on collecting experimental data from
the literature sea, we performed large-scale computer simulations over several years to generate a high-quality
dataset. Based on the current understanding of these data, we build an optimized physics-based machine
learning model with only four basic features. The model is able to make reliable predictions on new substances
and provides insights into the most critical features. It is found that the non-linear couplings of the energetic
parameters and geometric parameters are key for glass formation. This further demonstrates the complexity
of the long-standing GFA issue. More interestingly, the most important factor for glass formation is found to
be the coupling of energetic parameters and composition. This rationalizes the crucial role of local chemical
ordering in glass formation and crystallization of metallic alloys, which has been overlooked in the past. A
deeper understanding of the physics of GFA is desired in the future. Practically, generating and maintaining
a high-quality data warehouse for the GFA with extended variables are important for future study. This may
require the collaboration of the whole field.

Although here we focus on binary alloys for simulation convenience, the current study can be effectively ex-
tended to multi-component materials based on the acquired knowledge. With more components, there will
be more independent variables. For example, there will be 12 independent variables for a ternary system.
Even though this will greatly increase the sampling difficulty in molecular dynamics simulations, some opti-
mized high-throughput computational strategies may be developed. It would be interesting to see whether