Page 19 - Read Online

Zhou et al. J Mater Inf 2022;2:18 https://dx.doi.org/10.20517/jmi.2022.27 Page 13 of 21

Figure 12. (A) Engineering tensile stress-strain curves of as-cast and as-printed metastable Fe Mn Co Cr Si HEA. (B-D) Phase
[103]
40
20
15
20
5
maps of as-cast and as-printed samples before and after deformation . FCC: face-centered-cubic; HCP: hexagonal-close-packed.
summary, due to the layer-by-layer manufacturing feature and ultrafast cooling rate of L-PBF, multiple
strengthening mechanisms occur in as-printed HEAs, which can synergistically improve their mechanical
properties.
MACHINE LEARNING FOR ALLOY DESIGN AND AM
The compositions of HEAs vary over a wide range, which provides an opportunity to design new alloys with
enhanced properties, such as hardness, tensile property and corrosion resistance. However, it is very
challenging to obtain an alloy with desired properties by a “trial and error” method due to the complexity of
chemical compositions and phases. Compared with the experimental method, machine learning (ML)
provides a new approach to accelerating the discovery of new materials by building the relationship between
targeted properties and various materials descriptors [104-108] . Until now, many works have been conducted to
predict the possible phases in HEAs using ML algorithms, including logistic regression, random forest,
decision tree, K-nearest neighbor, support vector machine and artificial neural network (ANN)
[109-111]
approaches . However, how to choose suitable ML algorithms and descriptors remains a major
challenge in ML due to the numerous combinations between ML models and material descriptors. Based on
the genetic algorithm, Zhang proposed a framework to select the best combination of ML model and
[112]
material descriptor and demonstrated its efficiency for two-phase formation problems in HEAs . Deep
learning is a type of ANN algorithm generally possessing more than three hidden layers. Thus, it is an
efficient data-driven modeling tool for nonlinear system dynamic modeling and identification because of its
flexible structures and general approximation capabilities that can capture complex nonlinear behaviors.
Thus, this method has been widely applied to predict possible phases in HEAs [109-111,113] . For example,
Lee et al. proposed deep learning-based ANN methods for predicting the phase structures of HEAs, i.e.,
solid solution, intermetallic compound, and mixed and amorphous phases. Through the Bayesian
optimization for overall settings related to model architecture, training and regularization, a phase
[113]
prediction model with an unprecedented accuracy of 93.17% was established .