Page 129 - Read Online
P. 129
Chen et al. J Mater Inf 2023;3:10 https://dx.doi.org/10.20517/jmi.2023.06 Page 9 of 19
Table 6. List of the ML models based on SVM or ANN with good performance on small-sized databases
Target Size of database Algorithm Performance Ref.
Phase prediction 118 ANN Accuracy = 0.992 [127]
Phase prediction 401 ANN Accuracy = 0.943 [130]
Phase prediction 550 SVM Accuracy = 0.887 [134]
Phase prediction 322 SVM Accuracy = 0.9384 [139]
Phase prediction 391 ANN Accuracy = 0.92 [142]
Phase prediction 407 SVM Accuracy = 0.9743 [149]
Phase prediction 209 ANN Accuracy = 0.9297 [150]
Hardness prediction 155 SVM RMSE = 31 [122]
2
Hardness prediction 214 SVM R = 0.873 [146]
2
Hardness prediction 370 SVM R = 0.8836 [147]
2
Hardness prediction 53 ANN R = 0.8575 [151]
2
Strength prediction 231 ANN R = 0.9702 [152]
2
EHEA Design 321 ANN R = 0.9663 [79]
2
EHEA Design 100 SVM R = 0.916 [80]
ANN: Artificial neural network; EHEA: Eutectic high entropy alloys; RMSE: root mean square error; SVM: support vector machine.
[126]
Figure 4. Confusion matrix of the SVM, ANN models, and the predicted results. Reproduced with permission from Jaiswal et al. .
Copyright 2021, Elsevier. ANN: artificial neural network; SVM: support vector machine.
[147]
(RMSE) are usually used for regressors. It is noteworthy that the performance of the ML models should
be judged not only by existing data (i.e. data in the database) but also by the “unseen” data (data out of the
database). In addition to the above-mentioned numerical evaluations, experimental validation is therefore
[126]
needed that produces unseen data to test the predictability of the ML models. For instance, Jaiswal et al.
used two different ML models (i.e., SVM and ANN) for the phase prediction of the CoCuFeNi system.
x
While both ML models achieve a similar numerical accuracy (~0.85), it appears that the ANN model can
predict the result being consistent only with the experimental observation for low Ni content (x = 5, 10). In
contrast, the SVM model can predict the results correctly only for high Ni content (x = 15, 20, 25), as
illustrated in Figure 4.
Among the above-mentioned ML algorithms, the SVM and ANN are the ones that are widely used in the
[148]
design of HEAs, including EHEAs, due to their good performance on small-sized databases , as shown in
Table 6 [79,80,122,127,130,134,139,142,146,147,149-152] . Here, we note that the reported ML models for EHEAs with good
performance are mostly regressors, outperforming the classifier. This phenomenon could be attributed to
the data imbalance in the EHEA database (i.e. the number of EHEAs is significantly smaller than that of
[153]
non-eutectic HEAs) . In practice, data shortage and/or imbalance could be an issue, particularly for the
design of EHEAs. To mitigate the negative effect, people proposed a few methods, including (1) data
[154]
rebalancing with the Synthetic Minority Oversampling Technique (SMOTE) and the Tomek links for