Page 13 - Read Online

Li et al. J Mater Inf 2024;4:27 https://dx.doi.org/10.20517/jmi.2024.44 Page 7 of 20

Table 1. Compositions and processing conditions designed by an integrated design model based on specific requirements

Compositions (wt.%) Processing parameters
No.
C Cr W Si Mn V Ta Ti N NT (°C) Nt (min) TT (°C) Tt (min)
1# 0.140 8.64 1.30 0.30 0.57 0.320 0.180 0.200 0.0003 1120 21 720 62

2# 0.055 8.40 0.26 0.15 0.50 0.314 0.167 0.217 0.0030 1020 30 650 50

NT: Normalizing temperature; Nt: normalizing time; TT: tempering temperature; Tt: tempering time.

[32]
developed in our previous work and can be directly used in this study.

Microstructural model

The ideal microstructures of RAFM steels should avoid the presence of δ-ferrite, Laves, and Z-phase. δ-ferrite is regarded as an easy propagation site for
intragranular cleavage fractures . The coarsening of Laves phase significantly weakens microstructural stability and accelerates void growth [42,43] , adversely
[41]
affecting the mechanical properties of RAFM steels. The coarsening of Z-phase consumes fine MX-type precipitates during long-term testing and servicing,
impairing the strength of RAFM steels . In this section, the ML classification algorithms were applied to construct prediction models for identifying the
[21]
presence of δ-ferrite at NT (named as Model-I) and the presence of coarsening phases (i.e., Laves and Z-phase) at TT (named as Model-II). For Model-I, the

inputs were compositions and NT, while Model-II used compositions and TT as inputs.

Figure 2A displays the predictive performance of Model-I constructed by six common ML classification algorithms, with each data point averaged over 100
learning processes to reduce random error. It can be learned from Figure 2A that the Model-I model developed by the GBC algorithm achieves the highest Acc
value among the selected six ML algorithms. Figure 2B shows the predictive ability of Model-II constructed by six ML classification algorithms based on Data-

II dataset. Similar to the results in Figure 2A, Model-II constructed by the GBC algorithm has the highest Acc among the selected six ML algorithms, indicating
its best predictive ability. Based on Data-I dataset, Model-I was constructed by GBC algorithm and its detailed classification results are shown in confusion

matrix of Figure 2C. In the training sets, Model-I achieves the Acc of 100% in identifying both “with δ-ferrite” and “without δ-ferrite” classes. For the testing
sets, Model-I classifies “with δ-ferrite” and “without δ-ferrite” with Acc of 100% and 98.0%, respectively. Model-I exhibits high accuracy (> 90%), indicating that
it has excellent predictive ability for identifying the presence of δ-ferrite at NT and neither obvious overfitting nor under-fitting issues. The confusion matrix in

Figure 2D shows the detailed classification results of Model-II. As shown in Figure 2D, Model-II constructed by the GBC algorithm achieves the Acc of > 85.0%
in both the training and testing sets. The above analysis results demonstrate that the Model-I and Model-II constructed in this work have good robustness and

reliability.