J Mater Sci (2023) 58:4795–4808

Computation
COMPUTATION & & theory
THE ORY

Development of a machine learning model

for prediction of continuous cooling transformation
diagrams in welding heat-affected zone
Biao Zhang1, Baigang Wang1, Weihua Xue2, Asad Ullah3, Tianhao Zhang1, and
Hao Wang1,*

1
School of Materials Science and Engineering, University of Science and Technology Beijing, Beijing 100083, China
2
School of Materials Science and Engineering, Liaoning Technical University, Fuxin 123099, China
3
Department of Mathematical Sciences, Karakoram International University, Gilgit-Baltistan 15100, Pakistan

Received: 17 November 2022 ABSTRACT

Accepted: 17 February 2023 The traditional thermal simulation technique for the measurement of the weld
Published online: continuous cooling transformation diagrams in synthetic weld heat-affected
7 March 2023 zone (SH-CCT diagrams) is a complex and time-consuming process. Therefore,
it is necessary to develop a model that can rapidly and accurately predict SH-
Ó The Author(s), under CCT diagrams. In this article, an optimized machine learning (ML) model is
exclusive licence to Springer developed to predict the SH-CCT diagrams of various types of steel, which
Science+Business Media, LLC, consist of the Back Propagation (BP) neural network for the prediction of ferrite
part of Springer Nature 2023 transformation starting temperature curves (Fs), and Radial Basis Function
(RBF) neural network for the prediction of bainite, pearlite, and martensite
transformation starting temperature curves (Bs, Ps, Ms). The accuracy of the
present ML model was firstly verified by the excellent agreement between the
predicted transformation starting temperature curves and the measured values
on the internal dataset (Validation Set I). It was also observed that the predic-
tions obtained through the present model are better than those of the previous
models. Moreover, the excellent generalization ability was further verified by
the prediction performance on the external dataset (Validation Set II), in which
the R2 values were all above 0.93, and the MAE values were less than 7 K. This
successful demonstration indicates that the present model can accurately predict
the SH-CCT diagrams of various types of steel, which provides a certain ref-
erence for the optimization of steel welding process parameters and the pre-
diction of heat-affected zone (HAZ) microstructure.

Handling Editor: Ghanshyam Pilania.

Address correspondence to E-mail: hwang@ustb.edu.cn

https://doi.org/10.1007/s10853-023-08322-9

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

4796 J Mater Sci (2023) 58:4795–4808

GRAPHICAL ABSTRACT

Introduction that uses the Gleeble simulator to simulate the ther-

mal cycling process of the HAZ. By enlarging the
volume of the simulated and comparing it with the
Welding technology is a cost-effective way to join real microstructure of the HAZ, the evolution of the
steels together as a whole [1]. However, due to the microstructure and performance can be revealed
complex thermal cycling in the welding process, the [5, 6]. Although this method has high accuracy, it
microstructure of the heat-affected zone (HAZ) encounters some challenges such as long cycle, high
shows great inhomogeneity. This inhomogeneous cost, and high instrument operation requirements [7].
microstructure seriously weakens the strength and In addition, the traditional phase transformation
toughness of the welded joints [2–4]. As one of the models, based on the development of material ther-
effective strategies to understand the microstructure modynamics and dynamics have some difficulties in
variation of HAZ during the continuous cooling describing the transformation process of multiphase
process of steels, the continuous cooling transforma- microstructure and the mechanical properties of
tion diagram in synthetic weld heat-affected zone mixed microstructure, and cannot effectively describe
(SH-CCT diagram) reflects the quantitative relation- the quantitative relationship between temperature,
ship between temperature, time, and phase transfor- microstructure, and welding interface performance.
mation degree, and can be used to predict the Over the past few years, the technology of com-
microstructure and performance of HAZ under given bining data-driven machine learning (ML) with
process conditions. Therefore, an accurate SH-CCT material science has become an effective method to
diagram has great engineering significance for discover new materials and explore the structure–
determining the optimal welding parameters, property mapping relationship of new materials
revealing the microstructure evolution of HAZ, and [8–10]. Different from the materials simulation tech-
improving the mechanical properties of welded niques developed on the basis of mathematics and
joints. physics, ML technology can bypass the complex
The general strategy to obtain SH-CCT diagram is physical mechanism and use the existing historical
to apply the welding thermal simulation technique data and specific process conditions to construct the

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J Mater Sci (2023) 58:4795–4808 4797

model to predict the microstructure and performance developed ML model. Sect. ‘‘Conclusion’’ gives some
of the target alloys. Hence, ML technology provides a important conclusions of the current study.
convenient way to predict the SH-CCT diagrams
quickly and accurately.
Currently, several ML models have been reported Modeling
for the prediction of the Time–temperature-transfor-
mation (TTT) diagrams [11–14] and continuous Data collection, preprocessing, and cleaning
cooling transformation (CCT) diagrams of various
types of steel [15–21]. In contrast, the reports on The reliability of ML prediction largely depends on
predicting SH-CCT diagrams are quite limited so far. the quantity and quality of the dataset. In this study,
Recently, Geng et al. [22, 23] used various ML algo- 256 pairs of SH-CCT diagrams were screened from
rithms consisting of Random Forest (RF), K-Nearest the database of the MatNavi National Institute for
Neighbours (K-NN) and Random committee to Materials Science (NIMS) [25] to construct the model.
establish ML models for predicting SH-CCT dia- Moreover, 4 pairs of SH-CCT diagrams of different
grams for low-alloy steels and Ni-Cr-(Mo) steels types of steel were randomly selected from earlier
based on different steel classification datasets. How- work of Geng et al. [22, 23] (Validation Set I, included
ever, the strategy of dividing datasets according to in NIMS database), together with 4 pairs of diagrams
steel types leads to a small amount of data applied to from the literature [26–29] (Validation Set II, not
train the model, which reduces the prediction accu- included in NIMS database) are applied for the final
racy of the model. Furthermore, this model is only model validation. The alloying elements of all the
applicable to individual alloy systems, which is very collected steels, as well as the composition, are shown
unfavorable for the discovery of new steel materials. in supplementary Tables S-I to S-III (refer to elec-
Minamoto et al. [24] applied the RF algorithm to tronic supplementary material). Figure 1 shows a
develop an ML model and adopted a progressive typical SH-CCT diagram from NIMS database [25].
approach to predict the Ac3 temperatures critical From Fig. 1, the HAZ microstructure evolution from
cooling rates and transformation starting tempera- the Ac3 temperature can be accurately described
tures for a wide range of steels. However, this pro- according to the different cooling rates (CR). If the
gressive prediction approach leads to the error being cooling rate is below the critical cooling rates of
magnified step by step, which eventually affects the pearlite (Cp), austenite transformed to ferrite (C0-C1),
prediction accuracy of the SH-CCT diagrams. pearlite (C1-C2), bainite (C2-C3) and martensite
Besides, the generalizability of the model for new (Below C3), successively. (As per Region I, C0, C1, C2
samples outside the training set cannot be evaluated. and C3 are the intersection points of the cooling curve
Therefore, the existing ML models for predicting SH-
CCT diagrams have their limitations to be used 1200 C Si Mn Ni Cr Mo V Al B
NK-HT80 0.12 0.26 0.86 0.85 0.53 0.54 0.05 0.01 0.00 wt.%
widely.
In this study, in order to address the above prob- Ac3=1076 K
lems, an optimized ML model combining BP neural 1000
Ac1=992 K
network and RBF neural network is proposed to
A Fs
Temperature (K)

predict the SH-CCT diagrams of various types of Cf

Cp c0
F
800 Ps
steel. This work is organized as follows: Cz Bs
c1
c2
P
Ms
Sect. ‘‘Modeling’’ introduces the data collection and c3
Zw
preprocessing process, as well as the optimization
600
process of some important parameters of BP neural
IV III II I
networks and RBF neural networks. Sect. ‘‘Results M
and discussion’’ constructs a combined ML model for CR(K/s)
400 100 53 30 23 10 5 4 2 1 0.5
predicting the SH-CCT diagrams, and compares the
prediction results of the present model based on the
1 10 100 1000
internal and external validation sets with the exper-
Time (s)
imental values as well as the previous work to verify
the accuracy and generalization ability of the Figure 1 A typical SH-CCT diagram [25].

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

4798 J Mater Sci (2023) 58:4795–4808

and the transformation starting temperature curves Machine learning model

of ferrite (Fs), pearlite (Ps), bainite (Bs) and marten-
site (Ms)). The corresponding microstructures refer to The BP and RBF neural networks used in this paper
Supplementary Figure R1. As the cooling rate are both three-layer network structures, including an
exceeds Cp/Cf, implying that the pearlite/ferrite no input layer, a hidden layer and an output layer.
longer appears (As per Region II and III). If the Figure 2 show the schematic diagrams of these two
cooling rate is above Cz, the final microstructure is network structures. The input layer of the neural
transformed into martensite (As per Region IV). network includes composition (Supplementary
Therefore, factors related to the SH-CCT diagram Table S-I), transformation time, temperatures of Ac1,
included the composition, cooling rate (CR), Ac1 and Ac3, and cooling rate (CR), and the output layer are
Ac3 temperatures, as well as the transformation the phase transformation temperatures. The rela-
starting temperature curves of ferrite (Fs), pearlite tionship between input and output layers can be
(Ps), bainite (Bs) and martensite (Ms) are all collected expressed as:
as features to participate in the model construction. T ¼ Fðx; t; CRÞ ð2Þ
As discussed earlier, the strategy of dividing the
datasets by steel types of the existing models has To quantitatively describe the prediction perfor-
many drawbacks, such as the small amount of data mance of the model, mean square error (MSE), mean
applied to train the model, the insufficient prediction absolute error (MAE), and regression coefficient (R2)
accuracy and the weak generalization ability of the are applied to comprehensively evaluate the model
models. To avoid these problems, a new dataset prediction effect. (Eqs. 3–5).
partitioning strategy of targeting specific phase
1X N  2
transformation temperature curves is adopted in this MSE ¼ yi  y^i ð3Þ
N i¼1
work. In other words, if there are specified phase
transformation temperature curves in the graph, the 1X N  
MAE ¼ y^  yi  ð4Þ
relevant information about these transformation N i¼1 i
temperature curves will be collected in the local
PN  2
database regardless of whether other phase trans- i¼1 y^i  yi
2
formation temperature curves exist. This strategy also R ¼ PN  2 ð5Þ
i¼1 yi  y i
ensures the uniformity of prediction models for all
types of steel. Using the digital tool in Origin 2018, all where N represents the number of data, yi and y^i
the transformation temperature curves in the SH- represents the experiment values and the predicted
CCT diagrams of the partitioned datasets were values, respectively. The smaller MSE and MAE
transformed into time–temperature text data. Where values indicate the higher prediction accuracy of the
the time was expressed in logarithmic form, and the model, and the R2 converges to 1, implying the better
cooling rate (CR) can be transformed according to match between the prediction values and the exper-
Eq. (1) imental values.
A C3  T
CR ¼ ð1Þ Optimization of the back propagation (BP) neural network
t

Before training, the initial data should be prepro- BP neural network, as one of the most widely used
cessed to ensure its reliability. In the present work, neural networks, can freely adjust the network
Cook distance value [30] was used to judge the col- structure parameters according to the actual prob-
lected data. If the Cook distance value is too large, it lems. As shown in Fig. 2a, after receiving the input
indicates that the data have a greater impact on the data {x, t,..., CR}, the hidden layer neurons compare
model, and the abnormal data should be removed. the linear weighted values of the weights w and the
Finally, the preprocessed data were divided accord- input information with the initial biases h, and then
ing to the ratio of the training set, test set, and vali- processed by the hidden layer activation function f to
dation set as 0.7, 0.15, and 0.15. The final dataset obtain the activation values. Similarly, the activation
division results are shown in supplementary values are connected to the output layer through the
Table S-IV. weights q, after being processed by the output layer

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J Mater Sci (2023) 58:4795–4808 4799

Figure 2 Architecture of the BP and RBF neural network.

activation to obtain the output value T. Subsequently, 1200

Bayesian Regularization (BR)
the training algorithm continuously corrects the Levenberg Marquardt (LM)
1000
weights and biases of the network through back
Mean Square Error
propagation to minimize the error between the 800
measured value and the output value. Therefore, the
prominent parameters of BP neural network that 600
need to be considered include the training algorithm,
400
the activation function, and the number of neurons in
the hidden layer. In this work, 13 training algorithms
200
(i.e., gradient descent method, quasi-Newton
method, conjugate gradient method, Bayesian Regu- 0
5 10 15 20 25 30
larization and Levenberg Marquardt, etc.), and 3
Number of hidden layer neurons
activation functions (sigmoid, tansig, and purelin) are
selected to establish models for Fs, Bs, Ps, and Ms, Figure 3 Relationship between MSE values of training samples
respectively. The training algorithm and activation and the number of hidden layer neuron.
function with the minimum MSE are determined as
the optimal parameters. Finally, Bayesian Regular- training algorithm is used, the MSE values of the
ization (BR) algorithm is applied to determine the Fs model decrease with the increase in the number of
and Bs. Levenberg Marquardt (LM) algorithm is hidden neurons, and converges to 10 K when the
applied to determine the Ps and Ms. The sigmoid number of hidden neurons is greater than 9. In order
function is determined as the activation function to avoid overfitting phenomenon and improve the
between the input layer and the hidden layer, and the training speed, the number of neurons in the hidden
purelin function is determined as the activation layer is chosen as 9. The detailed model parameters
function between the hidden layer and the output are shown in Supplementary Table S-V.
layer. The predicted performance of the BP neural net-
On the basis of determining the training algorithm work with adjusted parameters is verified by the test
and activation function of the BP neural network, the set, and the initial weights and biases of the BP model
number of hidden layer neurons is further deter- are given random values by MATLAB. The predic-
mined. Since both the LM and BR algorithms are tion results are shown in Fig. 3.
participating in model training, two models are Aiming to avoid the degradation of prediction
established for each training algorithm. Figure 3 accuracy caused by the random assignment of the
illustrates the MSE values with the number of hidden initial weights and biases of the BP neural networks,
neurons for the model trained by the LM and BR various efficient global optimization methods have
algorithms. As can be seen, regardless of which been proposed to improve the prediction

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

4800 J Mater Sci (2023) 58:4795–4808

performance. In the present work, Particle Swarm q

X q
X kxi ci k2
Optimization (PSO) algorithm [31], Mind Evolution- f ð xÞ ¼ wi  uðxi ; ci Þ ¼ wi e 2r2 ð6Þ
i¼1 i¼1
ary Algorithm (MEA) [32], and Genetic Algorithm
(GA) [33] were selected to establish optimization
models for Fs, Bs, Ps, and Ms respectively. The set- where wi is the weight of the ith hidden node, q is the
tings of related PSO, MEA, and GA parameters are number of Gaussian functions,ci is the center point of
summarized in Supplementary Table S-VI. each activation function and r is the spread constant.
In addition, considering that there may be a certain kxi  ci k represent the Euclidean distance from xi to ci .
linear relationship and data noise among the input According to Eq. (6), the parameters of RBF neural
variables, which tend to negatively affect the output network to be solved mainly include the center point ci ,
variables, it is required to extract mutually indepen- weight wi and the spread constant r. More interestingly,
dent low-dimensional features from the high-di- ci can be sampled from the training samples. The weight
mensional features to reduce the dimensionality of wi can be adjusted by back propagation. Thus, the
the feature space and decrease the data noise. In this parameters of the present work are mainly to determine
study, Factor Analysis (FA) method is used to reduce the spread constant. The magnitude of the spread con-
the dimensionality of the feature space, and com- stant of RBF neural network determines how much the
bined the Factor Analysis (FA) method into PSO/ function decays with the change of distance. Therefore,
MEA/GA algorithm to further improve the predic- a suitable spread constant is the prerequisite for the
tion performance of the BP neural network. To show excellent prediction effect of RBF neural network.
the effect of dimensionality reduction on the predic- By considering this fact, a series of RBF neural net-
tion effect, both the non-dimensional reduction work models with progressively increasing spread
model and the dimensionality reduction model are constant are constructed for each transformation tem-
used to predict Fs, Bs, Ps and Ms. The prediction perature curve. The tenfold cross-validation is applied
performance obtained by the two models are shown to training ML model, which randomly divides the
in Fig. 4. training sets into 10 equal parts, one of which is
reserved for validation, the remaining parts are used for
Optimization of the radial basis function (RBF) neural training, and the average value of 10 test parameters
network taken as the final result. Finally, the ideal value of the
spread constant is determined by the R2 of each model.
Compared to BP neural network, as shown in Fig. 2b, Figure 5 illustrates the variation of R2 with the spread
the hidden layer neurons of the RBF neural network constant. It can be observed that with the increases in
adopt the Euclidean distance between the input the spread constant, R2 gradually increases, and con-
information and the center point as the independent verges to 1 when the spread constant reaches a certain
variable of the function, and use radial basis function value. Finally, the spread constant of RBF neural net-
as activation function to achieve the approximation of work for predicting Fs, Bs, Ps, and Ms are determined to
sample values to the center points. Therefore, it can be 82, 36, 56, and 82 respectively, the corresponding R2
approximate any nonlinear function globally and no values are 0.995, 0.993, 0.994, and 0.992 respectively.
local minimum problem exists. Moreover, RBF neural The predicted performance of the RBF neural network
network also has the advantages of good general- with optimized spread constant is verified by the test
ization ability, fast learning, and convergence speed. set, and the results are also superimposed in Fig. 4 for
Thus, RBF neural network is also applied to construct the convenience of comparison with the BP neural
predictive model. network predictions.
The common activation functions of the RBF neural
network include Gaussian function, multi-quadric
function, linear function, etc. [34]. In the present Results and discussion
work, Gaussian function uðxi ; ci Þ is used as the acti-
vation function. As a result, RBF neural network can Model construction
be expressed as follows:
From Fig. 4a, it can be observed that the BP, the RBF,
and the Optimized BP models by combining the

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J Mater Sci (2023) 58:4795–4808 4801

Figure 4 MAE and MSE values of eight models based on the test set.

Factor Analysis (FA) method into PSO/MEA/GA process taking twice as long as that of the RBF model.
algorithm (FA-PSO-BP, FA-MEA-BP, and FA-GA-BP) Considering this fact, the RBF model is the most
have good prediction performance for predicting the suitable choice to predict the Bs. For the prediction of
Fs. Among them, the values of MAE and MSE of the Ps, the MAE values of RBF, FA-PSO-BP and FA-GA-
BP and RBF models are less than 10 K. Compared to BP models are all less than 10 K (see Fig. 4c). Com-
the RBF model, the MSE of the BP model is 1, which pared to other models, RBF model has a lower MSE
is significantly superior to that of the RBF model. This value and excellent learning speed. Such advantages
also implies that the BP model with adjusted are incomparable to the other two models. Conse-
parameters has a significant advantage in predicting quently, the RBF model is also used for the prediction
the Fs. Therefore, the BP model is adopted in the of the Ps. Similarly, as depicted in Fig. 4d, the MAE
present work. As presented in Fig. 4b, both the RBF, and MSE values of the Ms predicted by RBF model
FA-PSO-BP and FA-GA-BP models are suitable for are 1 K and 12, respectively, which are much more
the prediction of Bs, and the MAE values of the three precise than the other models. Undoubtedly, the RBF
models are not greater than 5 K. In terms of training model is also applied for the prediction of the Ms. All
speed, however, the optimization process of the FA- the codes for the present model can be found on
PSO-BP and FA-GA-BP models are extremely time- GitHub (https://github.com/wbg-dev).
consuming, which results in the overall prediction

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

4802 J Mater Sci (2023) 58:4795–4808

Figure 5 The relationship between R2 and spread constant a Fs; b Bs; c Ps; d Ms.

Model validation [22, 23] are also presented in Fig. 6 for further com-
parison (dashed line). It can be observed that all the
Various transformation temperature curves predicted transformation temperature curves predicted by the
by the present model were compared with experi- present work and previous work are very consistent
mental data and previous evaluation results, the with the change trend of the experimental results, but
accuracy and excellent generalization ability of the the match between the prediction values of the pre-
present model are verified firstly. Figure 6 displays sent model and the experimental values is better than
the prediction results obtained by using the present those of the Geng’s model. Moreover, using the
model based on internal validation set (Validation Set Geng’s model requires first classifying the steels and
I). As can be seen that even though the compositions then selecting the appropriate prediction model for
of the 4 types of steel varies considerably (refer to the corresponding steel. The present model simplifies
supplementary Table S-II), the predicted phase tran- the classification process and only needs to input the
sition temperature curves (red lines) have the similar composition and the process parameters of the target
transformation pattern as the corresponding experi- steel, then the model can automatically plot the
mental curves (black lines). Figure 7 shows the transformation temperature curves. This convenience
regression performance between the predicted and is incomparable to other models.
experimental values. Once again, most of the data To visually describe the deviations directly existing
points are homogeneously distributed around the between the present prediction and previous work,
asymptotic lines, which sufficiently demonstrates the the prediction performance on the internal validation
high prediction accuracy of the present model. set (Validation Set I) of the present model as well as
Meanwhile, the results obtained by Geng et al.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J Mater Sci (2023) 58:4795–4808 4803

Figure 6 Predicted SH-CCT diagram by the present model, compared with the experimental data [25]and previous assessment [22, 23]
(dashed lines). a 1#; b 2#; c 3#; d 4#.

Geng’s models are summarized in Table 1. As pre- samples are essentially 0, which results in the R2
sented in Table 1, the MAE values of each transfor- deviates far from 1.
mation temperature curve in the 4 types of steel The ability to extend to new samples outside the
determined by the present model are less than 11 K, training datasets has been an important indicator for
which is significantly superior to the previous pre- evaluating the generalizability of established
diction, even though they claim that the errors machine learning [35]. Based on the external valida-
already meet industrial requirements. In addition, the tion set (Validation Set II), the generalization ability
R2 of all the transformation temperature curves in 1#, of the present model on different source datasets was
2#, and 3# reaches 0.98, except for the Ms in 1#, further verified, and the predicted results are shown
implying that the present model has excellent pre- in Fig. 8 and Fig. 9. It is evident from these fig-
diction performance and generalization ability. As for ures that the present predicted transformation tem-
Ms in 1#, Bs and Ms in 4#, the MSE and MAE values perature curves can account for most of the measured
are small (7.11/0.58, 79.92/3.21 and 13.10/0.92, data except those of Fs in the 6# and 8#. The accuracy
respectively), but the R2 is far away from 1. The of the Fs predicted by the present model gradually
reason for this deviation is due to the fact that these decreases when the CR is below 2 K/s and 6 K/s
transformation temperature curves are almost con- respectively, resulting in large deviations between
stant as the CR decreases, so the variances of the the predicted values and the measured values. The

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

4804 J Mater Sci (2023) 58:4795–4808

Figure 7 The predicted values versus the experimental values on Validation Sets I. a 1#; b 2#; c 3#; d 4#.

Table 1 Summary of the MSE/MAE/ R2 of the present work and previous work on Validation Sets I

Steel Fs Bs Ps Ms References

1# 51.22/5.22/0.9944 0.01/0.04/0.9995 126.41/10.99/0.9905 7.11/0.58/0.0581 This work

492.84/19.3 31.36/4.1 94.09/9.1 169/10.7 [22]
2# 5.45/1.83/0.9988 0.08/0.17/0.9998 – 4.74/0.64/0.9814 This work
102.01/8.4 92.16/6.7 – 234.09/13.3 [23]
3# 18.14/2.51/0.9985 2.16/0.47/0.9991 16.03/3.03/0.9991 0.01/0.03/0.9998 This work
645.16/20.3 342.25/15.6 174.24/10.8 912.04/23.2 [22]
4# 273.35/10.86/0.8451 79.92/3.21/0.5168 – 13.10/0.92/0.6823 This work
90.25/4.8 380.25/17.9 – 67.24/7.6 [23]

reason for this bias is due to the small amount of the measured data implies that the present results are
Fs datasets within this range. On the other hand, acceptable. Figure 10 displays the error statistical
different sources of data also may lead to prediction parameters between the predicted values and the
performance degradation. Moreover, the close con- experimental values. Although the data were from
sistency of the predicted Fs with the trend of the different sources, the R2 values predicted by the

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J Mater Sci (2023) 58:4795–4808 4805

Figure 8 Predicted SH-CCT diagrams by the present model, compared with experimental data [26–29]. a 5#; b 6#; c 7#; d 8#.

present model are all above 0.93, and the MAE values various transformation temperature curves predicted
are all less than 7 K. This further indicates that the by the present model are in good agreement with the
present model also exhibits excellent generalization experimental values, and the MAE values in the
ability to different source datasets. internal data set are all less than 11 K, which indi-
cates that the present model has excellent prediction
accuracy. Meanwhile, it is also observed that the
Conclusion prediction results obtained by the present model fit
well with the experimental ones as compared to the
In this paper, eight optimized neural networks were previous model. Moreover, the prediction perfor-
used to construct ML models to predict phase mance of the model on the external validation set
transformation temperature curves in SH-CCT dia- (Validation Set II) is consistent with the experimental
grams of various types of steel. The evaluation results values, and the R2 reaches 0.93, MAE value less than
show that the BP neural network was suitable for 7 K, which again verifies the excellent generalization
predicting Fs and the RBF neural network was suit- ability of the present model.
able for predicting Bs, Ps, and Ms. The prediction In summary, the optimization model developed in
accuracy of the present model was first validated by this paper based on BP and RBF neural networks can
evaluating the prediction performance on the internal achieve superior predictive performance in predict-
validation set (Validation Set I). The results show that ing SH-CCT diagrams for a wide range of steels.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

4806 J Mater Sci (2023) 58:4795–4808

Figure 9 The predicted values versus the experimental values on Validation Sets I. a 5#; b 6#; c 7#; d 8#.

the welding process while facing difficulties with

experimental approaches.

Acknowledgements
The authors acknowledge the financial support from
the National Key Research and Development Pro-
gram of China [No. 2016YFB0700505] and the
National Natural Science Foundation of China [No.
51571020].
Figure 10 Error statistical parameters predicted by present
models on Validation Sets II.
Author contributions
Moreover, the generalization ability for different
source samples were further confirmed. This model BZ contributed to Writing-Original draft preparation,
also provides a strategy for researchers studying the Methodology. BW contributed to Data curation,
microstructure evolution of the HAZ or optimizing Software. WX contributed to Formal analysis. AU
contributed to Writing-Review and Editing. TZ

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

J Mater Sci (2023) 58:4795–4808 4807

contributed to Investigation. HW contributed to [8] Butler KT, Davies DW, Hugh C, Olexandr I, Aron W (2018)
Conceptualization, Project administration, Funding Machine learning for molecular and materials science. Nat-
acquisition. ure 559:547–555
[9] Agrawal A, Choudhary A (2016) Perspective: materials
Declarations informatics and big data: realization of the ‘‘fourth para-
digm’’ of science in materials science. Apl Mater 4:53208
Conflict of interest The authors declare that they
[10] Hattrick-Simpers JR, Gregoire JM, Kusne AG (2016) Per-
have no conflict of interest.
spective: composition-structure-property mapping in high-
Data availability The data source files are available throughput experiments: turning data into knowledge. Apl
in the supplementary material. Mater 4:2522–2832
[11] Kundu M, Ganguly S, Datta S, Chattopadhyay P (2009)
Ethical approval Not Applicable. Simulating time temperature transformation diagram of steel
using artificial neural network. Mater Manuf Process
Supplementary Information: The online version 24:169–173
contains supplementary material available at http [12] Huang XY, Wang H, Xue WH, Xiang S, Huang HL (2020)
s://doi.org/10.1007/s10853-023-08322-9. Study on time-temperature-transformation diagrams of
stainless steel using machine-learning approach. Comp
Mater Sci 171:109282
References [13] Huang XY, Wang H, Xue WH, Ullah A, Xiang S, Huang
HL, Meng L, Ma G, Zhang GZ (2020) A combined machine
[1] Wu XY, Lin HT, Luo W, Jiang HT (2021) Microstructure learning model for the prediction of time-temperature-
and microhardness evolution of thermal simulated HAZ of transformation diagrams of high-alloy steels. J Alloy Compd
Q&P980 steel. J Mater Res Technol 15:6067–6078 823:153694
[2] Geng P, Morimura M, Ma H, Ma Y, Ma N, Liu H, Aoki Y, [14] Qiao L, Zhu J, Wang Y (2021) Modeling of alloying effect
Fujii H, Qin G (2022) Elucidation of intermetallic com- on isothermal transformation: a case study for pearlitic steel.
pounds and mechanical properties of dissimilar friction stir Adv Eng Mater 23:2001299
lap welded 5052 Al alloy and DP590 steel. J Alloy Compd [15] Chakraborty S, Das P, Kaveti NK, Chattopadhyay PP, Datta
906:164381 S (2018) MCDM towards knowledge incorporation in ANN
[3] Wu ZQ, Liu S, Yang ZB, Zhou HG, Li E, Pan HJ, Chen YN, models for phase transformation in continuous cooling of
Lu LW, Liu LF (2022) Simultaneously improved the strength steel. Multidiscip Model Ma 15:170–186
dramatically and eliminated HAZ softening of DP980 steel [16] Chakraborty S, Chattopadhyay PP, Ghosh SK, Datta S
pulsed-arc welding joints by PWHT. Mater Sci Eng A (2017) Incorporation of prior knowledge in neural network
837:142752 model for continuous cooling of steel using genetic algo-
[4] Song KJ, Wei YH, Dong ZB, Zhang XH, Zheng WJ, Fang K rithm. Appl Soft Comput 58:297–306
(2013) Numerical simulation of b to a phase transformation [17] Trzaska J, Jagiełło A, Dobrzański LA (2009) The calculation
in heat affected zone during welding of TA15 alloy. Comput of CCT diagrams for engineering steels. Arch Mater Sci
Mater Sci 72:93–100 39:13–20
[5] Albert SK, Matsui M, Hongo H, Watanabe T, Kubo K, [18] Ghosh SK, Chattopadhyay PP, Haldar A, Ganguly S, Datta S
Tabuchi M (2004) Creep rupture properties of HAZs of a (2008) Design of the directly air-cooled pearlite-free multi-
high Cr ferritic steel simulated by a weld simulator. Int J Pres phase steel from CCT diagrams developed using ANN and
Ves Pip 81:221–234 dilatometric methods. ISIJ Int 48:649–657
[6] Shao Y, Yan BY, Liu YH, Mao CL, Wei C, Liu YC, Yan ZS, [19] You W, Xu WH, Liu YX, Bai BZ, Fang HS (2007) Effect of
Li HJ, Liu CX (2019) Thermal simulation on double-pass chromium on CCT diagrams of novel air-cooled bainite
welding of a high Cr ferritic steel. J Manuf Prcesses 43:9–16 steels analyzed by neural network. J Iron Steel Res Int
[7] Geng XX, Chen Z, Wang SZ, Peng CK, Ullah A, Wang H, 14:39–42
Wu GL (2020) A data-driven machine learning approach to [20] Trzaska J, Dobrzański LA (2007) Modelling of CCT dia-
predict the hardenability curve of boron steels and assist grams for engineering and constructional steels. J Mater
alloy design. J Mater Sci 57:10755–10768https://doi.org/10. Process Technol 192:504–510
1007/s10853-022-07132-9 [21] Geng XX, Wang H, Xue WH, Song XC, Huang HL, Meng
L, Ma G (2020) Modeling of CCT diagrams for tool steels

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

4808 J Mater Sci (2023) 58:4795–4808

using different machine learning techniques. Comp Mater [30] Cook RD (1977) Detection of influential observation in
Sci 171:109235 linear regression. Technometrics 19:15–18
[22] Geng XX, Mao XP, Wu HH, Wang SZ, Xue WH, Zhang GZ, [31] Pratt L, Nicodemus S (1994) Proceedings of 1994 IEEE
Ullah A, Wang H (2022) A hybrid machine learning model international conference on neural networks (ICNN’94). In:
for predicting continuous cooling transformation diagrams in IEEE World congress on IEEE international conference on
welding heat-affected zone of low alloy steels. J Mech Sci neural networks vol 1, p 78–83
Technol 107:207–215 [32] Lei YM, Chen SJ (2009) A reliable parallel interval global
[23] Geng XX, Wang H, Ullah A, Xue WH, Ma G (2020) Pre- optimization algorithm based on mind evolutionary compu-
diction of continuous cooling transformation diagrams for tation. In: Fourth Chinagrid Conference p 205–209
Ni-Cr-Mo welding steels via machine learning approaches. [33] Nakhjavani OB, Ghoreishi M (2006) Multi criteria opti-
JOM 72:3926–3934 mization of laser percussion drilling process using artificial
[24] Minamoto S, Tsukamoto S, Kasuya T, Watanabe M, Demura neural network model combined with genetic algorithm.
M (2022) Prediction of continuous cooling transformation Mater Manuf Process 21:11–18
diagram for weld heat affected zone by machine learning. Sci [34] Mortazavi S, Ince A (2020) An artificial neural network
Technol Adv Mater Meth 2:402–415 modeling approach for short and long fatigue crack propa-
[25] CCT Diagram data base system national institute for mate- gation. Comp Mater Sci 185:109962
rials science, Japan, 2021 https://weldcct.nims.go.jp/Weld/ [35] Juneja R, Singh AK (2020) Guided patchwork kriging to
search/en/ develop highly transferable thermal conductivity prediction
[26] Wang JL, Di H (1994) Xue GL (1994) Mappihg for SH-CCT models. J Phys Mater 3:024006
diagram of low temperature pressure vessel steel
10MnNbVR. J Northeast Univer Nat Sci 15:175–179 Publisher’s Note Springer Nature remains neutral with
[27] Harbin institute of welding, ministry of machinery and regard to jurisdictional claims in published maps and
electronics industry (1990) Atlas of domestic low alloy steel institutional affiliations.
welding CCT. Beijing, China
[28] Jing TF, Wang MJ, Zheng YZ, Gao M, Lin ZH, Guo JR, He Springer Nature or its licensor (e.g. a society or other
GZ (1993) SH-CCT diagram of 09MnVTiN steel. Iron Steel partner) holds exclusive rights to this article under a
28:45–53 publishing agreement with the author(s) or other rightsh-
[29] Li Y, Yang CG, Li Q (2001) Measurement of the steel older(s); author self-archiving of the accepted manuscript
P46ONL1 SH-CCT. Mach Des Manuf 6:74–75 version of this article is solely governed by the terms of
such publishing agreement and applicable law.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

onlineservice@springernature.com