You are on page 1of 4

Materials Letters 291 (2021) 129573

Contents lists available at ScienceDirect

Materials Letters
journal homepage: www.elsevier.com/locate/mlblue

Modeling the quantitative effect of alloying elements on the Ms


temperature of high carbon steel by artificial neural networks
Xiao-Song Wang 1, P.L. Narayana 1, A.K. Maurya, Hong-In Kim, Bo-Young Hur, N.S. Reddy ⇑
School of Materials Science and Engineering, Gyeongsang National University, Jinju 52828, South Korea

a r t i c l e i n f o a b s t r a c t

Article history: Chemical composition affects the properties and the martensite start (Ms) temperature of steels. This
Received 21 September 2020 study predicts the Ms temperature of high carbon steel via artificial neural networks. Meanwhile, it
Received in revised form 17 November 2020 enables us to estimate the quantitative effect of alloying elements on the Ms temperature on a sizeable
Accepted 12 February 2021
selectable scale, which is the first time to release such results exactly. Compared to the previous formulas,
Available online 18 February 2021
this one is simple, visual, with high accuracy.
Ó 2021 Elsevier B.V. All rights reserved.
Keywords:
High carbon steel
Alloying element
Ms temperature
Quantitative effect
ANN

1. Introduction in Table S1 in Supplementary data. From most of which, a general


formula of linear regression can be extracted,
High carbon steels are the steel series with 0.6–1.0 wt% carbon X
and 0.30–0.90 wt% manganese. These steels are easily heat- y¼hþ xixi ð1Þ
i
treatable due to their high carbon content but have poor formabil-
ity, machinability, and weldability. Their applications include where y is the predictive Ms temperature of target steel. h is an
springs, high-strength wires, welding tools, hand tools, and offset parameter, which means a constant obtained for the Ms
agricultural tillage tools. temperature of pure iron. xi is the weight coefficient (quantitative
Ms temperature is one of the vital parameters in steel-making, effect) of each alloying element. xi is the content of alloying element
at which the steel begins to change its crystal structure from i. Except for the standard type, there are some formulas in particular
austenite into martensite. The dominant affecting factor to Ms tem- forms or containing transformed terms [4,11,16]. As researchers
perature is the chemical composition or alloying element of steel, claimed, these are deduced to modify the estimated values.
although other influences, such as austenitizing state, grain size, Regardless of the formulas, the disadvantage is that nearly each
plastic deformation, and magnetic field, play specific roles as well has a statement for its use range or restricted kind of alloying
[1,2]. Since the minor influences and inadequate relative data, the elements. It can’t be in good agreement with the experimental
factors mentioned above are scarce in the consideration list for result if any condition goes beyond the established scope. Thus,
most studies from the beginning [3–7]. Significant progress had conveniently predicting high carbon steels’ Ms temperature
been made for estimating or predicting the Ms temperature in becomes one of the issues that metallurgists face in operation; that
steels since 1944, initiated by Payson and Savage [3]. From then is one of this study’s aims. We also discuss the quantitative effect
on, scientists release formulas with some modification on their of alloying elements on the high carbon steels’ temperature, which
terms or coefficients. The applicable formulas in the whole or par- is significant for alloy design.
tial range for high carbon steels are collected [4–16] and tabulated
2. Methodology

ANN is one of the machine learning tools whose importance


⇑ Corresponding author. isn’t fully appreciated until a paper about the back-propagation
E-mail address: nsreddy@gnu.ac.kr (N.S. Reddy). algorithm by Rumelhart et al. in 1986 [17]. Using this method,
1
These authors contributed equally. one can study any non-linear systems [18]. A neural network

https://doi.org/10.1016/j.matlet.2021.129573
0167-577X/Ó 2021 Elsevier B.V. All rights reserved.
Xiao-Song Wang, P.L. Narayana, A.K. Maurya et al. Materials Letters 291 (2021) 129573

Fig. 1. The trial results of different parameters on the training, testing and total datasets:

Table 1
High carbon steel sets for the predictive result comparisons.

Data No. C Mn Si Cr Ni Mo V Co Al W Cu N Experimental Ms temperature (K)


35 0.62 0.82 2.01 0.07 0 0 0 0 0 0 0 0 523
36 0.84 0.62 0.23 0.04 0 0 0 0 0 0 0 0 494
39 0.61 0.59 0.24 0.94 0.16 0.33 0 0 0 0 0 0 533
45 0.865 0.4 0 0 0 0 0 0 0 0 0 0 488
47 0.635 0.88 0.28 0 0 0.25 0 0 0 0 0 0 518
49 0.6 0.88 0.23 0.8 0 0 0 0 0 0 0 0 528
51 0.915 0.75 0 0 0 0 0 0 0 0 0 0 489
55 0.6 0.72 0.2 0.17 0.2 0.03 0 0 0 0 0.12 0 543

contains several nodes grouped in three layers - input, hidden, and 3. Grouping: The dataset is divided into training and testing. The
output; wherein the input and output layer have their exact items proportion of the two is generally between 6:4 and 8:2. The
by the case. However, the configuration of other parameters (i.e., more training data are used, the higher partial accuracy the
the number of the hidden layer(s), the included nodes of each model has, but concurrently, the more limited its application
hidden layer, learning rate, momentum term, and iteration) should scope is. We adopted 6:4 here. Therefore, steel sets #1~#33
be determined by trial and error. In general, building an ANN has are for training, and the others (#34~#55) are for testing.
the following procedures: 4. Training: Do trials through the ANN using training data by
changing the parameters one by one, and then find out the most
1. Targeting: Make exact the input and output layer items, namely reliable result by comparing errors of the testing data. The trial
what you have and what you want. In this study, the input layer and error results for different parameters on the training, test-
refers to alloying elements’ content, and the output layer corre- ing, and total data are shown in Fig. 1.
sponds to the Ms temperature. 5. Building: Based on the training results, the optimal structure of
2. Collecting: Collection and construction of the database. Herein, the ANN is considered to be 12-2-1/0.5/0.6/11000, that is, three
fifty-five sets of high carbon steel data are found (Table S2 in layers with respective nodes 12-2-1, learning rate 0.5, momen-
Supplementary data), along with their chemical composition tum term 0.6, and 11,000 iterations, at which the resulting error
and Ms temperature. of the training and testing part is 6.87, and 17.25 degree,

2
Xiao-Song Wang, P.L. Narayana, A.K. Maurya et al. Materials Letters 291 (2021) 129573

Fig. 2. The comparison results of the predictive Ms temperature of high carbon steel sets.

Fig. 3. Step-wise prediction of Ms temperature and the quantitative effect of alloying elements of steel set #55.

3
Xiao-Song Wang, P.L. Narayana, A.K. Maurya et al. Materials Letters 291 (2021) 129573

respectively, and the mean square error (M.S.E.) is as low as CRediT authorship contribution statement
5.15  10-4 degree2. Construct the best-fit ANN model via the
parameters of the least deviation (average error percentage Xiao-Song Wang: Formal analysis, Writing - original draft,
0.0606% on the testing data). Writing - review & editing. P.L. Narayana: Investigation, Data
6. Applying: Calculating the Ms temperature by inputting reason- curation. A.K. Maurya: Investigation. Hong-In Kim: Investigation.
able contents of the alloying elements in the trained ANN model Bo-Young Hur: Funding acquisition. N.S. Reddy: Conceptualiza-
for estimation. tion, Methodology, Software, Writing - review & editing,
Supervision.
A virtual program of alloy addition system is constructed for the
ANN model using JAVA. The relationship between alloying con- Declaration of Competing Interest
tents and Ms temperature can be presented in a visual expression
through the model (Fig. S1 in Supplementary result). The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared
3. Results and discussion to influence the work reported in this paper.

3.1. Prediction comparison of the Ms temperature Acknowledgment

Eight steel sets randomly selected from the untrained database N.S. Reddy Acknowledges YSJ-TV for the inspiration and Y.K.K.
(testing part) are ready to compare the formulas and the ANN’s for helping in Model G.U.I. Development. The authors would like
prediction, listed in Table 1. Fig. 2 shows the predictive Ms temper- to appreciate the Engineering Research Institute’s support,
ature values, along with the linear fitting of the ANN data. The Gyeongsang National University.
figure indicates that the ANN model’s predictions are much more
accurate than those of the previous formulas. For the model, the Appendix A. Supplementary data
Pearson correlation coefficient (Pearson’s r) is up to 0.94999; thus,
the predictions are in good agreement with actual values. Supplementary data to this article can be found online at
https://doi.org/10.1016/j.matlet.2021.129573.
3.2. Quantitative effect of alloying elements
References
The relationship between composition and Ms temperature can
be quantitatively extracted using the ANN model. Typically, the [1] T. Sourmail, C. Garcia-Mateo, A model for predicting the Ms temperatures of
steels, Comput. Mater. Sci. 34 (2) (2005) 213–218.
tendencies or rough values of a few alloying elements’ effects on [2] S.M.C. van Bohemen, L. Morsdorf, Predicting the Ms temperature of steels with
the Ms temperature are probably known [12]. Through the well- a thermodynamic based model including the effect of the prior austenite grain
trained model, the impact of alloying elements on Ms temperature size, Acta Mater. 125 (2017) 401–415.
[3] P. Payson, C.H. Savage, Martensite reactions in alloy steels, Trans. Am. Soc.
can be understood relatively quickly and efficiently. It is also the Metals 33 (1944) 261–280.
first time for the accurate calculation of the topic in this field. [4] L.A. Carapella, Computing A‘‘ or Ms (transformation temperature on
For instance, consider the high carbon steel #55 as an example. quenching) from analysis, Metal Progress 46 (1944) 108.
[5] R.A. Grange, H.M. Stewart, The temperature range of martensite formation,
After adding the alloying elements virtually one by one, the steel’s
Trans. Am. Inst. Mining Metall. Eng. 167 (1946) 467–501.
Ms temperature will vary step by step from that of pure iron, [6] A.E. Nehrenberg, Phosphorus segregation in austenite in Ni–Cr steels, Trans.
shown in Fig. 3. The Ms temperature for pure iron is 668.7 K, which Am. Inst. Mining Metall. Eng. 167 (1946) 494.
is computed by setting the alloying elements at zero in the model. [7] E.S. Rowland, S.R. Lyle, The application of Ms points to case depth
measurement, Trans. Am. Soc. Metals 37 (1946) 27–47.
The Ms temperature for pure iron was deduced from the high [8] Y.X. Liu, Principle of heat treatment, China Mechanical Industry Press, Beijing,
carbon steels database. The quantitative effect of each alloying 1981.
element on the Ms temperature can be known after a simple com- [9] J. Barralis, G. Maeder, Métallurgie-Tome I: métallurgie physique, Collection
Scientifique ENSAM1982.
putation, as shown in the figure. The alloying elements Cu and Si [10] G. Krauss, Steels: heat treatment and processing principles, A.S.M
show a positive effect on the Ms temperature, whereas the C, Mn, International (1990).
Mo, Cr, and Ni have a negative impact (in order of strength, respec- [11] J. Zhao, Continuous cooling transformations in steels, Mater. Sci. Technol. 8
(11) (1992) 997–1003.
tively) [2]. These values of quantitative effect are exact in the [12] K. Ishida, Calculation of the effect of alloying elements on the Ms temperature
range, neither universal ones for all steels. The absolute and per- in steels, J. Alloy. Compd. 220 (1–2) (1995) 126–131.
centage errors of final predictive Ms temperature value upon the [13] A.V. Sverdlin, A.R. Ness, Chapter 2-The Effects of Alloying Elements on the Heat
Treatment of Steel, Marcel Dekker Inc., New York, 1997.
experimental one are only 3.5 degrees and 0.64%, respectively. [14] C. Capdevila, F.G. Caballero, C.G.D. Andrés, Determination of ms temperature
in steels a bayesian neural network model, ISIJ Int. 42 (8) (2002) 894–902.
4. Conclusion [15] U. Lorenz, Anwendung von Werkstoffmodellen auf die Phasenumwandlung
und die Austenitkonditionierung von Stählen, Institut für Eisenhüttenkunde
der RWTH Aachen, 2004
In the present study, we developed an user-friendly standalone [16] S.-J. Lee, K.-S. Park, Prediction of martensite start temperature in alloy steels
efficient ANN model for predicting the Ms temperature of high with different grain sizes, Metall. Mater. Trans. A 44 (8) (2013) 3423–3427.
[17] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning representations by back-
carbon steels compared to the existing models. The model is not propagating errors, Nature (London) 323 (9) (1986) 533–536.
limited to predicting the Ms temperature but also helps to realize [18] N.S. Reddy, J. Krishnaiah, H.B. Young, J.S. Lee, Design of medium carbon steels
the specific calculation of the quantitative effect of alloying by computational intelligence techniques, Comput. Mater. Sci. 101 (2015)
120–126.
elements on Ms temperature. The proposed method will become
an easy and efficient way for alloy design and guide actual exper-
iments or manufacturers.

You might also like