SCIENCE AND TECHNOLOGY OF ADVANCED MATERIALS

2019, VOL. 20, NO. 1, 972–978

https://doi.org/10.1080/14686996.2019.1671140

Data analytics approach for melt-pool geometries in metal additive

manufacturing
Seulbi Leea, Jian Pengb, Dongwon Shin b
and Yoon Suk Choia
a
School of Materials Science and Engineering, Pusan National University, Busan, Korea;
b
Materials Science and Technology Division, Oak Ridge National Laboratory, Oak Ridge, TN, USA

ABSTRACT ARTICLE HISTORY

Modern data analytics was employed to understand and predict physics-based melt-pool forma- Received 31 July 2019
tion by fabricating Ni alloy single tracks using powder bed fusion. An extensive database of melt- Revised 19 September 2019
pool geometries was created, including processing parameters and material characteristics as Accepted 19 September 2019
input features. Correlation analysis provided insight for relationships between process para- KEYWORDS
meters and melt-pools, and enabled the development of meaningful machine learning models Powder bed fusion (PBF)
via the use of highly correlated features. We successfully demonstrated that data analytics process; melt-pool; single
facilitates understanding of the inherent physics and reliable prediction of melt-pool geometries. track; machine learning;
This approach can serve as a basis for the melt-pool control and process optimization. correlation analysis
CLASSIFICATION
106 Metallic materials; 404
Materials informatics /
Genomics

1. Introduction thermo-mechanical properties, and so on. Extensive

research, as well as computational analysis, has been
Additive manufacturing (AM) offers many key benefits,
performed to clarify the effects of materials and process
which could change the industrial paradigm in various
parameters on melt-pool characteristics, and the under-
fields, as a tool-free, cost-efficient and digital approach
lying physics [2–7]. However, there are more than 130
to manufacturing [1]. Among the various metal AM
processing parameters associated with the PBF process
processes, powder bed fusion (PBF) offers advantages
[8,9], which interact intricately with various phase
for the fabrication of fully dense, near-net-shape metal-
changes over a wide temperature range during the pro-
lic parts directly using high-energy heat sources, such as
cess. It is difficult to understand the overall relationships
a laser and an electron beam. The PBF process com-
among the processing variables through well-controlled
prises a sequence of layer depositions by selectively
experiments or simulations using only a few variables.
melting a powder layer (together with the previously
Also, it is challenging to quantitatively determine which
deposited layer (or a substrate)), which leads to the
features have utmost priorities, and how relevant they are
formation of a series of melt-pools on a micrometer
to the characteristics of AM-processed parts.
spatial scale, formed by harsh solidification conditions
Herein, we introduce an emerging data analytics
on a millisecond temporal scale. Here, increasing the
approach to predict the multi-physics-based phenom-
reliability and stability of additively manufactured pro-
ena in AM, along with a scientific insight into the
ducts as a function of processing conditions is a grand
underlying mechanisms [10–15]. To predict melt-
challenge.
pool geometries, machine learning training was per-
The melt-pool geometry is important as a primary
formed using a database of PBF-processed single
criterion for optimizing processing conditions because it
tracks of Alloy 625 and Alloy 718 powders. Through
is an indirect evidence of interactions between processing
the correlation analysis, the machine learning models
parameters and intrinsic materials properties. Also, there
were improved using highly correlated features, which
is a need to understand geometry-related phenomena,
enabled us to interpret the underlying physical
such as defect formation, microstructure evolution,

CONTACT Yoon Suk Choi [email protected] School of Materials Science and Engineering, Pusan National University, Busan, Korea
This article has been republished with minor changes. These changes do not impact the academic content of the article.
The supplementary data for this article can be accessed here.
© 2019 The Author(s). Published by National Institute for Materials Science in partnership with Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Sci. Technol. Adv. Mater. 20 (2019) 973 S. LEE et al.

mechanisms. Finally, the results demonstrated that the beam diameter and layer thickness were used as pro-
reasonably trained data analytics approach prioritizes cess parameters for the PBF operation. Table 1 sum-
key materials/process parameters of the melt-pool for- marizes the laser parameters and powder bed
mation, with the physical relevance, and facilitates the properties under a wide range of process windows.
accurate prediction of melt-pool geometries for the All processes were performed at room temperature
AM process optimization. in an inert argon atmosphere.
For reliable data acquisition, sections in the mid-
dle of the 3-mm long single tracks were prepared for
2. Experimental and data analytics details the metallography. Figure 1 presents the representa-
tive optical metallographs in a cross-section view,
2.1. Data acquisition through systematic
showing the melting mode from conduction to key-
microstructural analysis
hole. As shown in Figure 2, the target output para-
Pre-alloyed spherical powders of Alloy 625 (21.4Cr- meters (will be referred to as targets, hereafter) were
9.0Mo-3.3Nb-3.6Fe-0.2Co–0.3Mn-0.3Ti-0.2Al-Ni, by set up such that they represent characteristics of
Chang Sung Co.) and of Alloy 718 (19.2Cr-3.1 melt-pool geometries: those measured in the sub-
Mo-5.0Nb-18.4Fe-0.2Co-0.05Mn-1.0Ti-0.5Al-Ni, by strate (width (w), depth (d), and area within the
Carpenter Inc.), were used to fabricate single tracks substrate (Asub) in Figure 2) and in the powder
on a substrate of the same material. The PBF machines bed (height (h) and area based on the height (Ah)
used in this study were Mlab and M2 cusing (Concept in Figure 2). By measuring these target values
laser) with maximum laser powers of 100 W and through image analysis, a geometry database of 472
400 W, respectively. The laser power, scan speed, melt-pools was constructed (Table 1).

Table 1. List of PBF process parameters for fabricating single tracks of alloy 625 and alloy 718 powders.
Material (machine)
Processing parameter Alloy 625 (M2) Alloy 718 (Mlab) Alloy 718 (M2)
Power (W) 120, 180, 240 70, 80, 90, 100 120, 180, 240
Scan speed (mm/s) 200, 400, 500, 700, 200, 400,
600, 800, 1000 800, 900, 1100 600, 800, 1000
Beam diameter (μm) 50, 100, 150 50 50, 100, 150
Layer thickness (μm) 25, 50 25, 35, 45 25, 50
Powder size distribution 18, 31, 49 23, 34, 45 15, 30, 47
– D10, D50, D90 (μm)
Total no. of the data set 175 117 180

Figure 1. Metallographs of melt-pool cross sections for (a) Alloy 625 and (b) Alloy 718 as a function of the scan speed. Here, the
laser power, beam diameter and layer thickness are 180 W, 50 μm and 25 μm, respectively.
Sci. Technol. Adv. Mater. 20 (2019) 974 S. LEE et al.

Figure 2. Schematic illustration of (a) single track deposition by the PBF process and (b) the melt-pool geometry parameters
(targets) from the cross-section view.

2.2. Data analytics meaningful machine learning models using highly

ranked features.
(1) Correlation analysis
(2) Machine learning
A total of 5 targets and 23 input features categor-
ized by process parameters and material-related prop-
In this study, six machine learning algorithms were
erties are listed in Table 2.
employed for generality (i.e. not limited to a specific
The correlation analysis between input features and
data analytic algorithm): Bayesian ridge regression
targets was performed using both the advanced maximal
(BR) [18], kernel ridge regression (KR) [19], linear
information coefficient (MIC) approach and the conven-
regression (LR) [20], nearest neighbors regression
tional Pearson’s correlation coefficient (PCC) approach.
(NN) [21], random forest regression (RF) [22], and
The PCC approach is suitable for seeking the linear
support vector machine (SVM) [23]. All of these are in
relationship between input features and target, while
the open source data analytic toolkit ASCENDS [24].
MIC approach can identify non-linear relationships of
Based on a 5-fold cross-validation, the predictability of
high-dimensional large dataset [10,15–17]. By using two
machine learning models was evaluated by calculating
different approaches, we expect to have insights for dif-
the coefficient of determination (R2) between the
ferent statistical aspects between input features and target
actual and predicted values as a function of the num-
properties and then to inspire domain expert for under-
ber of top-ranked features (5, 10, 15, 20 and all)
standing the physical phenomena [15]. In the PCC
associated with each correlation approach.
approach, coefficient value has between −1 and +1,
where a positive value indicates a direct relationship
and a negative correlation coefficient indicates
3. Results and discussion
a reciprocal relationship between input features and tar-
gets. The MIC belongs to the maximal information- Figure 3 shows the results of the correlation analysis for
based nonparametric exploration (MINE) [17], having the five targets using both MIC and PCC approaches.
between 0 to 1 coefficient value. To compare the results Here, asterisk marks on PCC correlation bars indicate
of both approaches, MIC values and absolute PCC values negative correlation coefficients, which mean inverse
were used. Based on the quantitative comparison of these influence. The rank order of the features shows
values, we can identify which physical phenomenon is a distinct difference between the melt-pool geometries
predominant for each melt-pool formation and develop associated with the substrate (w, d and Asub) and with the

Table 2. List of input features and targets used in the present study.
Classification Constituents
Input features Chemistry of powders Ni, Cr, Mo, Nb, Fe, Co, Mn, Ti, Al
Materials thermal property Solidus, liquidus, density, conductivity, thermal diffusivity, specific heat
Information Powder size distribution (D10, D50, D90), layer thickness
of the powder bed
Laser parameters Power, scan speed, energy density*,
beam diameter
Targets for the melt-pool geometry Measured in the substrate Width, depth, area within the substrate
Measured in the Height, area based on the height
powder bed
* Energy density is given by Eρ ¼ vPl (herein, Eρ is the surface energy density, P is the power, v is the scan speed, and l is the layer thickness of powder bed.)
Sci. Technol. Adv. Mater. 20 (2019) 975 S. LEE et al.

Figure 3. Results of correlation analysis using the MIC approach, compared with the absolute PCC values (bars marked by
patterns), for the five targets (w, d, Asub, h and Ah). In the plots, asterisk marks on the PCC bars indicate negative values.

powder bed (h and Ah). Also, the rank orders are differ- In Figure 4, the accuracies (R2-values) show higher
ent depending on the correlation approach and targets, predictability for the melt-pool geometries associated
as listed in supplementary Table 1. with the substrate (w, d and Asub), when the PCC
In Figure 3, the energy density, laser power, scan approach is employed, and higher predictability for
speed, and powder size distribution were highly ranked the melt-pool geometries associated with the powder
for most targets. Here, first three parameters are asso- bed (h and Ah) under the MIC approach (although
ciated with how much energy is emitted by the heat w and Asub are highly predictable using both
source. The powder size distribution (in the powder approaches). In general, the PCC approach is suitable
bed) is related to the energy absorptivity. The laser for capturing linear relationships between the input
beam interacts with the powder bed first, and multiple features and targets; whereas the MIC approach is sui-
scattering between the powders causes a higher laser table for capturing a wide range of relationships, includ-
absorptivity than on the bare substrate. In particular, ing linear, exponential, or periodic, or even all
the laser absorptivity tends to be chiefly dependent on functional relationships [17]. A possible interpretation
the powder size distribution in the top layer, which of the results is that melt-pool geometries measured in
affected the primary reflection of the laser [25]. the substrate have linear relationships with the most
To identify the dependence for selected features highly ranked features, whereas those measured in the
based on correlation analysis, we conducted the accu- powder bed have complex relationships with the most
racy analysis of 5 machine learning training cycles, as highly ranked features. Also, the equitability of the MIC
shown in Figure 4 (numeric values with standard approach seems to facilitate interpretation of the rela-
deviations are listed in supplementary Table2). tionships with noises [17] caused by surface defects,
Sci. Technol. Adv. Mater. 20 (2019) 976 S. LEE et al.

Figure 4. Accuracy analysis of the five targets (w, d, Asub, h and Ah) as a function of the number of the top-ranked features (5, 10,
15, 20 and all), determined by the two correlation approaches (MIC and PCC), for the six different machine learning algorithms (BR,
KR, LR, NN, RF and SVM). Here, R2-values are averages after 5 machine learning training cycles.

vaporization losses, height roughness, and so on. In the that the liquid metal under an incident laser moves
following, the main discussion will focus on correlation away from the bottom to the rear part of the melt-pool
approaches with better machine learning accuracies through a dynamical vortex fluid for heat dissipation.
(PCC for w, d and Asub, and MIC for h and Ah). In brief, the direct laser penetration is a major phe-
In Figure 3, the energy density, laser power, scan nomenon that determines d, and the heat transfer
speed, and powder size distribution ranked high for w, through the fluid convection is a major physical phe-
d and Asub based on the PCC approach. Chemistries nomenon that determines w. The correlation analysis
and material properties followed in the rankings; they performed in the present study accurately captured
show different PCC correlation values between such a difference.
w (about 0.17), and d and Asub (≤ 0.03). Here, the It is interesting in Figure 3 that the powder size
similar correlation tendency for d and Asub is probably distribution, chemistries, and material properties have
due to a proportionality between two targets. much higher correlation values for h and Ah than do
Considering the melt-pool formation under the laser the targets associated with the substrate (w, d and
irradiation, d is determined by direct laser penetration Asub) for both PCC and MIC approaches. Since the
(the laser cavity) resulting from the largest recoil pres- melt volume above the substrate is determined by the
sure. However, w is decided by the heat transfer, mainly amount of consolidated powders, the powder size dis-
through the convection of liquid flow surrounding the tribution is the highly correlated feature for h and Ah
laser cavity [2,3,7]. in terms of the powder packing density [26]. After the
This phenomenon was observed by Zhao et al [2], laser passes, the final h and Ah values will be decided
who conducted in situ observations of the melt-pool by the fluid characteristics during cooling. Since fluid
formation through the synchrotron X-ray imaging characteristics are intrinsically based on material
technique. The experimental evidence shows that the properties and chemistries, these features have high
depth of the laser cavity is similar to the d value of the correlation values with melt-pool geometries asso-
melt-pool, while the w value of the melt-pool continu- ciated with powder bed.
ously increases because the cavity has a constant width Based on the accuracies of the trained models in
(similar to the beam diameter) during melt-pool evo- Figure 4, the minimum numbers of top-ranked fea-
lution [2]. Herein, it can be seen that the PCC correla- tures, excluding noise-inducing features and including
tion value of the beam diameter for w is higher than physically key features [15], were determined to be 10
that for d, also having a positive value (Figure 3). The and 20 for melt-pool geometries measured in the sub-
computational study by Khairallah et al. [3] showed strate (w, d and Asub) and the powder bed (h and Ah),
Sci. Technol. Adv. Mater. 20 (2019) 977 S. LEE et al.

Figure 5. Accuracy analysis of the five targets (w, d, Asub, h and Ah) for the six optimized machine learning algorithms (BR, KR, LR,
NN, RF and SVM). Here, average R2-values and standard deviations are plotted after 5 machine learning training cycles.

respectively. Figure 5 shows the resulting accuracies 4. Conclusions

(R2-values) of the six optimized machine learning
models. Modern data analytics approaches were employed to
In Figure 5, for w, d and Asub, the machine learning check the feasibility of predicting melt-pool geome-
models show fairly high predictability, having a highest tries of PBF single-tracks and the physical relevance to
R2-value of more than 0.9, which is striking, considering the real phenomena, and the following conclusions
a limited number of melt-pool data available for train- were drawn.
ing. However, for h and Ah, the results show rather low
predictability, having R2-values between 0.75 and 0.85. (1) Machine learning models better predicted
The reason for the relatively low predictability of h and melt-pool geometries measured in the substrate
Ah seems to be associated with the complexity of with the PCC approach, and those measured in
dynamic powder motions. During the laser irradiation, the powder bed with the MIC approach. It
vapor-driven flow causes powder spatter ejection back- implies different interactions of input features
ward, upward, or even forward with respect to the scan and output targets by the laser irradiation to the
direction [3–5]. In particular, the development of powder bed and substrate.
dynamic vapor flow as a result of uneven temperature (2) The correlation analysis suggested that the
distribution in the melt-pool makes it even more diffi- depth, width and height of the melt-pool are
cult to predict powder motion [2–5]. The powder physically relevant with the laser penetration,
entrainment toward the melt-pool caused by the inter- fluid convection, and fluid properties of melted
action between the metal vapor and the surrounding powders, respectively.
gas pressure causes more complex powder motion (3) Machine learning models optimized with high
[2,4,5]. Also, the difficulty of controlling the layer thick- ranking features accurately predicted melt-pool
ness uniformity over the powder bed seems to give an geometries measured in the substrate, but
additional effect on the uncertainty of the powder bed showed relatively low predictability for those
properties [27]. Therefore, for the improvement of pre- measured in the powder, which seemed to be
dictability for h and Ah, it is suggested to incorporate associated with complex and dynamic powder
input features representing the gas- and vapor-driven motions.
powder behaviors, and related to the material properties (4) Input features prioritized by the correlation
associated with optics and rheological characteristics. analysis and machine learning training well
Also, we will consider the integrated input features captured the relevant physical phenomena to
(such as D90/D10 for the powder distribution, new form melt-pools. A strong possibility of using
energy density formula, and so on) and the new repre- modern data analytics to accelerate the PBF
sentative targets for defining the melt-pool geometries process optimization for the melt-pool control
having physically meaningful insights. was demonstrated.
Sci. Technol. Adv. Mater. 20 (2019) 978 S. LEE et al.

Acknowledgments in Manufacturing; 2005 Jun 13–16; Munich,

Germany.
The authors are grateful to Dr. J.-K. Hong of the Korea Institute [10] Shin D, Lee S, Shyam A, et al. Petascale supercomput-
of Materials Science (KIMS) and Dr. Y. Kim of KAMI Co. Ltd. ing to accelerate the design of high-temperature
for the sample fabrication throughout the project. alloys. Sci Technol Adv Mater. 2017;18:828–838.
[11] Kamath C. Data mining and statistical inference in
selective laser melting. Int J Adv Manuf Technol.
Disclosure statement 2016;86:1659–1677.
[12] Agrawal A, Deshpande PD, Cecen A, et al.
No potential conflict of interest was reported by the authors. Exploration of data science techniques to predict fati-
gue strength of steel from composition and proces-
sing parameters. Integr Mater Manuf Innov.
2014;3:1–19.
Funding [13] Mishra S, DebRoy T. Tailoring gas tungsten arc weld
This research was supported by the Industrial Strategic geometry using a genetic algorithm and a neural net-
Technology Development Program [10077677] and the work trained with convective heat flow calculations.
Technology Innovation Program [20000201] funded by the Mater Sci Eng A. 2007;454–455, 477–486.
Ministry of Trade, Industry and Energy (MOTIE, Korea). [14] Wang J, Nobakht AY, Blanks JD, et al. Machine
learning for thermal transport analysis of aluminum
alloys with precipitate morphology. Adv Theory
Simul. 2019;1800196:1–10.
ORCID [15] Shin D, Yamamoto Y, Brady MP, et al. Modern data
analytics approach to predict creep of high-temperature
Dongwon Shin http://orcid.org/0000-0002-5797-3423 alloys. Acta Mater. 2019;168:321–330.
[16] Shin D, Shyam A, Lee S, et al. Solute segregation at the
Al/θ′-Al2Cu interface in Al-Cu alloys. Acta Mater.
References 2017;141:327–340.
[1] Gorsse S, Hutchinson C, Gouné M, et al. Additive [17] Reshef DN, Reshef YA, Finucane HK, et al. Detecting
manufacturing of metals: a brief review of the char- novel associations in large data sets. Science.
acteristic microstructures and properties of steels, 2011;334:1518–1524.
Ti-6Al-4V and high-entropy alloys. Sci Technol Adv [18] Yu K, Tresp V, Schwaighofer A Learning Gaussian
Mater. 2017;18(1):584–610. processes from multiple tasks. Proceedings of the
[2] Zhao C, Fezzaa K, Cunningham RW, et al. Real-time 22nd International Conference on Machine Learning
monitoring of laser powder bed fusion process using (ICML-05); 2005; Bonn, Germany. p. 1012–1019.
high-speed X-ray imaging and diffraction. Sci Rep. [19] Zhang Y, Duchi J, Wainwright M. Divide and con-
2017;7:3602. quer kernel ridge regression: a distributed algorithm
[3] Khairallah SA, Anderson AT, Rubenchik A, et al. with minimax optimal rates. J. Mach. Learn. Res.
Laser powder-bed fusion additive manufacturing: 2015;16:3299–3340.
physics of complex melt flow and formation mechan- [20] Neter J, Kutner MH, Nachtsheim CJ, et al. Applied
isms of pores, spatter, and denudation zones. Acta linear statistical models. Vol. 4, Chicago: Irwin; 1996
Mater. 2016;108:36–45. [21] Cherkassky V, Ma Y. Comparison of model selection
[4] Bidare P, Bitharas I, Ward RM, et al. Fluid and par- for regression. Neural Comput. 2003;15:1691–1714.
ticle dynamics in laser powder bed fusion. Acta [22] Segal MR. Machine learning benchmarks and random
Mater. 2018;142:107–120. forest regression. Biostatistics. 2004;1–14.
[5] Matthews MJ, Guss G, Khairallah SA, et al. Denudation [23] Hsu C-W, Lin C-J. A comparison of methods for
of metal powder layers in laser powder-bed fusion multi-class support vector machine. IEEE Trans
processes. Acta Mater. 2016;114:33–42. Neural Netw. 2002;13:415–425.
[6] Leung CLA, Marussi S, Atwood RC, et al. In situ X-ray [24] Lee S, Shin D, Peng J ASCENDS: a data SCiENce
imaging of defect and molten pool dynamics in laser toolkit for non-data scientists; 2019. Available from:
additive manufacturing. Nat Commun. 2018;9:1355. https://github.com/ornlpmcp/ASCENDS.
[7] Ki H, Mohanty PS, Mazumder J. Modeling of laser [25] King WE, Anderson AT, Ferencz RM, et al. Laser
keyhole welding: part II. Simulation of keyhole evolu- powder-bed fusion additive manufacturing of metals;
tion, velocity, temperature profile, and experimental physics, computational, and materials challenges.
verification. Metall Mater Trans A Phys Metall Mater Appl Phys Rev 2. 2015;041304:1–26.
Sci. 2002;33A:1831–1842. [26] Lee YS, Nandwana P, Zhang W. Dynamic simulation
[8] Yadroitsev I. Selective laser melting: direct manufac- of powder packing structure for powder bed additive
turing of 3D-objects by selective laser melting of manufacturing. Int J Adv Manuf Technol. 2018;96
metal powders. Saarbrucken: LAP Lambert (1–4):1507–1520.
Academic Publishing; 2009. [27] Moges T, Ameta G, Witherell P. A review of model
[9] Rehme O, Emmelmann C Reproducibility for proper- inaccuracy and parameter uncertainty in laser powder
ties of selective laser melting products. Proceedings of bed fusion models and simulations. J Manuf Sci Eng.
the Third International WLT-Conference on Lasers 2019;141:1–14.