Document [original]

Testing for Welding, Joining or Additive Manufacturing Applications

Angelina Marko, Stefan Bähring, Julius Raute*, Max Biegler and Michael Rethmeier

Transferability of ANN-generated parameter sets

from welding tracks to 3D-geometries in Directed

Energy Deposition

https://doi.org/10.1515/mt-2022-0054

Abstract: Directed energy deposition (DED) has been in

industrial use as a coating process for many years. Modern

applications include the repair of existing components

and additive manufacturing. The main advantages of DED

are high deposition rates and low energy input. However,

the process is inﬂuenced by a variety of parameters affecting

the component quality. Artiﬁcial neural networks (ANNs)

offer the possibility of mapping complex processes such as

DED. They can serve as a tool for predicting optimal process

parameters and quality characteristics. Previous research

only refers to weld beads: a transferability to additively

manufactured three-dimensional components has not

been investigated. In the context of this work, an ANN is

generated based on 86 weld beads. Quality categories (poor,

medium, and good) are chosen as target variables to

combine several quality features. The applicability of this

categorization compared to conventional characteristics is

discussed in detail. The ANN predicts the quality category

of weld beads with an average accuracy of 81.5%. Two

randomly generated parameter sets predicted as “good”

by the network are then used to build tracks, coatings, walls,

and cubes. It is shown that ANN trained with weld

beads are suitable for complex parameter predictions in a

limited way.

Keywords: additive manufacturing; artificial neural network;

DED; quality assurance; welding parameter.

1 Introduction

In recent years, an increasing number of applications for

metal-based additive manufacturing have been developed.

Apart from technologies that use a powder bed, directed

energy deposition (DED) is meanwhile established in the in-

dustry. This process enables a high deposition rate by melting

metal powders or wires with focused thermal energy while

being deposited. This way, a material can be added to sub-

strates and existing surfaces [1], components can be repaired

[2], and structures can be built for additive manufacturing [3].

In the powder-based process, 3D structures are built using a

combinationofanozzle-blownpowderandalaser.Single

weld beads are deposited onto each other in a series of layers

to create customized components. The bead shapes are

mainly inﬂuenced by the process parameters such as laser

power,scanspeed,andpowderfeedrate.Dependingonthe

machine that is used, other parameters such as spot diameter,

focus distance, and shield- as well as carrier-gas-ﬂow are also

of importance. Furthermore, build strategy, cooling time, and

preheating of the substrate inﬂuence the welding process. All

mentioned aspects must be considered and adjusted based

on the used machine and material to provide the best com-

bination for additive manufacturing.

One material that is widely used in DED is the nickel-

based alloy Inconel 718. It is characterized by a high corrosion

resistance at high temperatures (up to 1000 °C). Thus, it can

be used to produce parts for gas turbines and turbochargers

as well as highly stressed components in aerospace [4].

Despite its high potential, up until now, only a few studies

have investigated the inﬂuence of process parameters on

geometries produced with Inconel 718. A study by Corbin

et al. [5] deals with the inﬂuence of the parameters such as

*Corresponding author: Julius Raute, Füge- und Beschichtungstechnik,

Fraunhofer-Institut für Produktionsanlagen und Konstruktionstechnik

IPK, Pascalstraße 8-9, Berlin, Germany,

E-mail: [email protected]r.de. https://orcid.org/0000-0002-

0280-7778

Angelina Marko, Technische Universitat Berlin, Berlin, Germany,

E-mail: [email protected]om

Stefan Bähring and Max Biegler, Füge- und Beschichtungstechnik,

Fraunhofer-Institut für Produktionsanlagen und Konstruktionstechnik

IPK, Berlin, Germany, E-mail: [email protected]

(S. Bähring), [email protected] (M. Biegler)

Michael Rethmeier, Technische Universitat Berlin, Berlin, Germany;

Füge- und Beschichtungstechnik, Fraunhofer-Institut für

Produktionsanlagen und Konstruktionstechnik IPK, Berlin, Germany;

and Bundesanstalt für Materialforschung und –prüfung (BAM), Berlin,

Germany, E-mail: [email protected]

Materials Testing 2022; 64(11): 1586–1596

laserpower,powderfeedrate,workingdistance,andsub-

strate preheating on the geometry of a weld bead. This study

concludes that laser power has the greatest inﬂuence on bead

width, and that working distance has the greatest inﬂuence

on bead height as well as the angle of repose. In addition,

substrate preheating was found to increase the effects of

varying laser power on bead height and width. By applying

linear regression, the data were used to create empirical

models to predict weld bead geometry. Kistler et al. [6]

investigated the effect of the same parameters on the

microstructure, fusion zone morphology, and hardness of a

weld bead. Microstructure, hardness, and geometry of the

fusion zone were considered as quality characteristics. It

was shown that laser power and scan speed have a linear

effect on the width and area of the meltzone. In this case, the

greatest effect was achieved by the laser power. Preheating

the substrates increased the width of the fusion zone and

resulted in more uniform hardness throughout the welding

process. Bax et al. [7] dealt with the signiﬁcance of laser

power, powder feed rate, and scan speed on weld beads as

well. For this purpose, speciﬁc process parameter maps

were created based on experimental tests and a semi-

empirical correlation from ﬁndings in the literature. It can be

shown that some correlations are valid regardless of ma-

chines and materials, but not everything is generally

transferable. It is pointed out that the evaluation of indi-

vidual beads is the prerequisite for high-quality coatings

and 3D structures.

For the evaluation of weld beads, bead width and bead

height, as well as burn-in depth, are usually considered.

Several investigations deal with the dependencies of these

target variables. In general, there are two main approaches:

Statistical design of experiments (DoE) and the application

of artificial neural networks (ANNs).

Graf et al. [8] used a full factorial experimental design

to describe the inﬂuence of the parameters such as laser

power, spot diameter, scan speed, and powder feed rate on

both the bead width and height. It was demonstrated that

laser power has the most signiﬁcant effect on bead width. By

contrast, scan speed and powder feed rate were shown to

have a statistically signiﬁcant inﬂuence on the bead height.

The prognosis based on the empirical model indicates

that, especially with regard to the bead width, there is still

potential for optimization. Similar research was performed

by Ermurat et al. [9] using the Taguchi optimization method.

Laser focus distance, working distance, feed gas ﬂow rate,

and scan speed were varied to apply minimally dimen-

sioned weld beads. Again, it was shown that signiﬁcant

inﬂuences on the bead geometry, in this case, mainly the

laser focus distance, can be identiﬁed using DoE. Jinoop

et al. [10] used the DoE approach as a full factorial

experimental design to investigate the inﬂuence of process

parameters on the geometry and quality of thin-walled

buildups.Limitvaluesforlaser energy and powder feed rate,

both per unit length, were determined. By evaluating the

observed effects, an optimal process window for a defect-

free buildup could be identiﬁed.

These publications show that the focus of the statis-

tical experimental design is to identify the influence of

process parameters on defined target variables to increase

the process knowledge. By contrast, when using neural

networks, the focus is on predicting the target variable as

accurately as possible based on input data. Since this is a

black box approach, the effects of individual input pa-

rameters are usually not determined. Song et al. [11]

compared both tools for the prediction of geometric

characteristics in DED and concluded that the use of sta-

tistical DoE is appropriate if the prediction does not

require a high degree of accuracy, and in addition, low

costs have to be considered. Neural networks are advan-

tageous when the data points are not based on a statistical

experimental design, and accuracy requirements are

comparably high. The predictive ability of neural net-

works has already been extensively studied, as shown by

the following publications. Saqib et al. [12] investigated

the inﬂuence of laser power, scan speed, powder feed

rate,focallength,andfocuspositionontheweldgeom-

etry for P420 ﬁne-grain structural steel. The predictive

abilities of the analysis of variance and a neural network

were compared. This work shows that the best match

between the predicted and actual values is achieved by

the neural network. An accuracy of 96.3% between pre-

dicted and experimentally determined validation data

was achieved. The 32 data points used during the exper-

iment were generated by a central composite design. A

similar approach was taken by Mondal et al. [13], who also

compared the analysis of variance to neural networks.

The material used was a nickel-based alloy. Based on a

Taguchi plan, nine experimental points were generated to

predict the bead width and the penetration depth. The

prediction performance of the neural network was 95.3%

for the bead width and 96.9% for the penetration depth.

Within the context of another publication, the bead height

could be predicted with an accuracy of 72% [14]. Guo et al.

[15] investigated the inﬂuence of laser power, scan speed,

and powder feed rate on the geometry and hardness of a

weld bead. Based on a 34experimental design, 81 exper-

imental data points were generated using a cobalt alloy.

Contrary to other research, the neural network was

designed to predict optimal process parameters. For this

purpose, bead geometry and hardness were used as in-

puts,whereaslaserpower,scanspeed,andpowderfeed

A. Marko et al.: Transferability of ANN-generated parameter sets in DED 1587

rate served as outputs of the network. The trained neural

network achieved a predictive accuracy of around 90%.

Moreover, Feenstra et al. [16] used ANN to determine the

process window in DED, regardless of the material. For

this purpose, 297 beads consisting of three different ma-

terials(Inconel265,316L,and Hastelloy X) were pro-

duced.Theresultsareusedtovisualizerelationships

between volumetric energy density and the energy

required to melt a given amount of added powder.

The principal approach for the application of ANN is

mostly identical. The existing database is divided up into

datasets for training, validation, and testing. Initially, all

weights of the neurons as well as all biases are set randomly.

In every training iteration, the network calculates output

values based on the training inputs. The error between those

values and the actual training output values is determined,

and neuron weights and biases are updated to decrease the

error. During training, weights and biases are constantly

optimized until a certain termination criterion is met.

Hyperparameters such as network topology, learning rate,

and number of epochs can be adjusted using the validation

data. When the optimal hyperparameters are found and

the training is completed, the network accuracy is ﬁnally

validated using test data that was not used during training.

The described learning procedure is known as supervised

learning. Here, in addition to the input data, the provision of

output data is of particular importance [17]. On the one

hand, neural networks can be trained to predict numerical

values, such as bead width and bead height. On the other

hand, for classiﬁcation tasks, different weld bead categories

can be determined.

The state-of-the-art shows that with ANN, the prediction

of process windows in DED related to bead height and bead

width works reliably. Transferability from predicted values

of weld beads to three-dimensional additive structures

has not yet been investigated. In addition, an evaluation

of when the respective target values are suitable for prac-

tical applications is missing. One possibility to implement

this is a categorization system based on optical quality

assessment. In production, manufactured components are

often divided up into the categories “good”and “poor”to

identify rejects. Weld beads can be categorized with the

same principle. The machine operator usually tests different

parameter combinations and classiﬁes the weld beads

based on practical experience. For the additive buildup,

beads rated as “good”are selected and used to create 3D

structures. This procedure is mimicked in this work by using

ANN. The results of the networks are validated by a buildup

of different geometries. The aim is to test whether ANN that

has been trained on weld beads can specify suitable

parameter sets for complex geometries.

2 Experimental approach

2.1 Material

The nickel alloy Inconel 718 was used both as the substrate

and the powder material. The dimensions of the substrate

plate are 300 ×300 ×10 mm³.

2.2 Experimental setup

All experiments were carried out using a TRUMPF TruLaser

Cell 7020 with three translational and two rotational axes

and a 2 kW Yb:YAG disk laser with a wavelength of 1030 nm.

The mounted ring jet nozzle has an ideal working distance

of 9 mm to the substrate.

First, to provide a database for the neural network,

86 weld beads were deposited. After that, the network

was trained to predict the bead quality based on the given

process parameters. To evaluate the transferability of the

network predictions within the welding process, additional

beads, as well as coatings, walls, and cubes, were deposited

with new promising parameter sets proposed by the ANN.

The parameters for the 86 initial weld beads were

selected based on the process window shown in Table 1.

Accordingly, laser power P, scan speed v, powder feed rate

m, and working distance w were varied. For this purpose, a

full factorial test plan based on the process limits was applied

and expanded by additional randomly chosen points.

Next, the weld beads were visually assessed and cate-

gorized into the categories “good”,“medium”,and“poor”.

Figure 1 provides an explanation of this categorization.

3 Design of the ANN

Training and evaluation of the ANN was carried out using

the programming language Python in combination with the

machine learning program library PyTorch. Before being fed

into the network, the input data were normalized to ensure

an even distribution and a uniform range of values. Apart

from that, the data were divided up into training, validation,

Table :Process window.

Lower limit Upper limit

Laser power, P  W W

Scan speed, v  mm min− mm min−

Powder feed rate, _

m.g min− g min−

Working distance, w mm  mm

1588 A. Marko et al.: Transferability of ANN-generated parameter sets in DED

and test data to prepare for the training phase and to

ensure a proper evaluation of the prediction accuracy. In

this study, a percentage ratio of 70/10/20 was chosen,

resulting in 60 beads in the training set, 9 beads in the

validation set, and 17 beads in the test set. The parameter

sets of four values, according to Table 1, were used as

input, whereas the quality category of the weld bead rep-

resented the output value (good, medium, and poor). By

trial-and-error method, the number of neurons in the hid-

den layer was set to ﬁve. The topology of the neural

network is visualized in Figure 2.

The simple and widespread ReLU (rectified linear unit)

function was chosen as the activation function of the

neurons. It determines if and how a neuron passes on an

incoming signal. The ReLU function sets the output as the

given input if it is positive. Otherwise, the output is set to

0 [18]. With cross-entropy loss, a common loss function

suitable for classiﬁcation tasks was applied during training

to calculate the error of the respective quality prediction.

For this purpose, the ideal probability distribution of each

class, for example, [1, 0, 0] if the example belongs to the

ﬁrst class, and the actual probability distribution with

numbers between 0 and 1, for example, [0.7, 0.2, 0.1] are

compared and processed. Moreover, the Adam optimizer

was used as the optimization algorithm to update the

weights and biases of the neurons as it is fast and applicable

for simple classiﬁcation tasks [19].

3.1 Transferability on other geometries

Two new randomly generated parameter combinations that

had been classified as “good”by the ANN were used to build

up more complex geometries. This was done to examine the

transferability of such parameter sets on coatings, walls,

and cubes. The respective buildup strategies of these

geometries are shown in Figure 3.

For the walls, the scan direction was changed after

every layer. While building up the cubes, the layer pattern

was rotated by 90°after every layer to vary the starting

points and to improve the inner structure of the cube. The

values of both parameter sets are listed in Table 2.

4 Results and discussion

4.1 Process window and database

All 86 weld beads are shown in Figure 4. Since these beads

serve as the basis for the neural network, the aim should be

Figure 2: Topology of the ANN.

Figure 1: Categorization of welding beads.

Figure 3: Buildup strategies of the geometries.

A. Marko et al.: Transferability of ANN-generated parameter sets in DED 1589

to achieve a relatively even distribution among the categories

good, medium, and poor. A total of 31 good, 26 medium, and

29 poor beads were generated. Of the poor beads, 11 did not

show sufﬁcient attachment to the substrate.

In the literature, the process window for welding

beads using DED has already been extensively studied. The

comparison with the beads generated in this work conﬁrms

the previous ﬁndings. So far, the existing research has

mostly concentrated on ﬁndings regarding the extent to

which the parameters affect the bead height, bead width,

and penetration. These target parameters are usually

considered separately from each other. For example, Corbin

[5] and Graf [8] showed that laser power affects the bead

width especially. This relationship is due to a larger melt

pool if the power input increases. In the case of bead height,

the powder feed rate has a signiﬁcant effect [8]. The more

powder is available for the melting process, the higher

the structure of the individual beads. However, it should

be noted that there is an interaction with the laser power:

if there is not enough energy to melt the powder, in

some cases, no sufﬁcient melt pool can be created on the

substrate. Thus, bonding of the bead is prevented. These

ﬁndings are necessary for the basic process understanding.

However, it is difﬁcult to estimate a process window without

any information about when a bead is well or poorly suited

for practical applications. In particular, a combined consid-

eration of all target variables to provide a suitable process

window for the user is not available.

One way of combining the target variables is categorizing

the beads. In practice, the operator often makes an

experience-based decision about whether a weld bead is

“good”or not. Generally, various factors can be included in

this evaluation. In the context of this work, a visual assess-

ment was carried out to indicate whether a welding bead is

suitable for further applications. Here, the geometric shape

and tarnish colors, which generally indicate an excessive

energy input, were considered (Figure 1). A correlation of

three optical weld bead categories to the four varied process

parameters is shown in Figure 5.

Thus, it appears that laser power has a decisive influence

on the categorization. Good beads could only be generated

from about 400 W upward, and a low energy input led to

poor bead quality. It is also clear that if the laser power input

is too high, the beads tend to receive worse ratings. However,

there is no clear limit to this effect. Rather, it is strongly

related to the scan speed and the powder feed rate. This can

be explained by the correlation between energy input and

thepowdermasstobemelted(Figure5).

Within the scope of this work, it has been shown that an

energy per unit length of at least 25 kJ m−1usually produces

good beads. This conﬁrms the observations known from prior

experiments [5]. Furthermore, it can be stated that a visual

assessment gives a good representation of the combination

of bead height and bead width. The powder feed rate also has

aclearinﬂuence on the evaluation of beads. The assumption

that too much powder does not ensure a good application is

conﬁrmed. A powder feed rate of 3.5 g⋅min−1to 10.5 g⋅min−1

tends to yield favorable results. However, especially with

higher powder ﬂow rates, the quality of the bead depends on

other factors, such as the laser power or the energy per unit

Table :Predicted parameters.

Parameter P

,prog

Parameter P

,prog

Laser power, P  W W

Scan speed, v  mm min− mm min−

Powder mass ﬂow, _

mg min−.g min−

Working distance, w  mm mm

Overlap, O %%

Figure 4: Overview of all welding beads.

1590 A. Marko et al.: Transferability of ANN-generated parameter sets in DED

length. Nevertheless, it is generally true that higher powder

quantities lead to smaller melt pools, since the introduced

powder attenuates the effective laser power, as Bax [7]

showed in his investigations. The scan speed has a direct

effect on the bead quality as well. Once more, the justiﬁcation

is based on the energy per unit length. For instance, if the

scan speed is too low, excessive heat input occurs locally.

Furthermore, velocities that are too high result in poor beads.

With an unfavorable scan speed, the welding beads either

get too voluminous or too smooth. In contrast to the work

of Corbin [5], no signiﬁcant inﬂuences on the bead height

and thus ultimately on the optical quality evaluation could

be detected regarding the working distance. This can be

observed by the fact that all quality categories are arbi-

trarily represented over the entire varied spectrum of the

parameter. The different observations are likely attribut-

able to the additional factor of powder feed rate, which

wasconsideredinthiswork.Itexertsagreaterinﬂuence

on the buildup of the bead than the working distance and

the associated inﬂuence on the powder ﬂow. However,

based on Corbin’sﬁndings, it is assumed that the factor

can still inﬂuence the additive buildup.

In summary, it was shown that the categorization of

weld beads reflects the process knowledge that exists to

Figure 5: Correlation between single parameters and categorization.

A. Marko et al.: Transferability of ANN-generated parameter sets in DED 1591

date. Process parameters declared as significant based on

prior experiments are also highly relevant for an optical

quality evaluation. Thus, using categorical quality classes as

theoutputofanANNcanbeconsideredreasonable.

Furthermore, it becomes clear that the process window is

limited by the process parameters. At the same time, the three

quality classes must tend to be equally represented in the

datasettocreateabalancedinputfortheANN.Ifonecate-

gory is overrepresented in the training data, the network

learns that the probability of this very category is higher and

the prediction accuracy of the network in a real application

scenario decreases. In addition, the parameter combinations

should not be too narrow to avoid the network from receiving

duplicate information for almost the same data point. Thus, it

can be concluded that only a limited number of data points

are available when using ANN in DED. Generally, it can be

assumed that such a limited database, as often also found in

the literature, is too small to ensure a high informative value

of the network.

4.2 Results of the ANN

In the first step, an ANN was created and trained according

to the speciﬁcations. After the training was completed, the

prediction accuracy was evaluated. For the evaluation, it was

considered that the accuracy of the network depended on the

random composition of the training and test data set. For this

reason, the mean accuracy value over several runs was

determined. After 20 runs, the mean prediction accuracy on

the test data was 81.5%. Thus, approximately 14 out of 17

beads were categorized correctly on average. All incorrect

categorizations were off by just onecategory.Theindividual

accuracies of the runs are shown in Figure 6.

The ANN trained within the scope of this work reached an

average accuracy of 81.5% regarding the bead quality based

on four input variables, such as laser power, scan speed,

powder feed rate, and working distance. The predictive ac-

curacy is significantly lower than that in similar use cases in

the literature. There are several reasons for this.

First, it should be noted that previous studies always

chose numeric characteristics to perform regressions with

ANN. For instance, most researchers predicted bead height

and bead width. Predictions of quantitative characteristics

of welding beads are usually easier to make because the

target values are distributed in a small range. Qualitative

assessments, as performed in the present work, are not

always unambiguous. Different operators may already

have divergent opinions about whether a welding bead is

too narrow or too flat for a specific application. Therefore,

their quality evaluation might vary between “good”

and “sufﬁcient”or “sufﬁcient”and “poor”.Makingaclear

assessment is difﬁcult even for experienced operators.

Accordingly, the error potential in a corresponding prediction

by an ANN is also high.

Another reason for the strongly deviating results of

prediction accuracy in the different runs is the limited amount

of data. In the studies regarding the applicability of ANN in

DED, the findings are generally based on a very small data-

base. Often, data were obtained based on a statistical

experimental design and then evaluated by an ANN. The

added value compared to statistical evaluation, for example,

by ANOVA or regression analysis, is not apparent here. With a

systematic data composition, the ANN recognizes, depending

on the topology, a very similar mathematical model to the one

generated by statistic evaluations. In addition, the creation of

a model via ANN has two disadvantages: The mathematical

model is not visible, since ANN function as a black box. Only

the result is presented, but not the calculation path. aThere-

fore, the impact of individual parameters cannot be assessed.

Accordingly, statistical models offer the essential advantage

of generating the process knowledge. Feenstra et al. already

pointed out the difﬁculties regarding the accuracy of neural

networks [16]. Despite having a nominal target size, the bead

Figure 6: Accuracy of the ANN at 20 runs.

1592 A. Marko et al.: Transferability of ANN-generated parameter sets in DED

widthcouldonlybepredictedwithanaccuracyof72–78%.

The reasons were seen on the one hand in the number of

data points and on the other hand in the complexity of the

parameters and their interactions with each other. This can be

conﬁrmed by the ﬁndings in the present work. Furthermore,

Song et al. [11] compared the predictive capabilities of

statistical DoE and ANN and concluded that the use of DoE is

justiﬁed, especially if the accuracy of the prediction does

not need to be very high and the system has to be designed

cost-effectively. By contrast, the use of ANN is reasonable if

the accuracy requirements are high, and if the database is not

based solely on a statistical experimental design. This can be

conﬁrmed by the present work. The data basis was also

initially built based on a full factorial experimental design,

but only through the extension by random experimental

points, the ANN could be trained properly. Especially for

categorization, it must be considered that enough data points

from all different categories are available during the training.

The data basis is decisive for the informative value of the

network. Conducted experiments show that the accuracy is

about 20% higher if an even distribution of categories in the

training data is ensured. This ﬁnding can also be applied to

quantitative target variables: If the numerical values of the

target variable are too close together in the training data, it

is difﬁcult for the network to correctly predict outliers in

practical application.

The importance of the data basis can be extrapolated

from the accuracy distribution in Figure 6. All networks were

trained and tested with the same topology on the same

database. However, the data points for training and testing

were randomly selected in each case. This has an enormous

inﬂuence on the accuracy of the neural network. Runs

showing correct predictions for almost all beads were

observed. Nevertheless, the capability of the network is

limited. If irregularities occur in the process, the probability

is high that the network will not be able to identify them. For

this reason, a general estimate of the capability of a network

is only possible if different runs are evaluated, and if

an average value is calculated. If the accuracy is too low,

optimization measures must be taken regarding network

topology, boundary conditions of the network, and data

basis.

4.3 Transferability to complex geometries

Beads, coatings, walls, and cubes were created based on

the new parameter combinations predicted as “good”by

the ANN. The results are shown in Figure 7.

The results show that a parameter set cannot be trans-

ferred to different geometries without further adjustment.

Although both parameter combinations predicted as “good”

work well for beads, it must be considered for coatings

that the overlap factor is added as an unknown influence.

In the present case, this parameter was chosen freely

based on previous findings. It is known that a desirable

overlap ratio is between 30 and 50%. Depending on the

bead geometry, it must be varied accordingly within this

interval. Due to the missing information in the training

data, quality predictions for coatings or other geometries,

such as cuboids, are difficult.

In the present case, however, it was shown that with the

selected overlap ratio, the parameter sets can be transferred

to coatings. This depends on, among other things, the fact

that the shape of the bead is decisive for good coatings.

Bax et al. have shown this already in their research [7].

The predicted beads have a uniform structure with good

substrate connection and withouttarnish. It can be observed

that the bead height of the parameter P

2,prog

is signiﬁcantly

lower than that of P

1,prog

. The bead height is especially

important in additive manufacturing. This can be observed

well for the geometries such as wall and cube. The walls are

judged from sufﬁcient to poor.

In the case of P

2,prog

,theinsufﬁcient bead height is

reﬂected in the wall structure. With 40 welded layers, a

height of 7 mm was achieved. It can be concluded that the

parameter set is not suitable for building wall structures. By

contrast, P

1,prog

yields more favorable results. An additive

structure with 40 layers achieved a wall height of 20 mm. In

the optical quality evaluation, however, it is only classiﬁed as

“medium”: The buildup shows a wavy contour and strong

powder adhesion. Accordingly, an ideal wall buildup could

not be achieved with any of the parameter sets predicted as

“good”and suitable for beads by the ANN. The optimal

parameter set for a wall structure depends on various factors.

In particular, the laser energy and the powder feed rate per

unit length are decisive, as shown in the work by Song et al.

[11].

They stated that these factors ideally range from

105 kJ m−1to 210 kJ m−1and 4 g m−1to 12.5 g m−1, respectively,

for an ideal buildup of thin walls using Inconel 718 material.

Such speciﬁcﬁndings are generally not easily transferable.

Especially, the laser beam systems and the materials used

play an important role [7]. However, the values certainly

provide an orientation range. In the case of the predicted

parameter sets of this work, the laser energy introduced per

unit length is signiﬁcantly lower. In detail, a laser power of

85 kJ/m is introduced with P

1,prog

and about 49 kJ/m with

2,prog

. This is well below the optimum range. Thus, it can

be concluded that the energy introduced is too low for an

optimum wall structure. The powder feed per unit length, by

contrast, shows that the predicted parameters of 13.3 g⋅min−1

A. Marko et al.: Transferability of ANN-generated parameter sets in DED 1593

1,prog

) and 5.48 g⋅min−1(P

2,prog

) are within or close to the

deﬁned optimum parameters. This factor describes whether

sufﬁcient powder is available for the ideal buildup. In

particular, it plays an essential role in the appearance of a

weld bead. Therefore, it can be explained that the powder

feeds per unit length show a better match with the optimum

range than the combination of laser power and scan speed.

It can also be deduced well as to why the thin-walled

structure could not be successfully implemented. For future

developments, it is therefore conceivable to provide such

additional information to the network as an additional input.

The buildups of the cubes show similar results: an

optimum structure was not achieved with either of the two

parameter sets. In the case of cube 1, an excessive amount of

energy and thus an excessive amount of heat is introduced.

This leads to strong tarnish and shape deviation. Regarding

the number of layers, it can be observed that a significantly

higher application rate is achieved with a lower number of

layers. Cube 1 was buildup with 20 layers and reached a

height of 16 mm. By contrast, cube 2 was deposited with 40

layers and has a height of only 10 mm. In addition, cube 2

also shows signiﬁcant elevations at the corner points. This

can be explained mainly by the buildup strategy. In the

present investigations, the pattern was rotated by 90°for

every new layer. It is advisable to build frames between

layers to avoid strong shape deviations, especially at the

corner points. Generally, it is important to bear in mind that

with complex components, an optimum buildup strategy

tailored to the geometry is necessary for achieving satis-

factory results in practice. In addition, cooling times and

preheating and postheating of the substrate play a decisive

role in a successful buildup. However, the ANN cannot take

any of this information into account for classiﬁcation tasks.

In conclusion, the prediction of welding parameters for

additive manufacturing by DED based solely on welding

beads does not work consistently. In additive manufacturing,

the beads are no longer buildup on a flat surface of the

substrate but are strongly influenced by the shapes of the

Figure 7: Transferability of the predicted parameters to complex geometries.

1594 A. Marko et al.: Transferability of ANN-generated parameter sets in DED

previous beads and layers. In this case, the buildup strategy

plays a decisive role to compensate for such unevenness.

To apply ANN successfully, the networks require more in-

formation. However, it must be considered that for any

additional information as input, the number of required data

points increases accordingly.

5 Conclusions

In this work, the prediction ability of an ANN that was trained

with welding beads was tested on different geometries. To

generate training data for the network, weld beads were

deposited and then classified into quality categories. Two

new random parameter sets classified as “good”by the

trained ANN were then transferred to more complex geome-

tries. The following conclusions can be drawn:

–ANN that are trained with welding beads can be applied

for parameter prediction in DED in a limited way. Process

understanding of bead deposition can be increased, and

an optimal process window can be identiﬁed. However,

the process conditions for coatings and especially for

Additive Manufacturing (AM) applications are signiﬁ-

cantly more complex, so that transferability is only

possible to a limited extent. Nevertheless, the knowledge

of optimized weld bead parameters is an important

prerequisite for a stable and economical process. It is

therefore recommended to investigate beads through

DoE to identify signiﬁcant inﬂuences. Based on this

knowledge, AI methods can be used to efﬁciently

implement quality assurance in DED. For instance,

quality-relevant characteristics such as mechanical

properties can be predicted to reduce downstream

inspections.

–The selection of an optimal buildup strategy is a ma-

jor challenge in additive manufacturing and thus, an

obstacle to the use of ANN. A variety of possible ap-

proaches are used with consideration for the geometry

and requirements for the mechanical characteristic

values. The ANN in their present form trained with

currently available data sets are not able to master this

complexity.

–One possibility to deal with the complexity of coatings

and AM applications and thus to better train the ANN is to

extend the data basis. Up to now, beads have always

been used as the data basis in the literature. A new

approach would be to use coatings or cubes. This way,

for instance, parameters such as overlap ratio and cool-

ing times could serve as an additional input for the ANN.

–Furthermore, it is possible to generate measurement

data during the process and feed that information to

the ANN. Today, many DED machines already have basic

sensor systems for process monitoring. For example, the

average temperature or melt pool size could contain

important information to signiﬁcantly improve quality

assessments by ANN even for complex geometries.

Author contributions: All the authors have accepted

responsibility for the entire content of this submitted

manuscript and approved submission.

Research funding: This research was funded by the German

federal ministry for economic airs and energy (BMWi)

through the industrial cooperative research association

(AiF) and the German Welding Association (DVS) under the

project entitled “Certify as you build –Quality assurance for

the directed energy deposition”with grant number of IGF

20.612 N. The authors would like to thank the BMWi, AiF,

and DVS for their support.

Conﬂict of interest statement: The authors declare no

conﬂicts of interest regarding this article.

References

[1] W. E. Frazier, “Metal additive manufacturing: a review,”J. Mater.

Eng. Perform., vol. 23, no. 6, pp. 1917–1928, 2014, https://doi.

org/10.1007/s11665-014-0958-z.

[2] B. Graf, A. Gumenyuk, and M. Rethmeier, “Laser metal deposition

as repair technology for stainless steel and titanium alloys,”

Phys. Procedia, vol. 39, pp. 376–381, 2012, https://doi.org/10.

1016/j.phpro.2012.10.051.

[3] T. Petrat, B. Graf, A. Gumenyuk, and M. Rethmeier, “Laser metal

deposition as repair technology for a gas turbine burner made of

Inconel 718,”Phys. Procedia, vol. 83, no. 2016, pp. 761–768,

2016, https://doi.org/10.1016/j.phpro.2016.08.078.

[4] E. Hosseini and V. A. Popovich, “A review of mechanical

properties of additively manufactured Inconel 718,”Addit.

Manuf., vol. 30, no. 100877, 2019, https://doi.org/10.1016/j.

addma.2019.100877.

[5] D. J. Corbin, A. R. Nassar, E. W. Reutzel, A. M. Beese, and

N. A. Kistler, “Effect of directed energy deposition processing

parameters on laser deposited Inconel® 718: external

morphology,”J. Laser Appl., vol. 29, no. 2, pp. 1–12, 2017.

https://doi.org/10.2351/1.4977476.

[6] N. A. Kistler, A. R. Nassar, E. W. Reutzel, D. J. Corbin, and

A. M. Beese, “Effect of directed energy deposition processing

parameters on laser deposited Inconel® 718: microstructure,

fusion zone morphology, and hardness,”J. Laser Appl., vol. 29,

no. 2, p. 22005, 2017, https://doi.org/10.2351/1.4979702.

[7] B. Bax, R. Rajput, R. Kellet, and M. Reisacher, “Systematic

evaluation of process parameter maps for laser cladding and

directed energy deposition,”Addit. Manuf., vol. 21, pp. 487–494,

2018, https://doi.org/10.1016/j.addma.2018.04.002.

[8] B. Graf, S. Ammer, A. Gumenyuk, and M. Rethmeier, “Design of

experiments for laser metal deposition in maintenance, repair

and overhaul applications,”Procedia CIRP, vol. 11, pp. 245–248,

2013, https://doi.org/10.1016/j.procir.2013.07.031.

A. Marko et al.: Transferability of ANN-generated parameter sets in DED 1595

[9] M. Ermurat, M. A. Arslan, F. Erzincanli, and I. Uzman, “Process

parameters investigation of a laser-generated single clad for

minimum size using design of experiments,”Rapid Prototyp. J.,

vol. 19, no. 6, pp. 452–462, 2013, https://doi.org/10.1108/rpj-

06-2011-0062.

[10] A. N. Jinoop, C. P. Paul, S. K. Mishra, and K. S. Bindra, “Laser

additive manufacturing using directed energy deposition of

Inconel-718 wall structures with tailored characteristics,”

Vacuum, vol. 166, pp. 270–278, 2019, https://doi.org/10.1016/j.

vacuum.2019.05.027.

[11] C. Song, L. Liu, Y. Yang, and C. Weng, “Prediction on geometrical

characteristics of laser energy deposition based on regression

equation and neural network,”IFAC-PapersOnLine, vol. 53, no. 5,

pp. 89–96, 2020, https://doi.org/10.1016/j.ifacol.2021.04.085.

[12] S. Saqib, R. J. Urbanic, and K. Aggarwal, “Analysis of laser

cladding bead morphology for developing additive manufacturing

travel paths,”Procedia CIRP, vol. 17, pp. 824–829, 2014, https://

doi.org/10.1016/j.procir.2014.01.098.

[13] S. Mondal, B. Tudu, A. Bandyopadhyay, and P. K. Pal, “Process

optimization for laser cladding operation of alloy steel using

genetic algorithm and artiﬁcal neural network,”Int. J. Comput.

Eng. Res., vol. 2, no. 1, pp. 2250–3005, 2012.

[14] S. Mondal, A. Bandyopadhyay, and P. K. Pal, “Application of

artiﬁcial neural network for the prediction of laser cladding

process characteristics at Taguchi-based optimized

condition,”Int. J. Adv. Manuf. Technol.,vol.70,nos.9–12,

pp. 2151–2158, 2014, https://doi.org/10.1007/s00170-013-

5393-z.

[15] S. Guo, Z. Chen, D. Cai, Q. Zhang, V. Kovalenko, and J. Yao,

“Prediction of simulating and experiments for Co-based alloy

laser cladding by HPDL,”Phys. Procedia,vol.50,pp.375–382,

2013, https://doi.org/10.1016/j.phpro.2013.11.058.

[16] D. R. Feenstra, A. Molotnikov, and N. Birbilis, “Utilisation of

artiﬁcial neural networks to rationalise processing windows in

directed energy deposition applications,”Mater. Des., vol. 198,

no. 109342, pp. 1–16, 2021. https://doi.org/10.1016/j.matdes.

2020.109342.

[17] R. Sathya and A. Abraham, “Comparison of supervised and

unsupervised learning algorithms for pattern classiﬁcation,”Int.

J. Adv. Res. Artif. Intell., vol. 2, no. 2, pp. 34–38, 2013. https://

doi.org/10.14569/IJARAI.2013.020206.

[18] M. Goyal, R. Goyal, P. Venkatappa Reddy, and B. Lall, “Activation

functions,”in Deep Learning: Algorithms and Applications,

Cham, Springer, 2020.

[19] D. P. Kingma and J. Ba, “Adam: a method for stochastic

optimization,”in Proc. of 3rd International Conference for

Learning Representations, San Diego, California, USA, 2015.

The authors of this contribution

Angelina Marko

Angelina Marko, born in 1987, receivedher M.Sc. degreein production

technology from the TU Berlin in 2015. After some practical experience

as a quality engineer in the industry, she is a scientific assistant at the

Fraunhofer Institute for Production Systems and Design Technology

IPK since 2016. Her focus is on laser metal deposition and quality

assurance for additive manufacturing.

Stefan Bähring

Stefan Bähring, born in 1996, is currently studying at the TU Berlin for

a master’s degree in production engineering with a specialization in

automation and information technology. Since 2020, he has been

working at the Fraunhofer IPK. His main focus is on quality assurance

in directed energy deposition using artiﬁcial intelligence.

Julius Raute

Julius Raute., born in 1993, studied Production Technology at the TU

Berlin and received his master’s degree in 2019. Since 2020, he has

been working as a researcher in the department of Joining and Coating

Technology at the Fraunhofer IPK,Germany. Thefocus of his work is on

laser-based directed energy deposition and electron beam welding of

Ni-based superalloys.

Max Biegler

Max Biegler, born in 1989, finished his studies in mechanical

engineering at the Technical University of Munich in 2015. For his

Doctoral Studies, he focused on the numerical modeling of welding

processes, including resistance spot welding and additive

manufacturing. Currently, he is the Head of the Department of Joining

and Coating Technology at the Fraunhofer IPK in Berlin.

Michael Rethmeier

Prof. Dr.-Ing. Michael Rethmeier, born in 1972, studied mechanical

engineering at the TU in Braunschweig, Germany. Afterward, he

worked at the same university, where he received his Ph.D. in 2003

and then became a project manager for production engineering and

concepts at the Volkswagen AG group research. In 2007, he got his full

professorship at the TU of Berlin in combination with being head of the

division “Welding Process Technologies”at the Federal Institute for

Materials Research and Testing BAM in Berlin. Additionally, he is the

division director of “Joining and Coating Technology”of the

Fraunhofer Institute for Production Systems and Design Technology

IPK in Berlin.

1596 A. Marko et al.: Transferability of ANN-generated parameter sets in DED