ANN-based modelling of dimensional accuracy in L-PBF

Abstract

Laser-based powder bed fusion (L-PBF) is one of the AM techniques that has continued to gain an increased market acceptance and penetration, particularly in a wide range of industrial applications that include automotive, aerospace, medical/dental and robotics. However, in spite of its unstoppable rise in popularity, the L-PBF process still poses some technological challenges that need addressing in order to improve the robustness and repeatability of parts produced. In particular, when compared to other conventional manufacturing technologies, AM and L-PBF in particular lags behind when it comes to being able to predict different quality marks of the parts produced, such as dimensional accuracy and surface quality. In this context, the aim within this paper is to implement a systematic approach to improve precision of printed parts through predictive process modelling and experimental-based study. The methodology of a comprehensive Design of Experiments (DoE) study to improve process knowledge of down-facing surfaces is presented along with the methodology used to generate the dataset with regards to dimensional inaccuracy. The acquired results are used to build process models based on Artificial Neural network (ANN). The prediction accuracy of the proposed model is discussed, and the feasibility of the proposed approach is demonstrated. The outcome of this paper helps understand the L-PBF method and the influence of the governing parameters to further develop high precision L-PBF processes.

Full text

Joint Special Interest Group meeting between euspen and ASPE
Advancing Precision in Additive Manufacturing
Ecole Centrale de Nantes, France, September 2019
www.euspen.eu
ANN-based modelling of dimensional accuracy in L-PBF
Amal Charles1, Ahmed Elkaseer1, 2, Mahmoud Salem1, 3, Lore Thijs4, Steffen Scholz1, 5
1 Institute for Automation and applied Informatics, Karlsruhe Institute of Technology, Karlsruhe, Germany
2 Faculty of Engineering, Port Said University, Port Said, Egypt
3 Faculty of Engineering, Ain Shams University, Cairo, Egypt
4 Direct Metal Printing Engineering, 3D Systems, Leuven, Belgium
5 Karlsruhe Nano Micro Facility, Hermann-von-Helmholtz-Platz 1, 76344 Eggenstein-Leopoldshafen, Germany
amal.charles@kit.edu
Abstract
Laser-based powder bed fusion (L-PBF) is one of the AM techniques that has continued to gain an increased market acceptance and
penetration, particularly in a wide range of industrial applications that include automotive, aerospace, medical/dental and robotics.
However, in spite of its unstoppable rise in popularity, the L-PBF process still poses some technological challenges that need addressing
in order to improve the robustness and repeatability of parts produced. In particular, when compared to other conventional
manufacturing technologies, AM and L-PBF in particular lags behind when it comes to being able to predict different quality marks of
the parts produced, such as dimensional accuracy and surface quality. In this context, the aim within this paper is to implement a
systematic approach to improve precision of printed parts through predictive process modelling and experimental-based study. The
methodology of a comprehensive Design of Experiments (DoE) study to improve process knowledge of down-facing surfaces is
presented along with the methodology used to generate the dataset with regards to dimensional inaccuracy. The acquired results are
used to build process models based on Artificial Neural network (ANN). The prediction accuracy of the proposed model is discussed,
and the feasibility of the proposed approach is demonstrated. The outcome of this paper helps understand the L-PBF method and the
influence of the governing parameters to further develop high precision L-PBF processes.
L-PBF; ANN; AM; dimensional accuracy; process parameters
1. Main section heading
Additive Manufacturing (AM) technologies represent a change
in the paradigm of the manufacturing Industry [1]. They are
considered a key enabling technology for the current trends of
automation and digitalisation within the manufacturing industry
known as Industry 4.0. Industry 4.0 comprises of various digital
technologies such as; the internet of things (IOT), Augmented
reality, virtual reality, autonomous robots, big data analytics,
machine learning and artificial intelligence (AI) [2].
AM technologies have seen increased adoption for industrial
production in recent years. Especially in the aerospace,
automotive and medical industry where the benefits of AM such
as design freedom, reduction in lead-time, supply chain
optimisation, mass customisation etc. are much appreciated [3-
5]. However, precision aspects of laser based powder bed fusion
(L-PBF) are still an area where there is considerable room for
improvement [6]. Therefore, there exists a need for effective
process modelling techniques for improving the predictability
and controllability of the process.
Previous research efforts have shown that the effect of L-PBF
process parameters on the quality of printed parts is
interdependent as well as non-linear and with a certain degree
of uncertainty and anisotropy [7, 8]. Therefore, it makes it
challenging to employ conventional modelling approaches for
this process. However, the usage of more computational
intelligence proves promising. The usage of Artificial Neural
Networks (ANN) is recognised as an effective tool to model,
identify and control non-linear processes.
ANNs are powerful tools that mimic the structures and
processes of biological neural systems. They are especially able
to process large input/output data sets and to visualise complex
non-linear associations within this data. Which then makes it
possible to create complex process models for the purpose of
optimisation and control.
More recently ANNs have been used to model many different
manufacturing and machining processes especially due to their
high degree of accuracy in prediction [9, 10]. They are also highly
suitable for processes that generate a large amount of
continuous data for big data analytics. Nevertheless, looking at
the literature, only few attempts to model the L-PBF process
using ANN can be found [11]. In this context, this paper presents
an attempt at using ANN for modelling of the L-PBF parts.
Especially with regards to the modelling of down-facing process
parameters to predict dimensional inaccuracy in down-facing
surfaces. Dimensional inaccuracies in down-facing surfaces are
primarily caused by the formation of dross. Dross formation

takes places due to the creation of a over-heated zone where
the laser scans over lose powder [12].
Following this introduction, the next section describes the
methodology that was followed for the printing of the parts, as
well as the method of analysis used for generating the data that
was used to train the ANN model. This is followed by a
description of the first results achieved in the training of the
model. Finally, conclusions derived from the results are
presented along with the direction of future work.
2. Methodology
2.1. Part fabrication
The test samples that were the focus of this study, were
designed to have a 45° overhanging surface with a thickness of
2.12 mm. The various process parameters that were considered
for these test pieces were four quantitative parameters (laser
power, scan speed, scan spacing and layer thickness) and one
qualitative (discrete) parameter (scan strategy). The various
levels of the process parameters can be seen in Table 1.
Table 1 List of process parameters and levels
Process Parameter
Levels
Laser power (W)
50, 90, 150, 210, 250
Scan speed (mm/s)
200, 465, 850, 1235, 1500
Scan spacing (μm)
50, 60, 75, 90, 100
Scan Strategy
Strips, Rectangular cells,
Hexagonal cells
Layer thickness (μm)
60, 90
The aforementioned process parameters were only varied
within the down-facing area of the part as show in Fig. 1. The
rest of the bulk of the part was printed using the same
parameters for all test pieces. This was done in order to make
sure that the thickness of the test pieces were solely affected by
the down-facing process parameters.
Figure. 1 Depiction of area where down-facing parameters were used
The parts were pre-processed on 3DXpert software and
printed using a 3DSystems ProX DMP320. The parts were printed
using the Ti6AL4V, which is a high strength Titanium alloy with
good mechanical and thermal properties [13]. All parts were put
through a stress relief heat treatment process before they were
removed from the build platform. This was done in order to
prevent any warpage.
2.2. Experimental design
The analysis of the thickness was carried out in order to
calculate the dimensional deviation as a result of the varied
process parameters. The measurements were obtained by
capturing an image of each printed part and performing an
image processing technique to measure the thickness. The
methodology of the image processing is described as follows.
The captured images are first grey-scaled and a threshold is
applied in order to detect the edges of the part. The developed
algorithm then scans the image vertically as well as horizontally
and extrapolates a straight line from the detected edge points.
Knowing the scale of the image makes it possible to calculate
precisely the distance between the lines, which gives the
thickness of the test pieces.
The error in the thickness was calculated by comparing the
measured value to the CAD design data. Thereby, providing the
data set fed into the ANN. The DoE can be seen in Table 2.
Table 2 Experimental design
Trial
Laser
Power (W)
Scan
Speed
(mm/s)
Scan
Spacing
(µm)
1
90
465
60
2
90
465
90
3
90
1235
60
4
90
1235
90
5
210
465
60
6
210
465
90
7
210
1235
60
8
210
1235
90
9
50
850
75
10
250
850
75
11
150
200
75
12
150
1500
75
13
150
850
50
14
150
850
100
15 - 24
150
850
75
Meaning the above table represents 24 samples printed using
one type of scanning pattern and layer thickness. The results for
dimensional error used for training the ANN consider 3 different
scanning patterns, which were fed into the ANN model as
discrete levels (-1, 0 and 1) and 2 layer thicknesses. Therefore, a
total of 144 (entailing 90 dataset and 9*6 replicates) samples
were fabricated and tested.
3. Results
3.1 Algorithm for Error Estimation for Metal AM
This section is devoted to the design of the Artificial Neural
Network (ANN) to model the relationship between the input and
output parameters based on the experimental results. ANN was
utilised in this study due to its high ability to tackle complex non-
linear problems, whether classification or regression problems,
which are very difficult to solve using other techniques. In this
paper, the proposed ANN model works to estimate the error of
metal additive manufactured parts. The data was split into input
parameters (features) data as “Laser Power, Scan speed, Scan
Spacing, Scan Strategy, and Layer Thickness” and label
(response) data as “Error%”.
The ANN Algorithm works as follows. Input data and label data
were stored into two matrixes 5x90 and 1x90, respectively.
Then, the entire data sets were divided randomly into training
data, validation data, and testing data. The proposed algorithm
was developed using the Neural Network toolbox in MATLAB.
Different ANN designs were tested and evaluated for the
minimum error between the expected and estimated results.
Accordingly, the proposed ANN was designed based on five

neurons for input layer, 1 hidden layer with three and one
neuron for the output layer as shown in Figure 2. The
Levenberg–Marquardt was selected as neural network
algorithm.
For the training phase of the ANN algorithm, random values
for initializing weight and bias of neural network, 1000 epochs
(training step) and 0.1 learning rate were set.
Figure. 2. Optimal structure of the proposed ANN
Further to the development of the training session, Figure 14.
The proposed training ANN model was updated with the newly
generated weights and bias that minimize the training error in
order to estimate the correct response “Errors%” of the 3D
printed parts.
As shown in Figure 3, the results of the training gradient was
calculated to be 0.05733 at epoch 11 and the best validation
results are at epoch 5 with a mean square error of 0.2855 as
shown Figure 5.
Figure. 3 Results of Training gradient
Figure. 4 Validation performance of neural network
Figure 5 shows the performance of the ANN algorithm
developed to predict the error of AM fabrication process. In
particular, the proposed ANN model successfully estimates the
error percentage in the MAM products for different input
parameters. This is clearly revealed considering the trend and
distribution of plotted data around the fit line in the training,
validation and testing phases of the proposed ANN model.
Figure. 5. Regression algorithm performance
4. Conclusions
The current work has utilised ANNs for predictive process
modelling of the L-PBF process. A DoE study was conducted and
the printed test pieces were measured for dimensional error %
using an image processing technique. The process parameters
and the measured error % were used to train a ANN. An ANN
model was then developed for the prediction of the error % of
the process based on the training data. This model was validated
and tested and it showed the best performance at a Mean
Squared Error of 0.2855 after 5 epochs. However, this current
ANN model is the result of a limited dataset for ongoing work. A
larger data sets are being prepped in order to train the ANN
model as it is even further. Therefore even more accurate
predictions are expected at the end of the current work, in order
to optimise the L-PBF using a genetic algorithm technique. The
results of this larger ANN training model and the identified
optimal processing parameters will be presented during the SIG
meeting after the work has been finalised. The current state of
the work along with the small error in prediction accuracy point
towards a promising direction for the usage of ANN for the
process modelling of L-PBF parts. Finally one can conclude that
developing of highly accurate predictive models for L-PBF
process are important in the context of greater adoption of AM
in the manufacturing industry. The application of artificial
techniques such as ANN to manufacturing processes have shown
much promise to develop precision, quality and predictability.
5. Acknowledgements
This work was done in the H2020-MSCA-ITN-2016 project
PAM2, Precision Additive Metal Manufacturing, which is funded
by The EU Framework Programme for Research and
Innovation—Grant Agreement No. 721383.
References
[1] M. Attaran, The rise of 3-D printing: The advantages of additive
manufacturing over traditional manufacturing, Business Horizons.
[2] J. Qin, Y. Liu, R. Grosvenor, A Categorical Framework of
Manufacturing for Industry 4.0 and Beyond, Procedia CIRP
52(Supplement C) (2016) 173-178.
[3] P. Liu, S.H. Huang, A. Mokasdar, H. Zhou, L. Hou, The impact of
additive manufacturing in the aircraft spare parts supply chain: supply
Predicted dimensional
inaccuracy
Predicted dimensional
inaccuracy
Predicted dimensional
inaccuracy
Predicted dimensional
inaccuracy
Actual measured dimensional
error
Actual measured dimensional
error
Actual measured dimensional
error
Actual measured dimensional
error

chain operation reference (scor) model based analysis, Production
Planning & Control 25(13-14) (2014) 1169-1181.
[4] C. Balletti, M. Ballarin, F. Guerra, 3D printing: State of the art and
future perspectives, Journal of Cultural Heritage 26 (2017) 172-182.
[5] D.R. Eyers, A.T. Potter, Industrial Additive Manufacturing: A
manufacturing systems perspective, Computers in Industry 92–93 (2017)
208-218.
[6] A. Witvrouw, J. Metelkova, R. Ranjan, M. Bayat, A. Charles, M.
Moshiri, J. Dos Santos Solheid, S. Scholz, M. Baier, S. Carmignato, P.
Stavroulakis, A. Shaheen, L. Thijs, D. De Baere, G. Tosello, Precision
Additive Metal Manufacturing, ASPE and euspen Summer Topical
Meeting 'Advancing Precision in Additive Manufacturing', Berkeley, CA,
USA, 2018.
[7] H. Shipley, D. McDonnell, M. Culleton, R. Coull, R. Lupoi, G. O'Donnell,
D. Trimble, Optimisation of process parameters to address fundamental
challenges during selective laser melting of Ti-6Al-4V: A review,
International Journal of Machine Tools and Manufacture 128 (2018) 1-
20.
[8] A. Charles, A. Elkaseer, L. Thijs, V. Hagenmeyer, S. Scholz, Effect of
Process Parameters on the Generated Surface Roughness of Down-
Facing Surfaces in Selective Laser Melting, Applied Sciences 9(6) (2019)
1256.
[9] S. Huang, H.-C. Zhang, Artificial neural networks in manufacturing:
Concepts, applications, and perspectives, 1994.
[10] A. Elkaseer, J. Lambarri, I. Quintana, S. Scholz, Laser Ablation of
Cobalt-Bound Tungsten Carbide and Aluminium Oxide Ceramic:
Experimental Investigation with ANN Modelling and GA Optimisation, in:
D. Dao (Ed.), Sustainable Design and Manufacturing 2018 – Proceedings
of the 5th International Conference on Sustainable Design and
Manufacturing (KES-SDM-18), Springer International Publishing, Cham,
CH2019, pp. 21-30.
[11] F.W. Baumann, A. Sekulla, M. Hassler, B. Himpel, M. Pfeil, Trends of
machine learning in additive manufacturing, International Journal of
Rapid Manufacturing 7(4) (2018) 310-336.
[12] A. Charles, A. Elkaseer, T. Müller, L. Thijs, M. Torge, V. Hagenmeyer,
S. Scholz, A Study of the Factors Influencing Generated Surface
Roughness of Downfacing Surfaces in Selective Laser Melting, World
Congress on Micro and Nano Manufacturing, Portoroz, SLO, September
18-20, 2018, Research Publishing Services, Singapore, 2018, pp. 327-330.
[13] J.E. Abu Qudeiri, A.-H.I. Mourad, A. Ziout, M.H. Abidi, A. Elkaseer,
Electric discharge machining of titanium and its alloys: review, The
international journal of advanced manufacturing technology 96(1-4)
(2018) 1319-1339.