Studying the Topography of Laser Cut Aluminium Using Latent
Space Produced by Deep Learning
Alexander F. Courtier
1
, Matthew Praeger
1
, James A. Grant-Jacob
1
, Christophe Codemard
2
,
Paul Harrison
2
, Ben Mills
1
and Michalis N. Zervas
1
1
Optoelectronics Research Centre, University of Southampton, University Road, Southampton, SO17 1BJ, U.K.
2
TRUMPF Lasers UK, 6 Wellington Park, Toolbar Way, Hedge End, Southampton, SO30 2QU, U.K.
Keywords: Deep Learning, Laser Cutting, Topography, Convolutional Variational Autoencoders, Neural Networks,
Convolutional Neural Networks.
Abstract: Modelling topography resulting from laser cutting is challenging due to the highly non-linear light-matter
interactions that occur during cutting. We show that unsupervised deep learning offers a data-driven capability
for modelling the changes in the topography of 3mm thick, laser cut, aluminium, under different cutting
conditions. This was achieved by analysing the parameter space encoded by the neural network, to interpolate
between output topographies for different laser cutting parameter settings. This method enabled the use of
neural network parameters to determine relationships between input laser cutting parameters, such as cutting
speed or focus position, and output laser cutting parameters, such as verticality or dross formation. These
relationships can then be used to optimise the laser cutting process.
1 INTRODUCTION
Fibre laser cutting is a materials processing technique
with many applications in industry. It offers many
advantages over competing techniques in terms of
precision, speed, and mechanical stability. Defects
can however be formed during cutting that limit the
final quality of the cut. These defects include
striations, seen as systematic ridges along the cutting
edge, as well as welts, seen as random depressions
along the sample.
The interactions causing these defects are poorly
understood due to their non-linearity (Arai, 2014), so
determining their relationship to input parameters is
challenging. Deep learning enables a data driven
approach to studying laser machining processes, with
much interest shown in recent years (Courtier et al,
2021; McDonnel et al, 2021; Stadter et al, 2020; Mills
and Grant-Jacob, 2021). Unsupervised learning
enables the use of unlabelled laser cut topographies
from which neural networks can extract their own
mathematical models. This enables the use of a latent
parameter space to model the relationships between
laser cutting input parameters, such as the cutting
speed or the focus position, and laser cutting output
parameters, such as verticality or dross formation.
2 EXPERIMENTAL METHODS
Sixty-five 3 mm thick grade 1000 aluminium samples
were cut with a 4 kW continuous wave disk laser. The
workstation was a TRUMPF TruLaser 1030 flatbed
cutting machine with a Precitec ProCutter cutting
head, with a 2.0x magnification focusing objective
and using nitrogen as the co-axial assist gas. The focal
spot size was 210 µm. Edges were measured using
interferometric profiling on a SmartWLI Compact
topographic profiler (GBS) using a Nikon 5x
Michelson interferometric objective lens (CF IC EPI
Plan TI) giving 1.34 µm spatial resolution, 0.57 µm
depth resolution and 3.4 x 2.8 mm field of view.
The focus position is defined as the distance
between the focus of the laser and the sample surface,
and the standoff distance is defined as the separation
between the laser cutting head and the work piece.
Each of these parameters have a dependence on each
other, for example standoff distance will impact the
effect of gas pressure.
Courtier, A., Praeger, M., Grant-Jacob, J., Codemard, C., Harrison, P., Mills, B. and Zervas, M.
Studying the Topography of Laser Cut Aluminium Using Latent Space Produced by Deep Learning.
DOI: 10.5220/0011631400003408
In Proceedings of the 11th International Conference on Photonics, Optics and Laser Technology (PHOTOPTICS 2023), pages 49-51
ISBN: 978-989-758-632-3; ISSN: 2184-4364
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
49

Figure 1: Application of a neural network for measuring the latent space of laser cut topography showing a) a schematic of
the experimental measurement process for collecting training data, and b) the concept of the prediction neural network with
a demonstration using topographic data, which also shows a concept of calculating the difference between latent vectors for
different parameters.
3 RESULTS AND DISCUSSION
Fig. 1 a) illustrates the laser cutting process with an
example of an experimental topography. Fig 1 b)
shows the concept of a Convolutional Variational
Autoencoder (CVAE) both conceptually and using
topographies. I.e. The encoder neural network (NN)
learns to compress the information contained within
the topographic data into alower number of
dimensions, more compact, representation (the latent
space vector). The decoder NN learns to perform the
reverse operation, rebuilding the original topographic
data as accurately as possible from the latent space
vector. The figure also shows a method for calculating
the difference between two latent vectors which can
be used to identify the effects of and interpolate
between different laser cutting parameters.
Our key result is shown in Fig. 2 which
demonstrates that latent vector arithmetic can be used
to predict laser machining topographies. When an
experimental topography is fed into our encoder
network, the output is a 1D vector whose parameters
represent the latent space governing the appearance
of the topography. By averaging the latent vectors of
many topographies that were laser machined under
the same conditions, we can produce a latent vector
that is representative of defects that occur under those
conditions. Vector arithmetic can then be used to
combine these representative latent vectors in order to
predict topographies that would result from
intermediate conditions. As the latent vector
parameters are correlated to the input topography, the
latent space can be mapped to determine the linearity
of the relationship between laser cutting parameters
and laser cutting defects. By comparing vector
properties such as the equivalent angle between latent
vectors (in multi-dimensional space) or the difference
in resultant magnitude of latent vectors, it is expected
that relationships between laser cutting input and
output parameters can be determined. Results of this
analysis will be discussed in more detail at the
conference.
4 CONCLUSIONS
In conclusion, a CVAE was trained using
unsupervised learning to model the appearance of
laser cut grade 1000 aluminium. The resultant
latent vectors were then used to model the changes in
PHOTOPTICS 2023 - 11th International Conference on Photonics, Optics and Laser Technology
50

Figure 2: Diagram showing the use of latent vector arithmetic to model the changes of defects from cutting speeds of 1 m/min
to 5 m/min. In these simulated topographies (i.e. produced by the Decoder NN), the focus position was 2.0 mm, the standoff
distance was 4.0 mm and the gas pressure was 9 bar.
appearance using the average difference between
vectors for different cutting conditions. The novelty
in this approach is the use of unsupervised learning to
model the relationships between laser cutting input
parameters, such as the cutting speed or the focus
position, and laser cutting output parameters, such as
verticality or dross formation. These relationships
could then be used to optimise the laser cutting
process by predicting output topographies for given
cutting parameters and for predicting parameter limits
such as maximum or minimum cutting speeds.
REFERENCES
Arai, T. (2014). Generation of Striations During Laser
Cutting of Mild Steel. SOP Transactions On Applied
Physics, 2014(2), 81-95. https://doi.org/10.15764/
aphy.2014.02010
Courtier, A., McDonnell, M., Praeger, M., Grant-Jacob, J.,
Codemard, C., & Harrison, P. et al. (2021). Modelling
of fibre laser cutting via deep learning.
Optics Express, 29(22), 36487. https://doi.org/10.1364/
oe.432741
McDonnell, M., Arnaldo, D., Pelletier, E., Grant-Jacob, J.,
Praeger, M., & Karnakis, D. et al. (2021). Machine
learning for multi-dimensional optimisation and
predictive visualisation of laser machining. Journal Of
Intelligent Manufacturing, 32(5), 1471-1483. https://
doi.org/10.1007/s10845-020-01717-4
Mills, B., Heath, D., Grant-Jacob, J., & Eason, R. (2018).
Predictive capabilities for laser machining via a neural
network. Optics Express, 26(13), 17245. https://doi.
org/10.1364/oe.26.017245
Stadter, C., Schmoeller, M., von Rhein, L., & Zaeh, M.
(2020). Real-time prediction of quality characteristics
in laser beam welding using optical coherence
tomography and machine learning. Journal Of Laser
Applications, 32(2), 022046. https://doi.org/10.2351/
7.0000077.
Studying the Topography of Laser Cut Aluminium Using Latent Space Produced by Deep Learning
51