Available online at www.sciencedirect.

com

ScienceDirect
Procedia CIRP 113 (2022) 410–416
www.elsevier.com/locate/procedia

ISEM XXI

Profile prediction in ECM using machine learning

Ming Wu1, Muhammad Hazak Arshad1, Krishna Kumar Saxena1, Jun Qian1, Dominiek Reynaerts1*

1
Department of Mechanical Engineering, KU Leuven & Members Flanders Make, Leuven, Belgium

* Corresponding author. E-mail address: dominiek.reynaerts@kuleuven.be

Abstract

ECM (Electrochemical Machining) with a micro-second pulsed power supply can decrease stray current corrosion and thus provide better surface
quality on top of higher machining accuracy. However, the use of a pulsed power supply makes it challenging to predict the final workpiece
profile based on the typical multi-physical-based FEM(Finite Element method) models. First, it is impractical to introduce high-frequency pulsed
electric current (ns - µs) into a multi-physical model because of time step limitations. Second, there arise problems of compatibility of the time-
steps within each physical module, as e.g. hydrogen bubbles have a lifetime of ms, the power supply would use µs, and thermal phenomena take
ms. Multi-scale models have been proposed, but the prediction accuracies are rather low and computation times are very long. In this article,
based on machine learning approaches, we present 3 data-driven ECM models for predicting the final workpiece profile when using pulsed-ECM:
the linear regression (LR) model, the neural network (NN) model, and the convolutional neural network (CNN) model. After training this data-
driven ECM model with different levels of pulse voltage and electrolyte flow conditions, predictions and experimental validation are conducted.
Experiments with parameters outside the training parameter window are also carried out to show the performance and general applicability of
our data-driven ECM model. The machine learning model shows a good generalizing ability, the CNN model presents a prediction MSE of 7.60.
The present results also demonstrate that more accurate predictions will be achieved when using in-processing data. Hence, the prediction
accuracy of data-driven models can be further improved on top of advancing in-processing monitoring systems.
© 2022 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the ISEM XXI

Keywords: Electrochemical micromachining; pulse ECM; machine learning

Nomenclature
1. Introduction
ECM Electrochemical Machining
FEM Finite Element Method With the advances in information technology, computational
IEG Inter electrode gap methods, and advanced sensing technology; it is possible to
LR Linear regression capture and process large amounts of data. With an increasing
NN Neural networks shift to industry 4.0, the machine tools are now equipped with
CNN Convolutional neural networks several advanced on-machine sensors which provide data
MSE Mean Square Error related to the production line[1], processing status and
quality[2], and machine-tool parameters[3]. Cloud computing
enables proper storage of this data and makes it available as and
when needed, along with edge computing, which facilitates data

2212-8271 © 2022 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the ISEM XXI
10.1016/j.procir.2022.09.192

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Ming Wu et al. / Procedia CIRP 113 (2022) 410–416 411

location and processing closer to the machine tool. With an complex 2D machining trajectories through the synchronized
enormous amount of data, it becomes complicated to process activation of XY-stage motion (LUMS software) and
and analyze big data and deduce useful processing information. microsecond pulsed voltage source (LabVIEW® MMI on NI®
Proper usage of these data streams can recognize highly
complex and non-linear patterns in data of different types and
from different sources. And developing data-driven
approaches, in turn, facilitates complex and dynamic
manufacturing technologies to extract useful information from
big data.
Electrochemical machining (ECM) is a popular
unconventional manufacturing technology to fabricate surface
microstructures[4] such as micro-cavities[5], micro-
protrusions[6], micro-holes[7], and micro-channels[8] on
conductive and difficult-to-cut materials. It offers promising
machining performance, especially in terms of high surface
finish[9], no process-related tool wear, and the absence of
thermal load on the workpiece. Multiphysics models[10] are
widely used methods for profile prediction and hence support
tool and experiments design. However, the nonlinear
phenomena inherent to the ECM process such as electrolyte
flow, passivation[9], bubbles, etc., make it challenging to
reliably predict machined profiles. The multiphysical
model[11] with well-tuned parameters can fit the
experimental results beautifully for a certain processing
condition but could have radically different performances in
other conditions. Furthermore, it is difficult to simulate the
ECM process across a wide range of processing parameters in
the state-of-the-art commercial FE software because of
several issues related to time-stepping, convergence,
excessive mesh distortion during moving mesh, model Figure 1: Experimental setup
insensitivity to a small change in parameters, etc.
The aforementioned data-driven models[3], on the other cRIO 9045). A stainless-steel tubular electrode (1.2 mm outer
hand, present the capability of adapting to various processing diameter and 0.35mm internal diameter) was used for
parameters and associated nonlinear phenomena for a machining experiments. The electrolyte was supplied through
manufacturing process, thereby providing a more generalized a 3D printed electrolyte guide with internal flow channels. The
model for profile prediction. material is PLA (Polylactic acid) and the flow channel is
In this article, we present a framework of a data-driven embedded inside the guide to avoid electrolyte flow towards
model for the profile prediction of machined features with the the spindle and force it to flow towards the tool electrode.
ECM process. Different supervised machine learning models The electrolyte used was 20% by weight aqueous sodium
are used to investigate the relationship between the dimensions nitrate.
of cavities and the applied pulsed current, voltage, electrolyte
flow rate, and environmental temperature. The applied pulsed
current and temperature are monitored and recorded during the
machining process. These machine learning models are
evaluated by an out-of-domain evaluation dataset, which
applies processing parameters that are never been fed to the
data-driven models and also investigates the feasibility of using
data-driven models over a wide range of the parameter set. The
advantages of using a data-driven model with simple training
with processing parameters are investigated.

2. Materials and Methods

2.1. Experimental Setup and methodology

The experiments for collecting the data were conducted on

an in-house developed hybrid laser-electrochemical machining
(LECM)[12]. The setup, as shown in Figure 1, can execute Figure 2: 3D printed electrolyte guide

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 412 Ming Wu et al. / Procedia CIRP 113 (2022) 410–416

2.2. Data Collection

Cavities were machined in two concentric circles trajectory

as shown in Figure 1 with outer radius 62.5 µm (trajectory #1)
and inner radius 12.5 µm (trajectory #2) in two repetitions. The
two concentric circles trajectory were used to improve
electrolyte flow into the extremely small IEG of 80um. The
dimensions of the trajectory were such small that they were less
than 1% of the internal diameter (0.65 mm) of the tubular
electrode tool with a diameter of 1.2 mm. Thus electrochemical
reactions following these trajectories could represent point
removal but with dynamic flushing conditions.
Average peak machining current and temperature (4
temperature sensors were implemented: 1 in the electrolyte
tank, 2 distributed in the lab to check the room temperature, and
1 attached in the downstream side of the machining area) were Figure 3: Different shapes of cavities observed in the ECM process with
logged every second through the NI-9223 module. The a tubular tool. a) cavity with a central hill; b) cavity without a central hill.
machined samples were cleaned in an ultrasonic bath (DI water
of 8 MΩ.cm) for 20 mins before acquiring the cavity profiles
using the Sensofar® S Neox 3D Optical Profiler. 2.3. Machine learning models
Inconel 718 was used as a workpiece for machining the
cavities due to its wide range of applications in the aerospace For predicting the dimension of cavities, three machine
industry. All experiments were done at a fixed inter-electrode learning models are employed.
gap (IEG) of 80µm and machining time of 14s. 1. The linear regression model.
As shown in Table 1, the processing parameters, i.e., Linear regression is a simple machine learning model with
voltage, electrolyte fflow rate, pulse duration, and duty cycle, good interpretability. A linear regression model is established,
are varied. The applied pulse current and temperature are as shown in Figure 4, to predict the cavity features listed in
monitored and recorded during the machining process. The Table 2. The prediction outputs y to the inputs x can be given
training dataset is built for training the machine learning as:
algorithms to predict the machining profile. A second dataset, 𝑦𝑦(𝑥𝑥) = 𝑤𝑤0 𝑥𝑥0 + 𝑤𝑤1 𝑥𝑥1 + ⋯ + 𝑤𝑤𝑖𝑖 𝑥𝑥𝑖𝑖 + 𝑏𝑏 (1)
i.e., the validation dataset, is recorded for evaluating how well where, 𝑤𝑤𝑖𝑖 is the weight corresponding to the feature 𝑥𝑥𝑖𝑖 and b is
these models are truly generalizing to the prediction and not the bias.
simply overfitting the training dataset. Experiments are carried In the linear regression model, the processing parameters
out with processing parameter combinations of all the values in and the recorded signal (the mean value during the processing
Table 1. For each combination, 5 cavities are machined. time) can be feature-fused, together as the input data. The linear
Therefore, the training dataset should contain: 34 × 5 = 405 regression model can also use the process parameters or the
cavities. However, the cavities can hardly be machined recorded signal independently as the input data.
properly with the lowest and highest flow rate because of the
insufficient electric conductivity and electrolyte stability in our
case. A large portion of cavities machined with the flow rate of
0.14 or 0.95 ml/s was dropped from the legit dataset. Finally,
140 cavities were collected for the training dataset, and
22 × 5 = 20 cavities were collected for the validation dataset.

Table 1 Processing parameters for training and validation

Training Validation

Voltage (V) 15 20 25 22.5 30

Flow rate (ml/s) 0.14 0.47 0.95 0.47
Pulse duration (μs) 8 20 50 20 50 Figure 4: The linear regression model applied to the ECM process.

Duty cycle (%) 20 30 50 50

2. The neural networks model.
The neural networks model (NN) is a classic neuron-based
As shown in Figure 3, two types of cavities are produced machine learning model that consists of multiple hidden layers.
with different processing parameters. The profile features,
including the a) diameter D and b) depth of the cavity H, the c) This neuron is a computational unit with input x operated by
diameter H_D and d) height of the center hill H_H, are matrices of weights W and bias b, the output for each neuron
measured as outputs for machine learning models. The cross- ℎ𝑤𝑤,𝑏𝑏 is calculated by:
sectional profiles of cavities in the following sections are ℎ𝑊𝑊,𝑏𝑏 = 𝑓𝑓(𝑊𝑊 𝑇𝑇 𝑥𝑥) = 𝑓𝑓(∑𝑊𝑊𝑖𝑖 𝑥𝑥𝑖𝑖 + 𝑏𝑏) (2)
interpolated by these features as well.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Ming Wu et al. / Procedia CIRP 113 (2022) 410–416 413

For electrochemical processes, there are many nonlinear presents the log channels, 1 from the electric current and 4 from
equations. The hidden layer in the neural network model can the temperature, and “14” presents the time series size.
describe nonlinear phenomena. In this paper, as shown in The feature map F can be defined by the convolution kernel
Figure 5, a neural network with 3 hidden layers is set up. K and input matrix I as 𝐹𝐹 = 𝐼𝐼 ⊛ 𝐾𝐾, where ⊛ denotes the
convolutional operation, the feature map can be regarded as a
collection of features produced by the kernel K, which means
the convolution kernel K extract certain information from the
input matrix I.
Therefore, a convolutional model with a convolution kernel
size of 2 × 2 is developed so that its feature maps should cover
the information both horizontally, which means features
collection cross the record channel, and vertically, which
means feature collection cross the sampling time.
All the machine learning models are deployed with
Tensorflow backend and solved on a Nvidia® GPU RTX 2070.
With the presented data-driven approaches for profile
prediction in ECM, the computational time only costs less than
1s to make predictions for all the validation datasets.
The MSE is computed from the squared error between the
𝑖𝑖 and observed ( 𝑦𝑦 𝑖𝑖 ) measurements for each j-th
predicted (𝑦𝑦̂)
experimental run of the validation dataset, as shown in equation
(3) Here n is the total number of experimental runs in the
validation dataset.
𝑛𝑛−1
Figure 5: The neural network model.
1 2
MSE (𝑦𝑦, 𝑦𝑦̂) = ∑(𝑦𝑦𝑖𝑖 − 𝑦𝑦̂)
𝑖𝑖 (3)
𝑛𝑛
𝑖𝑖=0
Similar to the linear regression model, the neural network
model can use the process parameter or the recorded signal (the In this paper, the information is monitored and collected in
mean value during the processing time) separately, along with different ways. It is hard to decide initially whether this
fusing them, as input data. information is redundant or essential for profile prediction. The
following sections will analyze the predictive capabilities of
different machine learning models.

3. Results and discussion

3.1. Prediction accuracy of machine learning models

The predicted MSEs of different algorithms with different

inputs are shown in Figure 7 and Table 2, where P denotes the
processing parameters and circles in Figure 7 denote the
outliers; S means the recorded signal data during the process.
For LR and NN models, S means the time average value of the
recorded signal. For the CNN model, S means the raw recorded
data.

Table 2 Overall MSE of different machine learning methods in the

validation dataset
LR NN CNN
P S P+S P S P+S S P+S
D 1.20 0.30 0.71 0.72 0.92 0.92 0.74 0.42
Figure 6: The convolutional neural network model.
H 17.8 18.0 9.1 12.6 9.2 14.8 9.4 13.0
3. The convolutional neural networks model. H_D 0.48 0.26 2.40 0.37 0.43 0.54 0.12 0.09
Convolutional neural network (CNN) is one of the typical H_H 19.6 19.4 23.1 22.8 21.2 20.8 20.3 16.9
algorithms of deep learning, including convolution layers and
Total 9.76 9.51 8.82 9.12 7.93 9.28 7.64 7.60
pooling layers, as shown in Figure 6. Instead of using the mean
value of the recorded signal during the processing time, CNN
There are several nonlinear phenomena, including
can use all the data through the recording time as inputs. The
electrochemical reactions, in the ECM process. Because of the
size of the input signal hence will be 5 × 14 , where “5”
inability to learn and express the nonlinear relationships, it can
be seen that the linear regression (LR) model has higher

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 414 Ming Wu et al. / Procedia CIRP 113 (2022) 410–416

prediction errors and depicts poor prediction accuracy than the

CNN model. After the feature fusion is performed as inputs for
the linear regression model sometimes, the error is even higher.
This indicates that using recorded signals may even be held as
noise for linear regression.

Figure 8: Machined cross-sectional profile with: voltage 30V, flow rate

0.47ml/s, pulse duration 20μs, duty cycle 50%
Figure 7: Comparison of different algorithms with different inputs

For the neural network (NN) model with different inputs,

when using the recorded signal solely as the input data, the
average prediction MSE is 7.93; whereas, when using a
combination of processing parameters and the mean value of
recorded signals, the average prediction MSE is 9.28, the
prediction error increased. This presents the NN model’s limits
of absorbing different information.
Since the processing parameters cannot be individually
considered as inputs for the convolutional neural network
(CNN) model, the results of fused-feature input data sets are
presented.
It can be seen that, when using fused feature inputs, the CNN
model represents the lowest prediction MSE, i.e. the best
performance. This indicates that CNN has a stronger ability to
analyze different data and adapt to nonlinearity. While the LR
and NN models can only use recorded signal data by
calculating their mean value, the CNN model can work with
Figure 9: Machined cross-sectional profile with: voltage 25, flow rate
record data with no subtraction. The more information fed to
0.47ml/s, pulse duration 8μs, duty cycle 50%
the machine learning model, the better prediction accuracy it
can present.
As shown in Figures 8, 9, the cavities in each figure were
machined in a certain processing condition but presented
3.2. Data-driven approaches in the unsatisfactory processing
profoundly different profiles. In the meantime, a significant
conditions
difference can be clearly noticed from the electric signal record.
This suggested some missing information that is essential to the
In the classical ECM model, all the analysis is based on ideal
ECM processing could be found in the recorded data.
and fixed machining conditions during the processing time,
which is heavy to be achieved and maintained in practice.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Ming Wu et al. / Procedia CIRP 113 (2022) 410–416 415

Figure 11: Predicted profile of different algorithms with different inputs

It can be seen that the predictions given by the CNN model

with P+S inputs fit the measurements most. In addition, this
Figure 10: Predicted profile of different algorithms with different inputs confirms the results of Section 3.1 that the data-driven
approaches with more complex architectures perform better;
As shown in Figures 10, 11, the identical processing the inputs with more information fed to the data-driven model
conditions can cause different profiles, and different algorithms perform better.
with different inputs showed a different predicted accuracy.
When only using the processing parameters as inputs, identical 3.3. Limitations of machine learning models
predictions will be presented, where is far from the ground
truth. When using the recorded signal data as inputs or one part Figure 12 demonstrated the predicted MSE for different
of inputs, the predictions become different, as the signal data is profile features. It can be seen that the diameter of the machined
different. cavity and the center hill have lower predicted errors, whereas
the prediction error of the cavity’s depth and the center hill’s
height is unexpecting high.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 416 Ming Wu et al. / Procedia CIRP 113 (2022) 410–416

References

[1] A. Perrot, R. Bourqui, N. Hanusse, F. Lalanne, and D. Auber, “Large

interactive visualization of density functions on big data infrastructure,”
2015, doi: 10.1109/LDAV.2015.7348077.
[2] R. Zhang et al., “Analysing the degree of sensitisation in 5xxx series
aluminium alloys using artificial neural networks: A tool for alloy
design,” Corros. Sci., vol. 150, pp. 268–278, Apr. 2019, doi:
10.1016/J.CORSCI.2019.02.003.
[3] L. Wen, X. Li, L. Gao, and Y. Zhang, “A New Convolutional Neural
Network-Based Data-Driven Fault Diagnosis Method,” IEEE Trans. Ind.
Electron., vol. 65, no. 7, pp. 5990–5998, Jul. 2018, doi:
Figure 12: Predicted MSE for different profile features 10.1109/TIE.2017.2774777.
[4] M. Wu, Z. Guo, J. Qian, and D. Reynaerts, “Fabrication of Surface Micro
One plausible reason is that the information collected now Letters by Electrolyte Jet Mask Machining,” Procedia CIRP, vol. 95, pp.
is not enough for data-driven approaches to faithfully predict 827–832, Jan. 2020, doi: 10.1016/J.PROCIR.2020.02.261.
the cavity’s depth or the center hill’s height. For instance, as [5] K. K. Saxena, J. Qian, and D. Reynaerts, “A review on process
the in-processing electrode flow condition in the IEG is capabilities of electrochemical micromachining and its hybrid variants,”
currently beyond reach, whereas there is a consensus that the Int. J. Mach. Tools Manuf., vol. 127, pp. 28–56, 2018, doi:
flow field in the inter-electrode gap (IEG) is crucial to the 10.1016/j.ijmachtools.2018.01.004.
machining process[11] and is sensitive to the bubbles [6] M. Wu, J. Liu, J. He, X. Chen, and Z. Guo, “Fabrication of surface
generation and machined workpiece surface profile, the effects microstructures by mask electrolyte jet machining,” Int. J. Mach. Tools
of electrochemical reactions principally influenced by the Manuf., vol. 148, p. 103471, Jan. 2020, doi:
electrolyte are not perceived by the machine learning models, 10.1016/j.ijmachtools.2019.103471.
hence the faithful predictions which cover all the processing [7] N. S. Qu, X. F. Zhang, X. L. Chen, H. S. Li, and D. Zhu, “Modified
scenarios in this investigation are practically challenging to be microscale pattern transfer without photolithography of substrates,” J.
presented. This is the inherent imperfection of data-driven Mater. Process. Technol., 2015, doi: 10.1016/j.jmatprotec.2014.11.040.
intelligence — no data, no intelligence. [8] M. Wu, K. K. Saxena, Z. Guo, J. Qian, and D. Reynaerts, “Fast
Fabrication of Complex Surficial Micro-Features Using Sequential
4. Conclusion Lithography and Jet Electrochemical Machining,” Micromachines 2020,
Vol. 11, Page 948, vol. 11, no. 10, p. 948, Oct. 2020, doi:
1. The machine learning model shows a good generalizing 10.3390/MI11100948.
ability. When using the out-of-domain datasets, the CNN [9] L. Meng, Y. Zeng, and D. Zhu, “Investigation on Wire Electrochemical
model presents a prediction MSE of 7.60. Micro Machining of Ni-based Metallic Glass,” Electrochim. Acta, 2017,
2. With data recorded during the machining process, more doi: 10.1016/j.electacta.2017.03.045.
accurate results will be achieved, however not all the [10]M. Wu, Zhongning Guo, J. He, and X. Chen, “Modeling and Simulation
algorithms can get the same advantages when adopting more of the Material Removal Process in Electrolyte Jet Machining of Mass
information. The extra information could be noise or irrelevant Transfer in Convection and Electric Migration,” in Procedia CIRP, 2018,
inputs for the LR and NN model. vol. 68, pp. 488–492, doi: 10.1016/j.procir.2017.12.079.
3. Despite the unsatisfactory processing conditions, good [11]D. Deconinck, S. Van Damme, and J. Deconinck, “A temperature
predictions can still be presented because of the employment of dependent multi-ion model for time accurate numerical simulation of the
recorded in-processing data. electrochemical machining process. Part II: Numerical simulation,”
4 The predicted accuracy is relatively low for the cavity’s Electrochim. Acta, 2012, doi: 10.1016/j.electacta.2012.02.079.
depth or the center hill’s height, which may be caused by the [12]K. K. Saxena, J. Qian, and D. Reynaerts, “A tool-based hybrid laser-
IEG electrolyte flow condition and requires further electrochemical micromachining process: Experimental investigations
investigation on the in-processing monitoring system. and synergistic effects,” Int. J. Mach. Tools Manuf., 2020, doi:
10.1016/j.ijmachtools.2020.103569.
Acknowledgments

This research was supported and partially financed by

Flanders Make.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.