The International Journal of Advanced Manufacturing Technology (2025) 136:775–787

[Link]

ORIGINAL ARTICLE

Optimization of single‑point diamond turning processes for single

crystal calcium fluoride: A surrogate model for surface roughness
prediction
Ke‑Er Tang1 · Yin‑Chung Huang1 · Wei‑Ting Lin1 · Yuan‑Chieh Cheng2 · Chun‑Wei Liu1

Received: 15 August 2024 / Accepted: 29 November 2024 / Published online: 9 December 2024
© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024

Abstract
Single-crystal calcium fluoride (­ CaF2) is essential in high-level optical systems due to its excellent optical properties, includ-
ing high penetration, laser damage resistance, and low dispersion across ultraviolet and infrared wavelengths. However, ­CaF2
is soft, brittle, and its anisotropic nature makes predicting surface roughness (Ra) during lens manufacturing challenging.
Traditionally, operators rely on manual adjustments and measurements to achieve desired surface finishes, a process that
is both time-consuming and costly. This study addresses these challenges by developing a model to predict surface rough-
ness in ultra-precision machining of ­CaF2. The machining depth was set to 200 nm to maintain ductile cutting conditions.
A central composite design (CCD) experiment gathered data on three critical machining parameters: spindle speed, feed
rate, and tool rake angle. The surface roughness prediction models were then constructed using a synthetic artificial neural
network (S-ANN), response surface methodology (RSM), and kriging. Evaluation results showed that the S-ANN model
achieved the highest accuracy (97.4% accuracy, 0.907 nm error), outperforming the RSM (90.4% accuracy, 1.343 nm error)
and kriging (87.1% accuracy, 1.931 nm error) models. The S-ANN model’s performance is suitable for practical application,
offering an effective method to predict ­CaF2 Ra, significantly reducing costs associated with single-point diamond turning.

Keywords Single-crystal calcium fluoride · Single-point diamond turning · Surrogate model · Synthetic artificial neural
network · Surface roughness

1 Introduction conductivity, and relatively poor physical properties (Table 1

­ aF2 leads to it having differ-
[3]). The anisotropic nature of C
Single-crystal calcium fluoride ­(CaF2) materials are widely ent ionic bond densities in various lattice orientations and
used in high-level optical lens systems for semiconductors varying macromechanical properties in different crystal
and optical inspection because of their excellent optical planes; thus, the high-precision faceted machining of C ­ aF2
properties in the ultraviolet and infrared wavelength bands, lenses is extremely challenging. With the growing require-
high penetration, high resistance to laser damage, and low ment of lenses with increasing precision and complexity in
dispersion [1]. The optical properties of ­CaF2 lenses depend the manufacturing industry, traditional optical processing
on their surface roughness (Ra) [2]. However, ­CaF2 is a soft methods such as grinding and polishing are increasingly
and brittle material with low hardness, low fracture tough- inadequate for achieving the high-precision processing
ness, a high coefficient of thermal expansion, low thermal required for complex surfaces.
Single-point diamond turning (SPDT) is used for ultra-
precision machining. It involves the use of a hydrostatic
* Chun‑Wei Liu guideway and an air spindle to isolate vibration during
weilu@[Link] machining, with the extremely high wear resistance and
1 thermal conductivity of the diamond tool facilitating the
National Tsing Hua University, Power Mechanical
Engineering, Section 2, Guangfu Road, Hsinchu 300044, achievement of high-speed ultraprecision turning. SPDT
Taiwan technology enables the achievement of nanoscale surface
2
AMETEK Precitech Inc., 44 Black Brook Road, Keene, roughness and the precision machining of complex surfaces
NH 03431, USA [4]. Therefore, some researchers have proposed using SPDT

Vol.:(0123456789)
776 The International Journal of Advanced Manufacturing Technology (2025) 136:775–787

Table 1  Optical, thermal, and mechanical properties of ­CaF2 and results obtained through testing and machining in the
Properties present study indicate that the (111) crystal plane of ­CaF2
is prone to exhibiting a unique surface pattern with a trefoil
Transmission range (μm) 0.12 ~ 12
shape during SPDT (Fig. 1(a)). This phenomenon is caused
Index of refraction (nd) 1.434 by changes in the cutting direction of the tool and the orien-
Abbe number (vd) 94.99 tation of the crystal during the face turning process, result-
Thermal expansion coefficient ­(K−1) 2 × ­10−5 ing in the crystal surface exhibiting different relative slip
Thermal conductivity (W/(m. K)) 9.71 conditions. Chen et al. [12] conducted experimental analy-
Knoop hardness (kg/mm2) 158.3 ses of different slip conditions by scribing the surface of a
Young’s modules (GPa) 75.8 workpiece at 30° intervals in the (111) crystal direction by
using a diamond scribe. The depths of scratches were meas-
ured to determine the deformation behavior of the work-
to fabricate high-precision complex C ­ aF2 lenses [5–8]. ­CaF2 piece surface. The results indicated that in the scratch angle
belongs to the isometric crystal system, and its laser damage from 0° to 120°, the scratch depth decreased from a high
threshold is higher along the (111) orientation than along the value (180 nm) to a low value (150 nm) and then increased
(110) and (001) orientations. Therefore, lenses fabricated again to a high value (180 nm). Thus, under the same load,
from the (111) plane of C­ aF2 have a longer lifespan than do a larger deformation occurred for these angles than for other
those fabricated from the (110) and (001) planes [9]. Con- angles, which made it easier to enter a brittle cutting condi-
sequently, the (111) plane of C
­ aF2 is often used as an optical tion at 0° to 120°. This phenomenon resulted in angularly
surface. However, the (111) surface is more anisotropic, and differentiated texture zones on the ­CaF2 surface; these zones
the conversion of the mechanical properties of different crys- comprised brittle and ductile cutting zones. The Ra values
tallographic orientations during the cutting process leads to of the brittle and ductile zones were approximately 200 and
machining difficulties, which results in high requirements 8 nm, respectively.
regarding surface finish quality in SPDT. Many preliminary studies have provided a general direc-
In SPDT, the ideal Ra value of a machined surface is influ- tion of development for the SPDT of ­CaF2 on the produc-
enced by the spindle speed, feed rate, and nose radius. The tion line. Selecting an appropriate critical depth of cut is
average Ra value is calculated using the following formula: one solution for reducing stress damage cracks in a material
subjected to SPDT. In cutting tests, especially for complex
Ra = 0.0321 × f 2 ∕R (1) surfaces, the critical depth of cut is not controllable, and the
where f is the feed per revolution, and R is the nose radius. depth of cut does not theoretically affect the ideal surface
However, the Ra value is often affected by various machining roughness. Therefore, to ensure that a cutting test is safe, the
factors, leading to inaccurate Ra predictions, especially in operator sets the depth of cut at the smallest jog distance of
the ultraprecision machining of C
­ aF2 through SPDT. Studies the machine. However, because the machine and workpiece
on the SPDT machining of ­CaF2 have indicated that changes have different shapes, the operator must subsequently deter-
in machining parameters such as spindle speed, feed rate, mine other suitable machining parameters through machin-
depth of cut, and tool rake angle can lead to brittle fracture ing experience and repeated testing. This process is tedious,
along mechanically fragile crystal orientations and consider- time-consuming, and costly. Therefore, a surface roughness
ably increase the Ra value [5–11]. Results from the literature prediction model should be developed for the SPDT of ­CaF2
to assist operators in selecting machining parameters rapidly.

Fig. 1  a Machining patterns of

the (111) surface of ­CaF2 and b
crystal orientations of ­CaF2
The International Journal of Advanced Manufacturing Technology (2025) 136:775–787 777

Improving process quality and reducing production costs used to determine relationships between different machining
through predictive modeling has become a key trend in the parameters. ANNs, which learn relationships between dif-
manufacturing industry. The SPDT of ­CaF2 is a costly and ferent parameters by mimicking the operation of biological
time-consuming method; therefore, an appropriate model neural networks, can capture complex details and patterns
should be developed for accurately predicting the Ra val- in a manufacturing process because of their excellent non-
ues of ­CaF2 to enable reductions in the cost and duration of linear learning capability. Natarajan et al. [20] used an ANN
the aforementioned process. Compared with full-factorial model to predict and optimize the Ra value on the basis of
experimental design, the central composite design (CCD) the spindle speed, feed rate, and depth of cut in the turning
of experiments enables all experimental combinations to of C26000 brass. Khorasani et al. [21] developed a dynamic
be tested through fewer experiments. In CCD, the center- ANN-based Ra monitoring system for the milling of Al 7075
point and axial-point approaches are used to determine the and St 52 workpieces. Moreover, Mia et al. [22] used an
variations in experimental parameters. Response surface ANN model with RSM to predict Ra, cutting force, and feed
methodology (RSM) and kriging methods are favored in rate in the turning of Ti-6Al-4 V alloy.
the modeling process of CCD because they have relatively This paper is structured as follows. Section 1 provides the
low data requirements. RSM and kriging models are often background of this study, detailing the challenges involved
used to approximate more complex physical models. RSM in the SPDT of C ­ aF2 and solutions for overcoming these
is a statistical method that is applied to experimental data to challenges. Section 2 describes the methodology and mod-
determine how various factors are associated with a response eling theories used in this study. Section 3 details the experi-
variable. Alauddin et al. [13] used RSM to develop a sur- mental equipment, design, and results. Section 4 presents
face roughness prediction model for steel milling, and Sahin a discussion on the results. Finally, Sect. 5 elucidates the
et al. [14] applied RSM to predict the Ra value of low-carbon conclusion and contributions of this study.
steels turned using coated carbide tools. Yang et al. [15] used
RSM to optimize the Ra value of titanium alloy subjected to 1.1 2. Research methods
turning. Amandeep et al. [16] combined RSM with CCD to
optimize the Ra value of a planar quartz substrate in a polish- Figure 2 displays a flowchart of the methodology used in
ing process. Kriging methods have also been widely applied this study. In the SPDT of ­CaF2, the main factor affecting
in the field of machining. Aramesh et al. [17] used a kriging the processing results is the crystal anisotropy, which results
method to predict Ra values and tool wear in the turning of in different crystal faces having different macromechanical
titanium metal matrix composites. Dang et al. [18] used a properties. This anisotropy often leads to transitions between
kriging model to study the nonlinear relationships between brittle and ductile cutting states during the cutting of C
­ aF2.
machining parameters and machining response characteris- An increase in the depth of cut can easily result in the brit-
tics (namely, material removal rate, Ra, and electrode wear) tle rupture of ­CaF2. To ensure that the ductile cutting state
in EDM (electric discharge machining). Nguyen et al. [19] is maintained, this study set the depth of cut to match the
used a kriging method to predict the Ra values of SKD61 minimum jog distance of the machine, namely 200 nm. This
and predict the machining energy, Ra, and productivity in study considered the effects of spindle speed, feed rate, and
SKD61 milling operations. tool rake angle on surface roughness in a CCD analysis.
In addition to RSM and kriging methods, which are sta- Measurement data on these factors were collected to con-
tistical methods, artificial neural networks (ANNs) can be struct RSM, kriging, and synthetic ANN (S-ANN) models

Fig. 2  Flowchart of the methodology used in this study

778 The International Journal of Advanced Manufacturing Technology (2025) 136:775–787

for predicting surface roughness, and the prediction perfor- number of influencing factors, and c is the number of experi-
mance of these models was compared. ments at the center point.

1.2 Central composite design N = 2k + 2 × k + c (2)

Three influencing factors were considered in this study:
CaF2 processing is costly; therefore, a method that enables spindle speed, feed rate, and tool rake angle. Experiments
the optimization of experimental parameters with a mini- at the center point were conducted thrice for the number
mum number of experiments is crucial for this process. of experiments to be reduced from 27 to 17. Moreover,
CCD can reduce the number of experiments while still pro- α is usually 1 when rotation of the shaft is not taken into
viding sufficient information for model fitting. In contrast account; thus, the experimental values were classified into
to full-factorial experimental design, CCD involves test- three levels: low, medium, and high. Finally, uncontrollable
ing all experimental combinations through a center-point factors such as tool wear, ambient temperature, and vibration
approach and axis-point approach [23]. A comparison of affected the output response; therefore, the experiments of
these approaches is illustrated in Fig. 3. When a second- this study were conducted in a random order to prevent error
order response model must be constructed in a three-factor accumulation from affecting specific parameter combina-
experiment, the full-factorial experimental design method tions, which would distort the experimental results.
requires the consideration of at least three factor levels, as
shown in the red circle, whereas CCD only requires the con-
sideration of two factor levels as well as the center points 1.3 Modeling methods
and axial points, as shown in the blue dot. In this schematic,
CCD experiments can be reduced by 10 sets of experiments 1.3.1 Response surface methodology
during the modeling process.
Thus, CCD was used in this study to construct a second- In RSM, statistical experimental design techniques and least-
order response surface model. In CCD, after the high and squares fitting methods are used to construct models. This
low levels of each control factor are determined, a center approach was originally proposed by Box and Wilson for
point and an axial point are added for replacing some of the model fitting in physical experiments and has been gradu-
experiments in the full factorial model. The center point ally applied in many other fields [24]. In RSM, experimental
enables the exploration of whether nonlinear error exists design and regression analysis are combined by systemati-
between the high and low levels and provides an independ- cally planning input factors, collecting corresponding output
ent error estimate. The experiment is usually repeated 3–5 responses, and then using regression analysis to fit the rela-
times at the center point to evaluate the experimental error. tionships between the input factors and the output responses.
The axial point, which is located on the axis at a distance of After crucial influencing factors are identified, an RSM
α from the center point, enables the examination of whether model can be used to predict output values corresponding to
a quadratic term exists in the model. The number of experi- various processing parameters or to obtain optimal response
ments (N) is determined using Eq. (2), in which k is the values for independent factors. The aforementioned step

Fig. 3  Comparison of full-

factorial experimental design
and CCD
The International Journal of Advanced Manufacturing Technology (2025) 136:775–787 779

is usually performed using a first- or second-order model where R is the matrix of correlation functions between
Eq. (3). known points Eq. (7); r(x) is the matrix of correlation func-
tions between unknown and known points Eq. (8); Y is the
∑
k
∑
k
∑
k
∑
k
value of each known point Eq. (9); Fs is the value of the
y = a0 + ai xi + aij xi xj + aii xi 2 + ε (3)
i=1 i=1 j>i i=1 function of each known point Eq. (10), and is the Kriging
algorithm Eq. (11).
where y represents the surface roughness; a is a coefficient
to be determined by the model; x is the input influence fac- ⎡ R(x1 , x2 ) ⋯ R(x1 , xp ) ⎤
tor; k is the number of influence factors; and ε is the model R=⎢ ⋮ ⋯ ⋮ ⎥ (7)
⎢ ⎥
residual. This study conducted RSM-based modeling using ⎣ R(xp , x1 ) ⋯ R(xp , xp−1 ) ⎦
the Minitab (State College, Pennsylvania, USA) statistical
software program. [ ( ) ( ) ( )]
r(x) = R x, x1 , R x, x2 , ⋯ , R x, xp (8)

[ ( ) ( ) ( )]
1.3.2 Kriging model Y = y x1 , y x2 , ⋯ , y xp (9)

The kriging model is a regression algorithm that uses known [( ) ( ) ( )]T

Fs = f x1 , f x2 , ⋯ , f xp (10)
information to model and predict the effects of unknown fac-
tors. This model was first proposed by Georges Matheron for
use in geostatistics to estimate the probability of a mineral −1
𝛽 ∗ = (Fs T R−1 Fs ) Fs T R−1 Y (11)
source in an area by collecting a small number of samples
[25]. The kriging model estimates the covariance of the sto-
chastic field by using the variance functions between known
points and between known and unknown points. The vari- 1.3.3 Synthetic ANN
ance functions are continuous functions all points except for
the origin; therefore, the stochastic field predicted by the ANN algorithms are conceptually based on how the human
kriging model is smooth, and this model’s prediction results brain operates. These algorithms have a multilayered net-
are continuously correlated with the known data. A typical work architecture with neurons as the basic unit. An ANN
kriging prediction model is expressed in Eq. (4). comprises an input layer, one or more hidden layers, and
an output layer [27]. The hidden layer processes the data
∑
p
( ) received from the input layer, which are transformed nonlin-
̂
y(x) = 𝛽i f xi + z(x) (4) early through thresholding by using a transformation func-
i=1
( ) tion. Finally, the output layer generates predicted values.
where ̂ y(x) is the predicted value; f xi is the function value Each ANN used in this study contains three hidden layers
of the known ( sample
) points; 𝛽i is the coefficient of determi- with eight neurons each, and it uses the rectified linear unit
nation of f xi ; and z(x) is a stochastic variable. The sto- activation function Eq. (12).
chastic variable has a normal distribution with a mean of 0 {
and a variance of 𝜎 2 , and this variables’ covariance matrix x, x ≥ 0
f (x) =
0, x < 0 (12)
is expressed as follows:
( ( ) ( )) ( )
cov z xi , z xj = 𝜎 2 R xi , xj (5) The performance of an ANN model can be effectively
improved by training multiple neural networks and combin-
( )
where R xi , xj is the correlation function between any two ing their predictions through averaging. The ANN model
sample points. This function determines the type of model produced after performing the aforementioned step is called
fit. Equation (2.3) can be rewritten as follows: an S-ANN (synthetic ANN) model [28]. In this study, five
neural networks were trained, and their predictions were
y(x) = f T (x)𝛽 ∗ + rT (x)R−1 (Y T − Fs 𝛽 ∗ )
̂ (6) averaged to reduce the risk of overfitting and improve model
performance. The overall architecture of the S-ANN model
where R is the matrix of correlation functions between
used in this study is displayed in Fig. 4. A single model
known points as in Eq. (7); is the matrix of correlation func-
might lead to overfitting because of the presence of noise or
tions between unknown and known points as in Eq. (8); Y is
random patterns in the training data. This problem can be
the value of each known point as in Eq. (9); r(x) is the value
mitigated by combining the predictions of several models.
of the function of each known point as in Eq. (10); and the
This approach weakens the effect of random patterns and
algorithm is as in Eq. (11) [26].
allows for more data features and patterns to be captured,
780 The International Journal of Advanced Manufacturing Technology (2025) 136:775–787

Fig. 4  S-ANN model used in

this study

Fig. 5  Experimental setup for

ultra-precision machining

thereby improving the accuracy and generalization ability Table 2  Experimental factors and measurement indices
of the overall model. Parameters Low level Center point High level

S (rpm) 500 1000 1500

2 . Experiments F (mm/min) 1 4 7
A (°) − 30 − 15 0
2.1 Experimental settings Measurement indices Ra (nm)

A five-axis ultraprecision machine (Freeform 705XG,

AMETEK Precitech, Keene, New Hampshire, USA) with
X, Y, Z, B, and C axes was used in this study. The work- setup is displayed in Fig. 5. Surface roughness (Ra) was
piece adopted in this study was a single crystal of C ­ aF2 measured using a surface profilometer (RTH Form Taly-
with the (111) crystallographic orientation; the crystal had surf, AMETEK Taylor Hobson, Leicester, UK).
a diameter and thickness of 26 and 8 mm, respectively. Hot
softening glue was used to glue the workpiece to an alu-
minum plate, and the workpiece was then attached to the 2.2 Design of experiment and analysis of variance
spindle of the ultraprecision machine by using a vacuum
system for face turning. During the machining process, The CCD method was used in this study to design the
paraffin mist was used as the cutting fluid; the nose radius machining parameters. In this method, spindle speed (S),
was fixed as 1 mm; the depth of cut was fixed as 0.02 μm; feed rate (F), and tool rake angle (A) were considered as
and the tool rake angle was adjusted using a self-designed influencing factors, and surface roughness (Ra) was con-
mechanism. The aforementioned depth of cut prevented sidered as the response variable. Experiments at the center
brittle cutting conditions from causing the rapid enhance- point was conducted thrice, for a total of 17 experiments
ment of product surface roughness. The experimental (Table 2). To validate model performance, an additional
eight experiments were conducted.
The International Journal of Advanced Manufacturing Technology (2025) 136:775–787 781

3 Results and discussion Figures 6 and 7 display the main effects and interaction
effects acting on Ra, respectively. Because a lower Ra value
3.1 Experimental results indicates better machining quality, a “smaller-the-better”
quality characteristic was selected. The plots in the aforemen-
Table 3 presents the results of the 25 experiments con- tioned figures indicate that Ra decreased and increased with
ducted in this study. an increase in the spindle speed and feed rate, respectively.
This study used the Minitab statistical software pro- These two trends are consistent with Eq. (1). The relationship
gram for experimental design and statistical analysis. between tool rake angle and Ra was nonlinear, with Ra being
Analysis of variance (ANOVA) was conducted in this lowest at a tool rake angle of approximately − 15°. The interac-
study to determine the effect of each experimental factor tion effects on Ra were not significant at the 0.05 level.
on the response variable. The results of ANOVA for Ra are
presented in Table 4, in which MS, SS, and DoF refer to 3.2 Prediction of surface roughness
the mean square, sum of squares, and degrees of freedom,
respectively. The F and P values obtained in the ANOVA 3.2.1 Prediction of surface roughness by using an RSM
suggested that spindle speed, feed rate, and square of the model
tool rake angle had significant effects at the 0.05 level on
the Ra value of ­CaF2 in SPDT. An RSM model for Ra prediction was constructed using
measurement data obtained in CCD. The constructed RSM
model is expressed in Eq. (13). The contour plots obtained
with the RSM model are displayed in Fig. 8(a), (b), and (c).
Table 3  Experimental results as per DOE In this figure, bluer regions indicate lower Ra. As displayed
in Fig. 8(a), at a fixed tool rake angle of − 15°, the surface
No Machining parameters Experi-
mental
roughness decreased with an increase in spindle speed and
results a decrease in feed rate. Moreover, the lowest surface rough-
S (rpm) F (mm/min) A (°) Ra (nm) ness was observed at a tool rake angle of approximately − 15°,
which is consistent with the main effect plot.
Training data
However, on the color bar of the RSM model, it can be
1 500 1 0 10.237
observed that the dark blue color is the surface roughness close
2 1000 4 0 15.125
to 0 nm, which is below the range of machining quality accord-
3 1500 1 0 4.762
ing to the experimental machining results and the machining
4 500 7 0 20.001
experience on the machine. In Fig. 8(a) and (c), there are dark
5 1500 7 0 12.249
blue areas, which represent the poor prediction of the RSM
6 1500 4 − 15 6.025
model in the edge range.
7 1000 4 − 15 6.553
8 1000 1 − 15 3.558 Ra =14.71 − 0.01406S + 3.32F + 0.925A + 0.000004S2
9 1000 7 − 15 7.704
− 0.247F 2 + 0.02584A2 − 0.000082SF
10 1000 4 − 15 5.674
11 1000 4 − 15 8.004
− 0.000107SA − 0.0056FA
(13)
12 500 4 − 15 11.591
13 1500 1 − 30 4.962
14 500 1 − 30 9.009
3.2.2 Prediction of surface roughness by using a kriging
15 1000 4 − 30 12.204
model
16 500 7 − 30 18.001
17 1500 7 − 30 15.207
A kriging model was also developed on the basis of the CCD
Testing data
data. The contour plots obtained with this model are shown in
18 700 2 0 11.927 Fig. 9. In the kriging prediction map, it can be observed that
19 1400 2 0 9.631 there is no dark blue Ra of 0 nm in the overall coarse prediction
20 700 6 0 18.520 color range. This means that the kriging model will give more
21 1400 6 0 12.865 reasonable results in terms of edge position and prediction.
22 700 2 − 20 5.178
23 1400 2 − 20 4.469
24 700 6 − 20 11.584
25 1400 6 − 20 9.810
782 The International Journal of Advanced Manufacturing Technology (2025) 136:775–787

Table 4  ANOVA for Ra Source DoF SS Adjusted MS F-value P-value

Model 9 360.508 40.056 13.82 0.001

Linear 3 231.604 77.201 26.63 0.000
Spindle speed 1 65.885 65.885 22.72 0.002
Feed rate 1 164.804 164.804 56.84 0.000
Rake angle 1 0.916 0.916 0.32 0.592
Square 3 123.116 41.039 14.15 0.002
Spindle speed × spindle speed 1 2.453 2.453 0.85 0.388
Feed rate × feed rate 1 13.207 13.207 4.56 0.07
Rake angle × rake angle 1 90.543 90.543 31.23 0.001
2-Way interaction 3 5.788 1.929 0.67 0.599
Spindle speed × feed rate 1 0.122 0.122 0.04 0.844
Spindle speed × rake angle 1 5.155 5.155 1.78 0.224
Feed rate × rake angle 1 0.511 0.511 0.18 0.687
Error 7 20.296 2.899
Lack-of-fit 5 17.527 3.505 2.53 0.307
Pure error 2 2.769 1.384
Total 16 380.804

3.2.3 Prediction of surface roughness by using an S‑ANN

model

The CCD data were employed to train and optimize an

S-ANN model for Ra prediction. The adopted S-ANN model
comprised five ANN models with an accuracy higher than
95%. Each of these models contained three hidden layers
with eight neurons each. The results of the five models are
averaged to obtain the final results, which are depicted in
Fig. 10.
The S-ANN model gives similar prediction trends as
the kriging model, which means that its prediction results
are more reasonable. However, compared with the kriging
model, the S-ANN model has more nonlinear characteristics
Fig. 6  Main effects plot for Ra in the Ra variation of the predicted trend. It is necessary to
test the accuracy of the two models through practical tests.

3.3 Model comparison

This study used the results obtained in eight additional

experiments to test the predictive ability of the adopted mod-
els on unknown data. Model performance was evaluated in
terms of mean absolute percentage error (MAPE) and root
mean square error (RMSE), which are calculated using Eqs.
(14) and (15), respectively. Smaller values of these parame-
ters indicate higher model accuracy. The performance results
of an ANN model (single ANN) and the S-ANN model are
presented in Table 5. Compared with the ANN model, the
S-ANN model had a higher coefficient of determination (R2)
and accuracy as well as lower RMSE and MAPE. Thus, the

Fig. 7  Interaction effects plot for Ra

The International Journal of Advanced Manufacturing Technology (2025) 136:775–787 783

Fig. 8  Results obtained with the RSM model: a variations in feed rate with spindle speed at tool rake angle of − 15°, b variations in tool rake
angle with spindle speed at a feed rate of 4 mm/min, and c variations in tool rake angle with feed rate at a spindle speed of 1000 rpm

S-ANN model had higher accuracy, lower error, and better captured the relationships between surface roughness and
generalization ability than did the ANN model. the three considered machining parameters. Nevertheless,
the S-ANN model outperformed the RSM model in terms
1 ∑ |experimental valuei − predicted valuei |
n
MAPE = of RMSE and MAPE (Table 6). The prediction error of the
n i=1 experimental valuei RSM model might have been caused by its limited ability to
(14) predict high-dimensional nonlinear relationships.
√ Although the kriging model and S-ANN model had
√ n
√1 ∑
RMSE = √ (experimental valuei − predicted valuei )2 similar R2 values, the kriging model’s RMSE, MAPE, and
n i=1 maximum error were higher than those of the RSM and
(15) S-ANN models. Thus, the kriging model exhibited over-
2 fitting, which resulted in deviations between its predic-
In the testing stage, the RSM model had a lower R value
tions and the actual processing results in certain parameter
(94.67%) than did the kriging and S-ANN models (99.27%
ranges.
and 99.64%, respectively). However, the MAPE and RMSE
The S-ANN model performed the best, with its R2,
of the RSM model were 9.6% and 0.878 nm, respectively,
MAPE, RMSE, and maximum error being 99.64%, 2.6%,
which were lower than those of the kriging model (12.9%
0.383 nm, and 0.907 nm, respectively. These results dem-
and 1.18 nm, respectively); thus, the RSM model accurately
onstrate that the S-ANN model has excellent model fitting
784 The International Journal of Advanced Manufacturing Technology (2025) 136:775–787

Fig. 9  Results obtained with the kriging model: a variations in feed rate with spindle speed at tool rake angle of − 15°, b variations in tool rake
angle with spindle speed at a feed rate of 4 mm/min, and c variations in tool rake angle with feed rate at a spindle speed of 1000 rpm

and prediction capabilities. The S-ANN model can handle and Ra in the aforementioned process were investigated by
complex nonlinear relationships by averaging the results of conducting ANOVA on the collected data. The results of
multiple ANN models, thus outperforming the RSM and ANOVA indicated that the spindle speed, feed rate, and tool
kriging models (Table 6). rake angle significantly influenced Ra value.
The R2 values of the S-ANN, RSM, and kriging models
were 99.64%, 94.67%, and 99.27%, respectively. Moreover,
4 Conclusion the accuracies of these models on unknown data were 97.4%,
90.4%, and 87.1%, respectively, with their maximum errors
This study developed an RSM model, a kriging model, and being 0.907, 1.343, and 1.931 nm, respectively. These errors
an S-ANN model for predicting surface roughness in the are acceptable in practical SPDT.
SPDT of ­CaF2 lenses. First, the depth of cut was defined Although the kriging model had an extremely high R2
as being equivalent to the minimum jog distance of the value, it had the lowest accuracy and largest errors among
machine to ensure ductile cutting. Subsequently, the CCD the three adopted models; thus, the kriging model exhibited
method was used to design experimental machining param- overfitting. Among the three models, the S-ANN model had
eters for the collection of experimental data in the SPDT of the best results in terms of R2, RMSE, MAPE, and maximum
­CaF2. The relationships of different machining parameters
The International Journal of Advanced Manufacturing Technology (2025) 136:775–787 785

Fig. 10  Results obtained with the S-ANN model: a variations in feed rate with spindle speed at tool rake angle of − 15°, b variations in tool rake
angle with spindle speed at a feed rate of 4 mm/min, and c variations in tool rake angle with feed rate at a spindle speed of 1000 rpm

Table 5  Comparison of single Model ANN_1 ANN_2 ANN_3 ANN_4 ANN_5 S-ANN
ANN models and ensembled
ANN model R2 99.21% 99.56% 99.30% 99.32% 99.26% 99.64%
MAPE 4.30% 4.30% 4.90% 4.20% 4.60% 2.60%
Accuracy 95.70% 95.70% 95.10% 95.80% 95.40% 97.40%
RMSE (nm) 0.539 0.471 0.536 0.534 0.507 0.383

error; thus, this model is most suitable for predicting Ra 2. The results of the ANOVA conducted in this study indi-
accurately in the SPDT of ­CaF2. cate that spindle speed, feed rate, and tool rake angle
The main contributions of this study are as follows: affect the surface roughness of ­CaF2 in the aforemen-
tioned process.
1. The number of experiments required for data collection 3. An S-ANN model, an RSM model, and a kriging model
was reduced using the CCD method, which can reduce were developed to model the relationships of spindle
the duration and cost of the SPDT process for ­CaF2. speed, feed rate, and tool rake angle with surface rough-
ness in the SPDT of ­CaF2. The R2 values of these models
786 The International Journal of Advanced Manufacturing Technology (2025) 136:775–787

Table 6  Comparison of No Experimental (nm) RSM (nm) Kriging (nm) S-ANN (nm)
experimental values and
prediction values for RSM, 18 11.927 12.365 13.310 12.046
kriging, and ANN model
19 9.631 8.288 8.802 8.724
20 18.520 17.512 18.500 18.320
21 12.865 13.205 13.400 12.368
22 5.178 5.923 6.866 5.289
23 4.469 3.344 3.677 4.505
24 11.584 11.518 10.560 11.742
25 9.810 8.709 7.879 9.712
R2 94.67% 99.27% 99.64%
MAPE 9.6% 12.9% 2.6%
Accuracy 90.4% 87.1% 97.4%
RMSE (nm) 0.878 1.18 0.383
Max. error (nm) 1.343 1.931 0.907

were 99.64%, 94.67%, and 99.27%, respectively, when 3. Technical Document of Corning Inc. Corning® Calcium fluoride
they were applied to new data, indicating their strong (CaF2) - Code 9575 (2014). [Link]
worldw​ ide/g​ lobal/d​ ocume​ nts/C
​ alciu​ m%2​ 0Fluo​ ride%2​ 0PI%2​ 0She​
generalization ability. Their accuracies were 97.4%, et%​20Sep​tember%​202014.​pdf
90.4%, and 87.1%, respectively. 4. Sun L, Duan W, Wu H, Chen M, Zeng J, Wu Y, Chen Y (2023)
4. Among the aforementioned models, the S-ANN had Investigation on the ultra-precision diamond turning of ZnSe
the highest accuracy (97.40%) and lowest RMSE aspheric surfaces using straight-nosed cutting tools. Journal of
Manufacturing Processes 104:108–122. [Link]
(0.383 nm); thus, this model is most suitable for pre- jmapro.​2023.​09.​01
dicting Ra accurately in the SPDT of ­CaF2. 5. Mizumoto Y, Aoyama T, Kakinuma Y (2011) Basic study on
ultraprecision machining of single-crystal calcium fluoride.
Procedia Engineering 19:264–269. [Link]
proeng.​2011.​11.​11010.​1016/j.​proeng.​2011.​11.​110
Author contribution All the authors contributed to the study concep- 6. Chen X, Xu J, Fang H, Tian R (2017) Influence of cutting
tion and design. Material preparation, data collection, and analysis parameters on the ductile-brittle transition of single-crystal cal-
were performed by Ke-Er Tang, Yin-Chung Huang, Wei-Ting Lin, cium fluoride during ultra-precision cutting. The International
Yuan-Chieh Cheng, and Chun-Wei Liu. The first draft of the manu- Journal of Advanced Manufacturing Technology 89:219–225.
script was written by Ke-Er Tang, and all the authors commented on the [Link]
previous versions of the manuscript. All the authors read and approved 7. H. Xu, X. Zhang, M. Xu, X. Li (2012). Study on the control of
the final manuscript. surface roughness in single point diamond turning. Paper pre-
sented at the 6th International Symposium on Advanced Opti-
Funding This work was supported by the Research Project of cal Manufacturing and Testing Technologies, Xiamen, China.
the National Science and Technology Council, Taiwan (NSTC [Link]
113–2221-E-007–058-MY3) and (NSTC 113–2218-E-007–024-). 8. Yan J, Tamaki J, Syoji K, Kuriyagawa T (2004) Single-point
diamond turning of CaF2 for nanometric surface. Int J Adv
Manuf Technol 24(9):640–646. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 007/​
Declarations s00170-​003-​1747-2
9. M. Azumi and E. Nakahata (2010).Study of relation between
Competing interests The authors declare no competing interests. crystal structure and laser damage of calcium flouride, Proc.
SPIE 7842, Laser-Induced Damage in Optical Materials: 2010.
[Link]
10. Hocheng H, Hsieh ML (2004) Signal analysis of surface rough-
ness in diamond turning of lens molds. Int J Mach Tools Manuf
References 44(15):1607–1618. [Link]
06.​003
1. He X, Yang L, Zhang K, Li R, Peng Y (2024) Research on the 11. GoelSinghSarepaka BBRV (2015) Optimizing single point
shear thickening polishing CaF2 with textured hollow polishing diamond turning for mono-crystalline germanium using grey
tool. J Manuf Process 119:193–203. [Link] relational analysis. Mater Manuf Processes 30(8):1018–1025.
jmapro.​2024.​03.​075 [Link]
2. Retherford RS, SabiaSokira RVP (2001) Effect of surface qual- 12. Chen L, Hu L, Xiao C, Qi Y, Yu B, Qian L (2017) Effect of
ity on transmission performance for (111) CaF2. Appl Surf Sci crystallographic orientation on mechanical removal of CaF2,
183(3–4):264–269. [Link] oi.​o rg/​1 0.​1 016/​S 0169-​4 332(01)​ Wear, Volumes 376–377, Part A. ISSN 409–416:0043–1648.
00587-6 [Link]
The International Journal of Advanced Manufacturing Technology (2025) 136:775–787 787

13. Alauddin M, El Baradie MA, Hashmi MSJ (1995) Computer- Ti-6Al-4V under cryogenic jets applied at flank and rake faces of
aided analysis of a surface-roughness model for end milling. coated WC tool. Int J Adv Manuf Technol 93:975–991. [Link]
J Mater Process Technol 55:123–127. [Link] doi.​org/​10.​1007/​s00170-​017-​0566-9
0924-​0136(95)​01795-X 23. Imanian ME, Biglari FR (2022) Modeling and prediction of sur-
14. Sahin Y, Riza Motorcu A (2005) Surface roughness model face roughness and dimensional accuracy in SLS 3D printing of
for machining mild steel with coated carbide tool. Mater Des PVA/CB composite using the central composite design. J Manuf
26:321–326. [Link] Process 75:154–169. [Link]
15. Yang A, Han Y, Pan Y, Xing H, Li J (2017) Optimum surface 065
roughness prediction for titanium alloy by adopting response 24. Box GEP, Wilson KB (1951) On the experimental attainment of
surface methodology. Results in Physics 7:1046–1050. [Link] optimum conditions. J Roy Stat Soc 13(1):1–45. [Link]
doi.​org/​10.​1016/j.​r inp.​2017.​02.​027 10.​1007/​978-1-​4612-​4380-9_​23
16. Singh A, Garg H, Lall AK (2017) Optical polishing process: anal- 25. Cressie N (1990) The origins of kriging. Math Geol 22:239–252.
ysis and optimization using response surface methodology (RSM) [Link]
for large diameter fused silica flat substrates. J Manuf Process 26. Koehler JR, Owen AB (1996) Computer experiments. Handbook
30:439–451. [Link] of. Statistics 13:261–308. [Link] oi.o​ rg/1​ 0.1​ 016/S
​ 0169-7​ 161(96)​
17. Aramesh M, Shi B, Nassef AO, Attia H, Balazinski M, Kishawy 13011-X
HA (2013) Meta-modeling optimization of the cutting process 27. Lu C, Teng D, Keshtegar B, Alkabaa AS, Taylan O, Fei CW
during turning titanium metal matrix composites (Ti-MMCs). (2023) Extremum hybrid intelligent-inspired models for accurate
Procedia CIRP 8:576–581. [Link] predicting mechanical performances of turbine blisk. Mech Syst
06.​153 Signal Process 190:110136. [Link] oi.o​ rg/1​ 0.1​ 016/j.y​ mssp.2​ 023.​
18. Dang X (2017) Constrained multi-objective optimization of EDM 110136
process parameters using kriging model and particle swarm algo- 28. Linares-Rodriguez A, Ruiz-Arias JA, Pozo-Vazquez D, Tovar-
rithm. Mater Manuf Processes 33(4):397–404. [Link] Pescador J (2013) An artificial neural network ensemble model for
1080/​10426​914.​2017.​12920​37 estimating global solar radiation from Meteosat satellite images.
19. Nguyen T (2019) Prediction and optimization of machining Energy 61:636–645. [Link] oi.o​ rg/1​ 0.1​ 016/j.e​ nergy.2​ 013.0​ 9.0​ 08
energy, surface roughness, and production rate in SKD61 mill-
ing. Measurement 136:525–544. [Link] Publisher's Note Springer Nature remains neutral with regard to
rement.​2019.​01.​009 jurisdictional claims in published maps and institutional affiliations.
20. NatarajanMuthuKaruppuswamy CSP (2011) Prediction and analy-
sis of surface roughness characteristics of a non-ferrous material Springer Nature or its licensor (e.g. a society or other partner) holds
using ANN in CNC turning. Int J Adv Manuf Technol 57:1043– exclusive rights to this article under a publishing agreement with the
1051. [Link] author(s) or other rightsholder(s); author self-archiving of the accepted
21. Khorasani A, Yazdi MRS (2017) Development of a dynamic manuscript version of this article is solely governed by the terms of
surface roughness monitoring system based on artificial neural such publishing agreement and applicable law.
networks (ANN) in milling operation. Int J Adv Manuf Technol
93:141–151. [Link]
22. Mia M, Khan MA, Dhar NR (2017) Study of surface roughness
and cutting forces using ANN, RSM, and ANOVA in turning of