Materials Characterization 221 (2025) 114781

Contents lists available at ScienceDirect

Materials Characterization
journal homepage: www.elsevier.com/locate/matchar

Press-forming of aluminum foam and estimation of its mechanical

properties from X-ray CT images using machine learning
Yoshihiko Hangai * , Yuki Sakaguchi , Kenji Okada , Yuuki Tanaka
Faculty of Science and Technology, Gunma University, Kiryu 376-8515, Japan

A R T I C L E I N F O A B S T R A C T

Keywords: Forming aluminum foam into the desired shape is essential for actual product application, but aluminum foam is
Cellular materials difficult to form. In this investigation, we attempted to press-form aluminum foam immediately after foaming the
Neural network precursor and employed a neural network to estimate the properties of the press-formed aluminum foam from X-
Mechanical properties
ray CT images. It was found that it was possible to press-form the aluminum foam immediately after foaming
X-ray computed tomography (CT)
Deep learning
while maintaining the pores. The resulting aluminum foam exhibited a similar compressive behavior to non-
press-formed aluminum foam. In addition, it was found that a neural network model for the estimation of
plateau stress from X-ray CT images of non-press-formed aluminum foam can be created by training on a dataset
of X-ray CT images and plateau stress obtained from actual compression tests of aluminum foam. From X-ray CT
images, it was also suggested that this neural network model can also be used to estimate the plateau stress of
press-formed aluminum foam that retains pores. That is, it was suggested that the neural network model created
utilizing X-ray CT images can be employed to estimate the properties of products even when they are press-
formed into complex shapes and their properties are difficult to evaluate.

1. Introduction precursor foaming process, a foaming agent is mixed with solid

aluminum, and when the precursor is heated, the foaming agent gen­
Aluminum foam, which is lightweight and exhibits excellent heat- erates gas to form pores in the molten aluminum and cause foaming.
insulating, shock-absorbing, and sound-absorbing properties, is ex­ When the temperature of the precursor exceeds the solidus temperature,
pected to be utilized in the production of building materials and auto­ the formation of pores is observed and foaming begins, followed by
motive components [1–5]. Forming aluminum foam into the desired rapid foaming when the precursor temperature exceeds the liquidus
shape is essential for actual product application, but it is difficult to form temperature [26,27]. It has been shown that aluminum foam can be
aluminum foam, which is mostly composed of pores with thin cell walls. formed after foaming but before the temperature decreases to below the
When aluminum foam is formed at room temperature, as is usually the liquidus temperature, while the foam is still soft, without damaging the
case with dense materials, it is immediately fractured at the cell walls. pores. However, it has not yet been clarified whether the properties of
Consequently, a variety of techniques have been explored, including press-formed aluminum foam are equivalent to those of conventional
laser processing at room temperature [6–9], compression forming aluminum foam. In addition, no technology has been established to
[10–13], incremental forming [14,15], bulge forming [16,17], extrusion ensure that properties similar to those of conventional aluminum foam
and rolling [18–20], and compression forming in a solid–liquid coex­ can be achieved after press-forming. In particular, the mechanical
isting state after heating from room temperature [21,22]. However, the properties of aluminum foam are often determined by compression tests.
forming rate has not been high enough to achieve the large deformations However, when aluminum foam is press-formed into complex shapes, it
possible by press and roll forming, and there have been few studies in is difficult to obtain specimens, and the compression test itself cannot be
this area. performed, making it difficult to evaluate the mechanical properties.
Recently, Hangai and co-authors have shown that in the precursor In recent years, attempts have been made to apply machine learning
foaming process, press and roll forming can be performed immediately to the estimation of the properties of porous material, but these have
after foaming while maintaining the pore structures [23–25]. In the been limited to estimating properties by inputting quantified

* Corresponding author.
E-mail address: [email protected] (Y. Hangai).

https://doi.org/10.1016/j.matchar.2025.114781
Received 24 September 2024; Received in revised form 23 December 2024; Accepted 26 January 2025
Available online 31 January 2025
1044-5803/© 2025 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-
nc/4.0/).
Y. Hangai et al. Materials Characterization 221 (2025) 114781

information such as experimental conditions, manufacturing conditions, mm shift perpendicular to the tool direction at the end of each pass. This
and pore structures [28–33]. Among these studies, Hangai and co- is to obtain a large-area precursor. Following four traverses to ensure the
authors attempted to employ deep learning to estimate the mechanical thorough mixing of the powder into the A1050 plates, four additional
properties of aluminum foam, utilizing only X-ray computed tomogra­ traverses were conducted at the same location [38]. Next, from the
phy (CT) images of the aluminum foam [34–36]. Zhuang et al. have stirred region of FSW, as shown in Fig. 1(c), a precursor of 30 mm × 30
attempted to estimate the mechanical properties of aluminum foam mm × 10 mm thickness was machined. Note that FSW is a solid-state
utilizing deep learning from two-dimensional aluminum foam surface joining process, and it is assumed that the tool was traversed at a tem­
structure images and simulations [37]. This means that we can expect perature lower than the decomposition temperature of TiH2 [45–48],
the mechanical properties of aluminum foam to be estimated without and since it is a short process, TiH2 was not completely decomposed
actually performing compression tests. during FSW.
In this study, we first conducted compression tests on aluminum
foam that had been press-formed immediately after foaming and on non- 2.2. Preparation of non-press-formed aluminum foam
press-formed aluminum foam to investigate whether press-formed
aluminum foam exhibited similar compressive behavior to that of non- As illustrated in Fig. 2(a), the obtained precursor was introduced into
press-formed aluminum foam. In addition, an attempt was made to es­ a stainless-steel mold of 32 mm × 32 mm × 30 mm height and subjected
timate the properties of press-formed aluminum foam from its X-ray CT to heating by three line-type halogen lamps from both the top and the
images with the help of a neural network. We prepared a large number bottom. The voltage and current were approximately 160 V and 8 A,
of non-press-formed aluminum foam specimens, and performed X-ray respectively. By heating from both sides, the thick precursor was heated
CT imaging and compression tests on them. A neural network model for as uniformly as possible to ensure that the timing of foaming was the
estimating mechanical properties was created using supervised learning same throughout the precursor. A video camera was used to observe the
from a dataset of X-ray CT images and mechanical properties. Using this foaming process from the side during the foaming of the precursor. Fig. 2
model, we investigated the possibility of estimating the mechanical (b) shows the actual heating and foaming process. When the foaming
properties of press-formed aluminum foam from its X-ray CT images. If behavior was observed and foamed aluminum appeared from the mold
this estimation can be realized, it is expected that the properties of press- as shown in Fig. 2(b), the lamps were turned off, heating was termi­
formed aluminum foam can be evaluated without performing new me­ nated, and the mold was air-cooled. Fig. 2(c) shows the resulting
chanical properties evaluation tests by creating a model from non-press- aluminum foam. A 20 mm × 20 mm × 20 mm compression test spec­
formed aluminum foam, for which there is a large amount of data imen, as shown in Fig. 2(d), was cut from the center of the aluminum
available. foam using an electric discharge machine. A total of 64 compression test
specimens were prepared. The average porosity of the 64 compression
2. Materials and methods test specimens was 85.8 %, ranging from 78.4 % to 93.0 %.

2.1. Preperation of aluminum foam precursor 2.3. Preparation of press-formed aluminum foam

In this study, aluminum foam was synthesized via the precursor The precursor obtained in Section 2.1 was foamed as described in
foaming method employing friction stir welding (FSW) [38–42]. By Section 2.2. Note that in order to achieve a high pressing ratio, the mold
FSW, precursors can be fabricated from inexpensive plates. A mixture was changed to a low-height, 35 mm × 35 mm × 25-mm-high stainless-
comprising a foaming agent and a pore structures stabilizer, each in steel mold. Immediately after foaming, the foamed samples were press-
powder form, was applied between four 3-mm-thick plates of industrial- formed with a water-cooled copper plate, and the mold was air-cooled.
grade aluminum (A1050) to form stacked plates, as illustrated in Fig. 1 The size of the copper plate was 50 mm × 50 mm × 15 mm thick. The
(a). The foaming agent was composed of titanium hydride (TiH2, water-cooled copper plate was used to solidify the aluminum foam as
Kojundo Chemical Laboratory Co., Ltd., Sakado, Japan) with a particle rapidly as possible to prevent pore floating and pore coalescence after
size of less than 45 μm. The pore structures stabilizer was composed of the press-forming process. A compression test specimen of 20 mm × 20
alumina (Al2O3, Kojundo Chemical Laboratory Co., Ltd., Sakado, Japan) mm × 20 mm, the same dimensions as described in Section 2.2, was cut
with a particle size of approximately 1 μm. The pore structures stabilizer from the aluminum foam sample after press-forming using an electric
enhances the thickening effect of melted aluminum during the precursor discharge machine. A total of 19 compression test specimens were
foaming process, thereby preventing the pores from floating or coa­ prepared.
lescing into a larger pore. The amounts of TiH2 and Al2O3 were
respectively 1 mass% and 5 mass% of the A1050 mass in the volume 2.4. Imaging process using X-ray CT
stirred by a traversing tool during FSW. Next, the FSW tool was traversed
over the powder-dispersed area of the stacked plates, as illustrated in As illustrated in Fig. 3(a), the cross-sectional images of the
Fig. 1(b). The FSW tool had a 9.3-mm-long tapered M8 – M4 threaded compression specimens were acquired using a microfocus X-ray CT
probe with a shoulder diameter of 19 mm manufactured from tool steel. system perpendicular to the direction of compressive loading. Images of
During FSW, the tool rotated at 2200 rpm and traversed 100 mm/min. 512 pixels × 512 pixels were obtained with a tube voltage of 80 kV and a
Four rows were traversed by the multi-pass method [43,44], with a 7.5 current of 30 μA. Fig. 3(b) shows a sample X-ray CT image. The inner
black area shows the pores, the white area is aluminum, and the outer
black area is the atmosphere. The pixel equivalent length was approxi­
mately 60 μm/pixel, and the slice pitch was set to the same value. Note
that the specimen was observed to be tilted in the obtained X-ray CT
image, so it was corrected by rotating the specimen through image
processing to obtain the image shown in Fig. 3(c). Next, an X-ray CT
image of the specimen was trimmed to 310 pixels × 310 pixels, as shown
by the red dotted line in Fig. 3(c), resulting in the image shown in Fig. 3
(d). The sample was trimmed from the original 20 mm on each side to
approximately 19 mm on each side. This is because in the preliminary
test, the neural network model described below was created using an
Fig. 1. Precursor fabrication method by FSW. image of only the sample part without the surrounding atmosphere,

2
Y. Hangai et al. Materials Characterization 221 (2025) 114781

Fig. 2. (a), (b) Foaming process of precursor. (c) Obtained aluminum foam. (d) Obtained compression test specimen.

Fig. 3. (a) X-ray CT imaging direction. (b) Obtained X-ray CT image. (c) Rotation of X-ray CT image. (d) Trimming of X-ray CT image.

which resulted in a more accurate estimate. However, this trimming Thus, the apparent number of specimens was augmented by eight
operation becomes unnecessary when a large number of samples are [51,52]. A total of 512 datasets (the dataset of 64 compression speci­
available and the training data for the neural network model is mens × 8 image data and plateau stresses) were obtained. Matlab release
improved. The number of stacked images was also set to 310 continuous 2022a was used to create a neural network model. Of the 64 non-press-
images, corresponding to the trimmed size of the cross section. Image formed compression specimens, 63 × 8 datasets corresponding to the 63
processing such as that involving rotation and trimming was performed compression specimens were employed as training data, while 1 × 8
using ImageJ software. A group of 310 pixels × 310 pixels × 310 images datasets corresponding to the remaining specimens were utilized as test
obtained in this way was used as image data for the neural network. data. Once the training phase was complete, tests were performed on the
neural network model with the designated test datasets. Eight plateau
stresses were estimated from the tests performed on each of the eight
2.5. Compression test procedure
datasets. Subsequently, the average of the eight estimates was employed
to define the estimated plateau stress σ p of the corresponding specimen.
The compression specimens were weighed and measured after being
In this way, the training and test data specimens were varied sequen­
subjected to X-ray CT inspection, and the porosity p was calculated from
tially, and the estimation was performed for all specimens.
these measurements prior to the compression test. Porosity is the volume
Next, a neural network model trained on a dataset of non-press-
fraction of pores in a sample [49] and is an indicator of weight reduc­
formed compression specimens (64 compression specimens × 8 sets)
tion. The test was performed using a universal testing machine in
was created to estimate the plateau stress of the press-formed
accordance with Japanese Industrial Standards [50] at 1.2 mm/min
compression specimens. For the test data, similar to the non-press-
compression rate. Using the load and displacement output from the
formed compression specimens, 310 pixels × 310 pixels × 310 image
testing machine, stress σ was calculated as the load divided by the initial
data were obtained from the X-ray CT images of each press-formed
cross-sectional area (20 mm × 20 mm) of the compression specimen,
compression specimen and padded by a factor of 8. Estimations were
and strain ε was calculated as the displacement divided by the initial
made for each of the eight specimens, and the average of the eight es­
height (20 mm) of the compression specimen. The compressive defor­
timates was determined as the estimated plateau stress σp for that
mation behavior was recorded by a video camera during the compres­
specimen.
sion test. In this study, the plateau stress was evaluated as a
Similar to the plateau stress estimation, the porosity estimation was
representative parameter of the mechanical properties predicted by the
also performed in the same manner. A dataset corresponding to X-ray CT
neural network. However, other parameters, such as energy absorption,
images and porosity was obtained, and a similar procedure was used to
could be evaluated similarly to the present study. The plateau stress was
create a model trained on a dataset of non-press-formed compression
determined by calculating the average stress at ε = 20–30 % within the
specimens (64 compression specimens × 8 sets). Using the created
σ –ε curve [50].
model, we investigated whether the porosity of press-formed aluminum
foam could be estimated from X-ray CT image data alone.
2.6. Neural network modeling method
3. Results and discussion
First, as a preliminary evaluation, we examined the estimation ac­
curacy of the neural network model using a non-press-formed 3.1. Press-forming results
compression test specimen. Supervised learning was performed on 64
non-press-formed compression specimens using a set of 310 pixels × 310 The precursor was foamed in the mold, as shown in Fig. 4(a). The
pixels × 310 image data obtained by the X-ray CT inspection of each foaming time was approximately 200 s. Then, as shown in Fig. 4(b), a
specimen and plateau stresses. To augment the quantity of training data, water-cooled copper plate was press-formed by hand against the foamed
the two-dimensional cross-sectional images, as represented in Fig. 3(d), sample immediately after heating was stopped. Aluminum foam was soft
were rotated by 90◦ , 180◦ , and 270◦ rightward and inverted using enough immediately after heating and could be press-formed with
ImageJ image processing software. That is, for each compression spec­ almost no resistance. Because the degree of foaming varied slightly from
imen, eight 310 pixels × 310 pixels × 310 image data were obtained.

3
Y. Hangai et al. Materials Characterization 221 (2025) 114781

Fig. 4. (a) Before pressing (after foaming). (b) Press-forming. (c) Press-formed aluminum foam. (d) Press-formed compression specimen.

sample to sample, the amount of pressing varied. In the video image of image. Fig. 5(e) shows the σ –ε curve obtained from the compression
the foaming process shown in Fig. 4(a), the height of the mold is defined tests. First, there was an elastic region, followed soon after by a plateau
as h0 and the part of the foamed aluminum protruding from the mold is region where the stress was almost constant. Sequential changes in the
defined as h. Press-forming was performed by pressing the water-cooled deformed layers caused the appearance of these plateau regions. When
copper plate to the top of the mold. Therefore, the amount of pressing all of the layers had been deformed, a densification zone appeared at the
was estimated as h/(h + h0). The average amount of pressing of the 19 end, where the entire layer was densified and the stress increased
press-formed samples obtained was 30.7 %. Although it is possible to rapidly. The plateau stress of the specimen was 3.48 MPa.
apply a larger amount of pressing to aluminum foam immediately after Figs. 6(a)–(c) show the compression test of press-formed aluminum
foaming [53], pressing was limited to the top of the die to maintain the foam with a porosity of p = 86.9 %. The compressive deformation
size of the compression specimens obtained in this study. An aluminum behavior shows that, as with the non-press-formed aluminum foam
foam sample after press-forming is shown in Fig. 4(c). By press-forming, shown in Fig. 5, the deformed layers changed sequentially as they were
the mountain-shaped aluminum foam, as shown in Fig. 4(a), could be compressed. Fig. 6(d) is an X-ray CT image of a longitudinal section of a
formed into a flat surface. A compression test specimen was cut from this specimen before compression. The X-ray CT image shows that pores are
sample, as shown in Fig. 4(d). Pores were observed on all surfaces of the present throughout the specimen, even after press-forming, as in non-
compression test specimens, and no pores appeared to have been flat­ press-formed aluminum foam. Furthermore, the σ–ε curve shown in
tened by the press-forming process in certain locations. This is consistent Fig. 6(e) also shows an elastic region, a plateau region, and a densifi­
with the results of the in-situ observation of press-forming by X-ray cation region, as in Fig. 5(e). The plateau stress of this specimen was
transmission [24], and suggests that press-forming is possible while 2.21 MPa. The same trend was observed for all other press-formed
maintaining pores. The average porosity of the 19 compression test specimens as well as the non-press-formed specimens. The compres­
specimens obtained was 84.3 %, ranging from 80.1 % to 89.1 %, and was sion behavior and σ –ε curve of aluminum foam specimens prepared by
similar to that of non-press-formed aluminum foam, described in section other methods found in the literature also show similar trends [54,55].
2.2. Therefore, it was shown that, even after press-forming, the pores were
maintained and similar compression deformation behaviors and σ –ε
curves to those of non-press-formed aluminum foam were observed.
3.2. Compression test results Generally, when aluminum foam is press-formed at room temperature,
only one layer is deformed and the pores in that area collapse, as in the
Figs. 5(a)–(c) show the compression test of non-press-formed compression test described in this section. In contrast, as shown in Fig. 4,
aluminum foam with a porosity of p = 82.8 %. The compression test it is possible to press-form the aluminum foam immediately after
shows that the deformation was not uniform throughout the specimen, foaming while maintaining pores throughout the sample.
but that the deformed layers changed sequentially. Fig. 5(d) shows an X-
ray CT image of a longitudinal section of a compression specimen before
compression. The vertical cross-sectional X-ray CT image shown in Fig. 3
(d) was reconstructed into a longitudinal section by image processing,
and compression tests were performed in the vertical direction of this

Fig. 5. Compression test results of non-press-formed aluminum foam with a porosity of p = 82.8 %.

4
Y. Hangai et al. Materials Characterization 221 (2025) 114781

Fig. 6. Compression test results of press-formed aluminum foam with a porosity of p = 86.9 %.

3.3. Verification of plateau stress estimation of non-press-formed where N is the number of specimens, 64 in this study. The average error
specimens calculated using these equations was 0.37 MPa and the maximum error
was 1.74 MPa. Although there is a certain amount of variation, it can be
Fig. 7 shows the relationship between the plateau stress σ a obtained seen that the estimation had a certain degree of accuracy. The largest
from the actual compression test and the estimated plateau stress σ p , errors occurred for the specimens at σa = 6.14 MPa, where the plateau
which is the result of creating a neural network model trained on 63 of stress was high. The reason for this is that the number of specimens with
the 64 non-press-formed specimens and then estimating the plateau high plateau stress was small and an insufficient amount of data was
stress on the remaining sample. The dotted line in the graph shows the collected, and when this specimen was selected for testing, the majority
relationship σ a = σ p . The closer the resulting plots are to this line, the of the training data contained specimens with plateau stress lower than
greater the accuracy of the estimation. Although there is some variation, this plateau stress, so the estimated value was lower than the actual
it can be seen that there is a tendency for the data to be gathered along a plateau stress. Estimation accuracy is expected to improve as the num­
straight line. ber of samples increases in the future.
The absolute value of the error between the estimated plateau stress
σ p and the plateau stress σa obtained from the actual compression test of
3.4. Results of plateau stress estimation for press-formed specimens
specimens, σer i , was calculated using
⃒ p ⃒
er ⃒
σi = σi − σi , a⃒
(1) Fig. 8 shows the relationship between the plateau stress σa obtained
from the actual compression test and the estimated plateau stress σp ,
where i is the number specimens, which ranges from 1 to 64 in this which is the result of estimating the plateau stress for each of the 19
study. The average error σ er for the 64 specimens was calculated using press-formed specimens using the neural network model trained on the
data of the 64 non-press-formed specimens. The average error is 0.33
N
1 ∑ MPa and the maximum error is 0.95 MPa, indicating that the estimation
σer = σer , (2)
N i=1 i is as accurate as that of the non-press-formed specimens. The average
computation time for creating this neural network model was 2091 s for

7 7
/ MPa
/ MPa

6 6

5 5
Predicted plateau stress,
Predicted plateau stress,

4 4

3 3

2 2

1 1

0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Actual plateau stress, / MPa Actual plateau stress, / MPa

Fig. 7. Relationship between plateau stress σa obtained from the actual Fig. 8. Relationship between plateau stress σ a obtained from the actual
compression test and plateau stress σp estimated using the machine learning compression test and plateau stress σ p estimated for press-formed specimens
model for non-press-formed specimens. The dotted line in the graph shows the using the machine learning model trained on the data of the non-press-
relationship σ a = σ p . formed specimens.

5
Y. Hangai et al. Materials Characterization 221 (2025) 114781

the three trials. In contrast, the average computation time for the esti­ 95
mation of 8 padded datasets together for a single sample was 1.36 s for
the 5 samples. That is, by creating a model that can estimate the plateau

Predicted porosity, p p (%)

stress from X-ray CT images of non-press-formed aluminum foam sam­ 90
ples with the same degree of foaming of the same alloy, it is possible to
estimate the plateau stress of press-formed aluminum foam by nonde­
85
structive observation in a short computation time. In addition, the
plateau stress estimation for the non-press-formed aluminum foam
sample shown in Fig. 7 and that for the press-formed aluminum foam 80
sample shown in Fig. 8 were similar in accuracy. Therefore, by
increasing the number of samples for training the neural network model
created for the non-press-formed aluminum foam shown in Fig. 7 and 75
improving the estimation accuracy of the model, it is expected that the 75 80 85 90 95
estimation accuracy for the press-formed aluminum foam will also Actual porosity, p a (%)
improve.
Fig. 10. Relationship between actual porosity pa and porosity estimated using
machine learning model, pp.
3.5. Plateau stress estimation for a press-formed specimen using a model
created from a press-formed specimen formed aluminum foam specimens and a dataset of porosity p. The
model was then used to estimate the porosity of 19 press-formed
In the previous section, it was shown that the plateau stress of the aluminum foam specimens. The average error is 0.7 % and the
press-formed specimens can be estimated using the neural network maximum error is 1.6 %, indicating that the estimation is relatively
model trained on the data of the 64 non-press-formed specimens. In this accurate. Generally, porosity can also be obtained by binarizing the X-
section, we attempted to conduct training and estimation on 19 press- ray CT image to separate the aluminum and pore areas. However, slight
formed specimens and compared the results with those in section 3.4. variations is caused owing to the variances in the threshold for binar­
Fig. 9 shows the relationship between the plateau stress σ a obtained ization and changes depending on the resolution of the image [56].
from the actual compression test and the estimated plateau stress σ p , Therefore, the estimation accuracy can be improved if the X-ray CT
which is the result of the same preliminary validation as described in image can be used as-is for estimation.
Section 3.3, where a training model was created for 18 of the 19 press- These results suggest that the neural network model can be used to
formed specimens and the model was used to estimate the plateau stress estimate properties such as the plateau stress and porosity of aluminum
for the remaining specimen. The average error is 0.38 MPa and the foam from X-ray CT images obtained by nondestructive observation.
maximum error is 1.36 MPa, indicating that the estimation results are This method can be applied not only to aluminum foam fabricated by
comparable to those shown in Fig. 8 in Section 3.4. Press-formed spec­ press-forming immediately after precursor foaming, but also to
imens are often assumed to have complex shapes, and it is difficult to aluminum foam fabricated by other methods that can fabricate complex
obtain compression specimens from complex-shaped aluminum foam, shapes using molds, such as the casting method [55] and the method of
making it difficult to evaluate the properties and create a neural network foaming precursor in a mold [57,58]. Therefore, it is expected that it will
model. Even if the compression specimens can be obtained and a model be possible to evaluate the properties of aluminum foam formed into
can be created, long computing times are still required. By pre-creating a complex shapes without having to perform actual property tests. In
neural network model using aluminum foam that is not press-formed as addition, quality is assured by inspecting all products through nonde­
in this study, it is assumed that the estimation can be performed in a structive product characterization.
short time without having to re-create the model for each product.
4. Conclusions
3.6. Porosity estimation
In this study, we attempted to press-form aluminum foam immedi­
Fig. 10 shows the relationship between the actual porosity pa and the ately after foaming the precursor and used a neural network to estimate
estimated porosity pp which is the result of porosity estimation using a the properties of the press-formed aluminum foam from X-ray CT im­
neural network model trained on X-ray CT images of 64 non-press- ages. The following results were obtained.

(1) It was found that it was possible to press-form the aluminum foam
7 immediately after foaming while maintaining the pores. The
/ MPa

6 resulting aluminum foam exhibited a similar compressive

behavior to non-press-formed aluminum foam.
5 (2) It was found that a neural network model for estimating plateau
Predicted plateau stress,

stress from X-ray CT images of non-press-formed aluminum foam

4
can be created by training on a dataset of X-ray CT images and
3 plateau stress obtained from actual compression tests of
aluminum foam.
2 (3) It was suggested that the neural network model for estimating
1 plateau stress from X-ray CT images of non-press-formed
aluminum foam can also be used to estimate the plateau stress
0 of press-formed aluminum foam that maintains pores from X-ray
0 1 2 3 4 5 6 7 CT images. In addition, the plateau stress of press-formed
Actual plateau stress, / MPa aluminum foam can be estimated in less computation time than
that needed to create a neural network model.
a
Fig. 9. Relationship between plateau stress σ obtained from the actual (4) Similarly, porosity can be estimated from X-ray CT images by
compression test and estimated plateau stress σ p for press-formed specimens creating a neural network model. That is, it was suggested that
using machine learning model trained on the press-formed specimens.

6
Y. Hangai et al. Materials Characterization 221 (2025) 114781

the neural network model created using X-ray CT images can be [16] H. Nassar, M. Albakri, H. Pan, M. Khraisheh, On the gas pressure forming of
aluminium foam sandwich panels: experiments and numerical simulations, CIRP
used to estimate the properties of products even when they are
Ann. 61 (2012) 243–246, https://doi.org/10.1016/j.cirp.2012.03.116.
press-formed into complex shapes and their properties are diffi­ [17] H. Mata, A.D. Santos, M.P.L. Parente, R.A.F. Valente, A.A. Fernandes, R.N. Jorge,
cult to evaluate. Study on the forming of sandwich shells with closed-cell foam cores, Int. J. Mater.
Form. 7 (2014) 413–424, https://doi.org/10.1007/s12289-013-1136-9.
[18] W.Y. Kim, W.J. Kim, H. Utsunomiya, Accelerated formation of an ultrafine-grained
CRediT authorship contribution statement microstructure in closed-cell aluminum foam after extrusion and differential speed
rolling, Mater. Trans. 58 (2017) 291–293, https://doi.org/10.2320/matertrans.L-
M2016838.
Yoshihiko Hangai: Writing – review & editing, Writing – original
[19] M. Weiss, B. Abeyrathna, M. Pereira, Roll formability of aluminium foam sandwich
draft, Resources, Funding acquisition, Conceptualization. Yuki Saka­ panels, Int. J. Adv. Manuf. Technol. 97 (2018) 953–965, https://doi.org/10.1007/
guchi: Visualization, Software, Investigation. Kenji Okada: Validation, s00170-018-1945-6.
Software. Yuuki Tanaka: Writing – review & editing, Validation, [20] T.R. Neu, K. Heim, W. Seeliger, P.H. Kamm, F. García-Moreno, Aluminum foam
sandwiches: a lighter future for Car bodies, JOM (2024), https://doi.org/10.1007/
Methodology. s11837-024-06460-2.
[21] Z.-y. Liu, Y. Cheng, Y.-x. Li, X. Zhou, X. Chen, N.-z. Wang, Shape formation of
closed-cell aluminum foam in solid–liquid–gas coexisting state, Int. J. Miner.
Declaration of competing interest Metall. Mater. 25 (2018) 974–980, https://doi.org/10.1007/s12613-018-1647-y.
[22] Z. Liu, Y. Cheng, Y. Li, N. Wang, X. Zhou, Study on deformation of closed-cell
aluminum foam in different solid–liquid–gas coexisting state, Met. Mater. Int. 27
The authors declare that they have no known competing financial (2021) 403–412, https://doi.org/10.1007/s12540-019-00576-w.
interests or personal relationships that could have appeared to influence [23] Y. Hangai, M. Ohashi, R. Nagahiro, K. Amagai, T. Utsunomiya, N. Yoshikawa, Press
the work reported in this paper. forming of aluminum foam during foaming of precursor, Mater. Trans. 60 (2019)
2464–2469, https://doi.org/10.2320/matertrans.MT-M2019192.
[24] Y. Hangai, D. Kawato, M. Ando, M. Ohashi, Y. Morisada, T. Ogura, et al.,
Acknowledgments Nondestructive observation of pores during press forming of aluminum foam by X-
ray radiography, Mater. Charact. 170 (2020) 110631, https://doi.org/10.1016/j.
matchar.2020.110631.
This work was financially supported partly by Light Metal Education [25] Y. Hangai, K. Suzuki, M. Ohashi, H. Mitsugi, K. Amagai, Roll forming of aluminum
Foundation, Mitutoyo Association for Science and Technology (MAST) foam immediately after precursor foaming, Res. Eng. 10 (2021) 100224, https://
and Gunma University for the promotion of scientific research. doi.org/10.1016/j.rineng.2021.100224.
[26] I. Duarte, J. Banhart, A study of aluminium foam formation - kinetics and
microstructure, Acta Mater. 48 (2000) 2349–2362, https://doi.org/10.1016/
Data availability S1359-6454(00)00020-3.
[27] Y. Hangai, K. Takada, H. Fujii, Y. Aoki, Y. Aihara, R. Nagahiro, et al., Foaming of
A1050 aluminum precursor by generated frictional heat during friction stir
Data will be made available on request.
processing of steel plate, Int. J. Adv. Manuf. Technol. 106 (2020) 3131–3137,
https://doi.org/10.1007/s00170-019-04834-4.
References [28] Z.M. Yaseen, R.C. Deo, A. Hilal, A.M. Abd, L.C. Bueno, S. Salcedo-Sanz, et al.,
Predicting compressive strength of lightweight foamed concrete using extreme
learning machine model, Adv. Eng. Softw. 115 (2018) 112–125, https://doi.org/
[1] F. García-Moreno, Commercial applications of metal foams: their properties and
10.1016/j.advengsoft.2017.09.004.
production, Materials 9 (2016) 85, https://doi.org/10.3390/ma9020085.
[29] T. Nguyen, A. Kashani, T. Ngo, S. Bordas, Deep neural network with high-order
[2] T. Wan, Y. Liu, C. Zhou, X. Chen, Y. Li, Fabrication, properties, and applications of
neuron for the prediction of foamed concrete strength, Comput. Aided Civ. Inf.
open-cell aluminum foams: a review, J. Mater. Sci. Technol. 62 (2021) 11–24,
Eng. 34 (2019) 316–332, https://doi.org/10.1111/mice.12422.
https://doi.org/10.1016/j.jmst.2020.05.039.
[30] M. Dudzik, A.M. Stręk, ANN architecture specifications for modelling of open-cell
[3] C. Ji, H. Huang, T. Wang, Q. Huang, Recent advances and future trends in
aluminum under compression, Math. Probl. Eng. 2020 (2020) 2834317, https://
processing methods and characterization technologies of aluminum foam
doi.org/10.1155/2020/2834317.
composite structures: a review, J. Manuf. Process. 93 (2023) 116–152, https://doi.
[31] E. Avalos-Gauna, Y.Y. Zhao, L. Palafox, P. Ortiz-Monasterio-Martinez, Porous metal
org/10.1016/j.jmapro.2023.03.015.
properties analysis: a machine learning approach, JOM 73 (2021) 2039–2049,
[4] D. Chen, K. Gao, J. Yang, L. Zhang, Functionally graded porous structures:
https://doi.org/10.1007/s11837-021-04695-x.
analyses, performances, and applications – a review, Thin-Walled Struct. 191
[32] A.E. Rodríguez-Sánchez, H. Plascencia-Mora, A machine learning approach to
(2023) 111046, https://doi.org/10.1016/j.tws.2023.111046.
estimate the strain energy absorption in expanded polystyrene foams, J. Cell. Plast.
[5] W. Fu, Y. Li, Fabrication, processing, properties, and applications of closed-cell
58 (2022) 399–427, https://doi.org/10.1177/0021955x211021014.
aluminum foams: a review, Materials 17 (2024) 560, https://doi.org/10.3390/
[33] H.S. Ullah, R.A. Khushnood, F. Farooq, J. Ahmad, N.I. Vatin, D.Y.Z. Ewais,
ma17030560.
Prediction of compressive strength of sustainable foam concrete using individual
[6] Y.P. Kathuria, A preliminary study on laser assisted aluminum foaming, J. Mater.
and ensemble machine learning approaches, Materials 15 (2022) 3166, https://
Sci. 38 (2003) 2875–2881, https://doi.org/10.1023/a:1024488503856.
doi.org/10.3390/ma15093166.
[7] A. Guglielmotti, F. Quadrini, E.A. Squeo, V. Tagliaferri, Laser bending of aluminum
[34] Y. Hangai, K. Okada, Y. Tanaka, T. Matsuura, K. Amagai, R. Suzuki, et al.,
foam sandwich panels, Adv. Eng. Mater. 11 (2009) 902–906, https://doi.org/
Classification of mechanical properties of aluminum foam by machine learning,
10.1002/adem.200900111.
Mater. Trans. 63 (2022) 257–260, https://doi.org/10.2320/matertrans.MT-
[8] F. Quadrini, A. Guglielmotti, E.A. Squeo, V. Tagliaferri, Laser forming of open-cell
M2021130.
aluminium foams, J. Mater. Process. Technol. 210 (2010) 1517–1522, https://doi.
[35] Y. Hangai, S. Ozawa, K. Okada, Y. Tanaka, K. Amagai, R. Suzuki, Machine learning
org/10.1016/j.jmatprotec.2010.04.010.
estimation of plateau stress of aluminum foam using X-ray computed tomography
[9] A. Changdar, S.S. Chakraborty, Laser processing of metal foam - a review, J. Manuf.
images, Materials 16 (2023) 1894, https://doi.org/10.3390/ma16051894.
Process. 61 (2021) 208–225, https://doi.org/10.1016/j.jmapro.2020.10.012.
[36] Y. Hangai, Y. Sakaguchi, Y. Kitahara, T. Takagi, O. Kenji, T. Yuuki, Plateau stress
[10] D. Contorno, L. Filice, L. Fratini, F. Micari, Forming of aluminum foam sandwich
estimation of aluminum foam by machine learning using X-ray computed
panels: numerical simulations and experimental tests, J. Mater. Process. Technol.
tomography images, Int. J. Adv. Manuf. Technol. 132 (2024) 5053–5061, https://
177 (2006) 364–367, https://doi.org/10.1016/j.jmatprotec.2006.04.028.
doi.org/10.1007/s00170-024-13670-0.
[11] J. Banhart, H.W. Seeliger, Aluminium foam Sandwich panels: manufacture,
[37] W. Zhuang, E. Wang, H. Zhang, Prediction of the compressive mechanical
metallurgy and applications, Adv. Eng. Mater. 10 (2008) 793–802, https://doi.org/
properties and reverse structural design of two-dimensional mesoscopic aluminum
10.1002/adem.200800091.
foam based on deep learning methods, J. Mater. Sci. 59 (2024) 11416–11439,
[12] F. Gagliardi, L. Filice, D. Umbrello, R. Shivpuri, Forging of metallic foams to
https://doi.org/10.1007/s10853-024-09866-0.
reproduce biomechanical components, Mater. Sci. Eng. A 480 (2008) 510–516,
[38] Y. Hangai, T. Utsunomiya, M. Hasegawa, Effect of tool rotating rate on foaming
https://doi.org/10.1016/j.msea.2007.07.045.
properties of porous aluminum fabricated by using friction stir processing,
[13] Z.-Y. Cai, X. Zhang, X.-B. Liang, Multi-point forming of sandwich panels with egg-
J. Mater. Process. Technol. 210 (2010) 288–292, https://doi.org/10.1016/j.
box-like cores and failure behaviors in forming process: analytical models,
jmatprotec.2009.09.012.
numerical and experimental investigations, Mater. Des. 160 (2018) 1029–1041,
[39] Q. Pang, Z.L. Hu, J.S. Song, Preparation and mechanical properties of closed-cell
https://doi.org/10.1016/j.matdes.2018.10.037.
CNTs-reinforced Al composite foams by friction stir welding, Int. J. Adv. Manuf.
[14] K.P. Jackson, J.M. Allwood, M. Landert, Incremental forming of sandwich panels,
Technol. 103 (2019) 3125–3136, https://doi.org/10.1007/s00170-019-03765-4.
J. Mater. Process. Technol. 204 (2008) 290–303, https://doi.org/10.1016/j.
[40] R. Shandley, S. Maheshwari, A.N. Siddiquee, S. Mohammed, D.L. Chen, Foaming of
jmatprotec.2007.11.117.
friction stir processed Al/MgCO3 precursor via flame heating, Mater. Res. Express.
[15] R. Matsumoto, H. Tsuruoka, M. Otsu, H. Utsunomiya, Fabrication of skin layer on
7 (2020) 026515, https://doi.org/10.1088/2053-1591/ab6ef0.
aluminum foam surface by friction stir incremental forming and its mechanical
[41] I.G. Papantoniou, D.E. Manolakos, Fabrication and characterization of aluminum
properties, J. Mater. Process. Technol. 218 (2015) 23–31, https://doi.org/
foam reinforced with nanostructured γ-Al2O3 via friction stir process for enhanced
10.1016/j.jmatprotec.2014.11.030.

7
Y. Hangai et al. Materials Characterization 221 (2025) 114781

mechanical performance, Int. J. Adv. Manuf. Technol. 130 (2024) 5359–5368, [50] JIS-H-7902, Method for Compressive Test of Porous Metals, Japanese Standards
https://doi.org/10.1007/s00170-024-13044-6. Association, 2016.
[42] D. Lohani, B.S.S. Daniel, Quasi-static deformation mechanism and compressive [51] P. Chlap, H. Min, N. Vandenberg, J. Dowling, L. Holloway, A. Haworth, A review of
properties of aluminum foams fabricated by friction stir processing, J. Mater. Eng. medical image data augmentation techniques for deep learning applications,
Perform. (2024), https://doi.org/10.1007/s11665-024-09508-1. J. Med. Imaging Radiat. Oncol. 65 (2021) 545–563, https://doi.org/10.1111/
[43] Y.S. Sato, S.H.C. Park, A. Matsunaga, A. Honda, H. Kokawa, Novel production for 1754-9485.13261.
highly formable mg alloy plate, J. Mater. Sci. 40 (2005) 637–642, https://doi.org/ [52] P. Kokol, M. Kokol, S. Zagoranski, Machine learning on small size samples: a
10.1007/s10853-005-6301-1. synthetic knowledge synthesis, Sci. Prog. (2022) 105, https://doi.org/10.1177/
[44] K. Nakata, Y.G. Kim, H. Fujii, T. Tsumura, T. Komazaki, Improvement of 00368504211029777.
mechanical properties of aluminum die casting alloy by multi-pass friction stir [53] Y. Hangai, M. Ando, M. Ohashi, K. Amagai, Fabrication of two-layered aluminum
processing, Mater. Sci. Eng. A 437 (2006) 274–280, https://doi.org/10.1016/j. foam with closed-cell and open-cell structures and shaping of closed-cell layer by
msea.2006.07.150. press forming immediately after foaming, Metals 11 (2021) 140, https://doi.org/
[45] D.H. Yang, B.Y. Hur, D.P. He, S.R. Yang, Effect of decomposition properties of 10.3390/met11010140.
titanium hydride on the foaming process and pore structures of Al alloy melt foam, [54] L.J. Gibson, Mechanical behavior of metallic foams, Annu. Rev. Mater. Sci. 30
Mater. Sci. Eng. A 445-446 (2007) 415–426, https://doi.org/10.1016/j. (2000) 191–227, https://doi.org/10.1146/annurev.matsci.30.1.191.
msea.2006.09.064. [55] T. Miyoshi, M. Itoh, S. Akiyama, A. Kitahara, ALPORAS aluminum foam:
[46] H. Wang, D.F. Zhu, S. Hou, D.H. Yang, T.G. Nieh, Z.P. Lu, Cellular structure and production process, properties, and applications, Adv. Eng. Mater. 2 (2000)
energy absorption of AlCu alloy foams fabricated via a two-step foaming method, 179–183, https://doi.org/10.1002/(SICI)1527-2648(200004)2:4<179::AID-
Mater. Des. 196 (2020) 109090, https://doi.org/10.1016/j.matdes.2020.109090. ADEM179>3.0.CO;2-G.
[47] G. Yuan, Y. Li, X. Zhou, L. Hu, Preparation of complex shaped aluminum foam by a [56] Y. Hangai, S. Maruhashi, S. Kitahara, O. Kuwazuru, N. Yoshikawa, Nondestructive
novel casting-foaming method, Mater. Lett. 293 (2021) 129673, https://doi.org/ quantitative evaluation of porosity volume distribution in aluminum alloy die
10.1016/j.matlet.2021.129673. castings by fractal analysis, Metall. Mater. Trans. A 40 (2009) 2789–2793, https://
[48] Y. Hangai, T. Takagi, S. Koyama, R. Suzuki, Y. Kamakoshi, Refoaming of deformed doi.org/10.1007/s11661-009-0009-9.
aluminum foam fabricated by precursor foaming process using remaining foaming [57] I. Duarte, M. Vesenjak, M.J. Vide, Automated continuous production line of parts
agent and densification using friction stir welding, Mater. Lett. 351 (2023) 135008, made of metallic foams, Metals 9 (2019) 531, https://doi.org/10.3390/
https://doi.org/10.1016/j.matlet.2023.135008. met9050531.
[49] JIS-H-7009, Glossary of Terms Used in Porous Metals, Japanese Standards [58] J. Banhart, Light-metal foams - history of innovation and technological challenges,
Association, 2016. Adv. Eng. Mater. 15 (2013) 82–111, https://doi.org/10.1002/adem.201200217.