Journal of Biomaterials and Nanobiotechnology, 2011, 2, 98-101
doi:10.4236/jbnb.2011.21013 Published Online January 2011 (http://www.SciRP.org/journal/jbnb)
Predicting the Mechanical Properties of BHA-Li2O
Composites Using Artificial Neural Networks
Hasan Huseyin Celik1*, Oguzhan Gunduz2,6*, Nazmi Ekren3,6, Zeeshan Ahmad4, Faik Nuzhet Oktar5,6
1School of Technical Education, Department of Electronic and Computer Education, Marmara University, Istanbul, Turkey;
2Department of Metal Education, Marmara University, Istanbul, Turkey; 3 School of Technical Education, Department of Electrical
and Electronic Engineering, Marmara University, Istanbul, Turkey; 4School of Pharmacy and Biomedical, Science Unive r sity of
Portsmouth, Portsmouth, UK; 5School of Health Related Professions, Department of Medical Imaging Technics, Marmara University,
Istanbul, Turkey; 6Nanotechnology and Biomaterials Application & Research Centre, Marmara University, Istanbul, Turkey.
Received October 13th, 2010; revised November 20th, 2010; accepted November 28th, 2010.
ABSTRACT
In this study the mechanical properties of bovine hydroxyapatite (BHA)-Li2O composites are predicted using artificial
neural networks (ANN) and then compared with obtained experimental values. BHA was mixed with lithium carbonate
(Li2CO3) and sintered at various temperatures between 900-1300°C. Selected experimental values obtained for the
compression strength, microhardness and density were used to define and train the ANN system. Intermediate data val-
ues not used to train the ANN model were then used to compare and determine the reliability of the ANN system. The
results demonstrate the viable potential in using the ANN approach in predicting mechanical properties even with lim-
ited data sets.
Keywords: Artificial Neural Network, Hydroxyapatite, Li2O Composites, Bioceramic
1. Introduction
Hydroxyapatite (HA) is an attractive material used in
several hard tissue orthopedic applications. Its crystallo-
graphic and chemical properties closely resemble those
of bone and tooth minerals [1]. However, because of its
poor mechanical properties, HA is often used in devices
as a coating, where the bioactive properties of the mate-
rial can still be utilized, whilst benefiting from the me-
chanical properties of the substrate (e.g. stainless steel).
The mechanical properties of HA, however, may be im-
proved by additives, which can modify the sintering
process, allowing the resultant material to achieve higher
densities [2]. One such additive that is of interest is lith-
ium and lithium based composites, which have demon-
strated reasonable biocompatibility [3]. The relative
knowledge concerning the effects of lithium compounds
in such orthopedic systems on living organism is scarce;
and for this reason reduced quantities of such materials
must be utilized e.g. Fanowich et al. have utilized mini-
mal quantities of lithium in their research (0.2, 0.4 and
0.6 wt. %) [2,4].
Artificial Neural Networks (ANN) are parallel proc-
essing models which can be deployed in predicting
outcomes based on input-output data (parameters and
relationships), demonstrating a degree of robustness
and self-learning capability [5]. The accuracy and reli-
ability of predicted outcomes can be increased based on
the number of input values. In this way accumulation
of data sets into an ANN system can enhance the out-
come and predictability as the likelihood of any value
will be based on previous experimental data. This
makes the ANN system robust as it is an adaptive sys-
tem which can change based on information it has al-
ready acquired or will obtain. This study describes the
use of ANN to predict the mechanical properties of
bovine hydroxyapatite (BHA)-Li2O composites. Pa-
rameters of compression strength, microhardness and
density were produced by experiments in various mix-
ture rates and temperatures. The model used was based
on a feed-forward system based on sintering tempera-
ture and material composition.
BHA lithium carbonate composite, like other bioac-
tive composites are key mater ials in the develop ment of
several orthopedic applications. However, sintering
temperatures and compositional values for such mate-
rials occupy considerable time periods (impact on me-
chanical properties) and therefore the outlook provided

Predicting the Mechanical Properties of BHA-Li2O Composites Using Artificial Neural Networks 99
by ANN systems can be extremely valuable. If a suffi-
cient number of data sets can be provided then pre-
dicted data from ANN systems can suffice; reducing
preliminary testing to determine mechanical properties
for each type of composite or sintering temperature. In
addition to orthopedics, bioceramics and their compos-
ites are also finding applications in other areas of bio-
materials and biotechnologies (such as drug delivery
and tissue engineering) and these predictive methods
may also assist in enhancing these frontiers.
2. Materials and Methods
2.1. Materials
BHA, was prepared in similar fashion to an earlier study
[6]. Lithium carbonate (Li2CO3) was purchased from
(Sigma Aldrich) and was added to BHA in the composi-
tions of 0.25, 0.5, 1 and 2 wt % Li2CO3. Ball-milled
BHA- Li2CO3 powders were then pressed to cylindrical
compacts (British Standard, No. 7253) which were then
sintered in an open atmospheric furnace for 4 hours (Na-
bertherm HT 16/17, Lilienthal, Germany) at 900, 1000,
1100, 1200, and 1300°C. D ensity, Vick ers microhard ness,
and compression strength were measured for all samples.
2.2. Mechanical Testing
Compression strength tests on sintered samples were
carried out using a universal testing machine (DVT.e,
Devotrans Inc. Istanbul, Turkey; speed 2 mm/min). The
densities were determined using the Archimedes method
and microhardness (TUKON, Wilson Instruments, Group
of Instron, Darmstadt, Germany; 200 gr. load) was
measured three times and a mean value was taken.
2.3. Artificial Neural Networks
In a feed forward ANN system, the input data is proc-
essed from input to output. The neurons are classified in
three layers called input layer, hidden layer and output
layer. The feed forward ANN structure is illustrated in
Figure 1. The network input(s) and output(s) of the hid-
den and output layers are denoted as Xi, Yj, Ok. The
ANN process is governed by the following equations:
00,1,2 i
Nh
netXW N
jiij
i,


 (1)

j
Ynet
j
f (2)
00,1,2,
Mo
jjk
j
netY W jM



k (3)

0,1,2,
kk
Ofnet kP (4)
Figure 1. Feedforward neural network architecture.
here W is weight of connections and f(.) is neuron active-
tion function that can be selected as hard limiter, linear or
nonlinear. Because of th e continuity, the activation func-
tion is frequently selected as a sigmoidal function. It can
be defined as:


-net
fnet e

 (5)
The parameter λ describes the rate of the activation
function. In supervised training, ANN applications re-
quire a training data set to learn the relationsh ip between
inputs and outputs. The training set should consist of
sufficient number of samples that define a process. Oth-
erwise, insufficient learning can limit the performance of
the ANN approach.
3. Results and Discussion
Sintering of BHA-Li2CO3 composites results in the for-
mation of BHA-Li2O [6]. Table 1 illustrates experimen-
tal values for density, Vickers microhardness and com-
pression strength for defined sintering temperatures and
Li2CO3 compositions. For 0.25% and 0.50%, it is seen
that density values increased with increments to tem-
perature and Li2CO3 content. The highest values com-
pression strength were obtained at 1300°C; ~73.75 MPa
for 0.25% and ~75.23 MPa for 0.50%. This can be at-
tributed to the occurrence of a glass phase, which has a
wetting effect through the grain boundaries. The lowest
compression strength values were observed at 1200°C;
~2.25 MPa for 1% and ~5.92 MPa for 2% Li2CO3 addi-
tions. Evidently, high amounts of Li2CO3 addition cause
a decrease in compression strength. For compacts con-
taining 0.25% Li2CO3, the Vickers hardness values in-
creased with increasing sintering temperature. However,
the same outcome was not observed for samples with 0.5,
1 and 2% quantities because of increased porosity.
Experimental values obtained for BHA composites
consist of five sets of sintering temperatures (ranging
from 900-1300oC) and within these, properties such as
density, microhardness and compression strength will

Predicting the Mechanical Properties of BHA-Li2O Composites Using Artificial Neural Networks 100
Table 1. Mechanical Properties of BHA-Li2O c o mposites with different Li2CO3 content.
T[oC] 0.25% 0.50% 1% 2%
d

HV d

HV d

HV d

HV
900 2.104 36.02 41.9 2.816 56.01 114.892.33 38.43 59.53 1.742 15.02 NA
1000 2.192 38.64 53.12 2.826 57.82 88.24 2.109 17.13 65.08 1.661 12.43 NA
1100 2.215 43.11 62.46 2.852 59.32 102.452.24888.67 49 1.713 10.57 NA
1200 2.587 47.81 120.26 2.869 61.84 108.842.36525.27 51.68 1.651 8.92 NA
1300 2.844 70.75 175.83 2.758 72.23 149.092.06722.75 49.01 1.696 9.83 NA
*NA = not applicable
vary. Three temperature sets (900, 1000, 1100oC) were
used to train the ANN model. A Back-propagation algo-
rithm was used in the training process with a sigmoid
type of activation function. Here, ANN models have two
inputs and one output (Figure 2). The inputs consisted of
sintering temperature and mixture rate (Figure 3) and the
output is a predicted value (for a selected property) based
on the training obtained from data provided. Training
experiments were performed individually for the predic-
tion of each property, which is now set as an outcome
(density, microhardness and compression strength).
Part of the training phase is to examine the different
number of neuron networks (or connections). The output
value for a property (density, microhardness and com-
pression strength) can be enhanced by increasing the
number of neurons or data entry during the training stage.
After the training step, the ANN was used to predict me
chanical properties of the last two groups (1200 and
1300ºC) which were withheld during the training process.
A comparison between the predicted (using ANN) and
obtained results are illustrated in Table 2.
ANN based predictions of the compression strength at
1200 and 130 0°C have an average maximum error rate of
~17%. Prediction error rates of density and microhard-
ness obtained were ~7% and ~2.5% respectively. When
comparing the total prediction errors of the two tested
data sets (1200°C and 1300°C) , it is observed that data
obtained for 1300°C is less accurate when compared to
those obtained for 1200ºC, and this can be seen by com-
paring the predicting and corresponding values for den-
sity and compression strength. However, the reverse is
Figure 2. Defined ANN model prediction.
(a)
(b)
(c)
Figure 3. ANN predicted values for a) density, b) compres-
sion strength, c) microhardness.
true for microhardness predictions.
The results of ANN pred iction for density, microhard-
ness and compression strength are also illustrated in
Figures 3 a, b and c; where the trend between the pre-
dicted and obtained values is clearly evident, although
there are differences in the actual values.
4. Conclusion
According to the findings, the ANN approach can supply
reasonable predicted values fo r limited mixture rates and
temperature. Because of non-linear changes to density,
compression strength and microhardness, the ANN ap-

Predicting the Mechanical Properties of BHA-Li2O Composites Using Artificial Neural Networks 101
Table 2. Obtained values compared with ANN predictions (and relative errors).
1200oC 1300oC
Mixture rate Measurement ANN
Prediction
Relative
error % Measurement ANN Prediction Relative error %
0.25 2.587 2.27 13.96 2.844 2.32 22.59
0.5 2.869 2.84 1.02 2.758 2.84 2.89
1 2.365 2.15 10.00 2.067 2.11 2.04
Density(g/cm3)
2 1.651 1.67 1.14 1.696 1.66 2.17
0.25 47.81 49.73 3.86 70.75 58.85 20.22
0.5 61.84 59.24 4.39 71.23 57.69 23.47
1 5.27 3.98 32.41 2.75 3.62 24.03
Compression Strength
(MPa)
2 8.92 11.09 19.57 9.83 12.65 22.29
0.25 120.26 117.63 2.24 175.83 177.05 0.69
0.5 108.84 115.61 5.86 149.09 148.47 0.42
microhardness hardness (HV)
1 51.68 49.654 4.09 49.01 49.84 1.67
proach may not provide very accurate numerical predic-
tions for values outside of the mixture rate and tempera-
ture limits provided during training. Prediction perform-
ance can be expanded by increasing number of experi-
mental data used in th e ANN learning process. Th e ANN
approach can provide quick predictions for mechanical
properties of BHA-Li2O composites, reducing time and
cost on sample preparation and analysis.
5. Acknowledgements
This study was supported with FEN-D-040310-0046
numbered project from Marmara University (BAPKO –
Scientific Research Projects Committee).
REFERENCES
[1] P. V. Landuyt, F. Li, J. P. Keusterma ns, J. M. Strey dio, F.
Delannay and E. Munting, “The Influence of High Sin-
tering Temperatures on the Mechanical Properties of
Hydroxylapatite,” Journal of Materials Science: Materials
in Medicine, Vol.6, No.1, 1995, pp. 8-13.
doi:10.1007/BF00121239
[2] M. A. Fanovich, M. S. Castro Porto and J. M. Lopez,
“Improvement Of The Microstructure and Microhardness
of Hydroxyapatite Ceramics by Addition Of Lithium,”
Materials Letters, Vol.33, No.5-6, 1998, pp. 269-272.
doi:10.1016/S0167-577X(97)00115-8
[3] P. Valerio, F. N. Oktar, G. Goller, S. Ozyegin, A. P. M.
Shainberg, A.M. Goes and M.F. Leite, “Biocompatibility
evaluation of lithium-hydroxyapatite composite,” Key
Engineer Materials, Vol. 309-311, 2006, pp. 1121-1124.
doi:10.4028/www.scientific.net/KEM.309-311.1121
[4] M. A. Fanovich, M. S. Castro and J. M. P. López,
“Analysis of the microstructural evolution in hydroxy-
apatite ceramics by electrical characterisation,” Ceramics
International, Vol.25, No.6, 1999, pp. 517-522.
doi:10.1016/S0272-8842(97)00087-4
[5] S. Haykin, “Neural Networks,” Prentice Hall, New Jersey,
1999.
[6] L. S. Ozyegin, F. N. Oktar, G. Goller, E. S. Kayali and T.
Yazici, “Plasma sprayed bovine hydroxyapatite coatings,”
Materials Letters, Vol.58, No.21, 2004, pp. 2605-2609.
doi:10.1016/j.matlet.2004.03.033