J Electroceram (2009) 22:291–296

DOI 10.1007/s10832-007-9394-x

Optimizing design of the microstructure of sol–gel derived

BaTiO3 ceramics by artificial neural networks
Huiqing Fan & Laijun Liu

Received: 27 March 2007 / Accepted: 11 December 2007 / Published online: 23 December 2007
# Springer Science + Business Media, LLC 2007

Abstract Modeling and application of artificial neural piezoelectric bodies, as well as for thin film electronic
network (ANN) technique to the formulation design of materials. It is known that the resultant microstructure of
BaTiO3-based ceramics were carried out. Based on the the ceramics is one of the key factors for the properties of
homogenous experimental design, the results of BaTiO3- electronic ceramics. The relative permittivity of BaTiO3
based ceramics were analyzed using a three-layer back ceramics can be increased to about 3500 if the grain size is
propagation (BP) network model. Then the influence of controlled at ∼1 μm, so a fine-grained structure is desirable
sintering temperatures, holding time, donor additives to enhance the dielectric performances [1]. Therefore, how
(La2O3, MnO2, Ce2O3) and sintering aids (Al2O3–SiO2– to obtain BaTiO3 ceramics with homogeneous and smaller
TiO2 (AST)) on the average grain size (da), the degree of particle size becomes more and more important. However,
grain uniformity given by the ratio of the maximal grain the grain growth is always difficult to control in the
size to the average grain size (dmax/da), and the relative sintering procedure. Many problems may be encountered
density (Dr) of doped BaTiO3 ceramics system was in processing high-purity stoichiometric and nanocrystalline
investigated. The optimized results and experiment data BaTiO3 powders. Meanwhile, the sintering behavior of
were expressed and analyzed by intuitive graphics. Based BaTiO3 is very sensitive to the sintering temperature, the
on input data and output data, the sintering behavior of sintering time, the sintering-aids (e.g. AST), the original
BaTiO3 nano-powder was explained well. Furthermore, the powders and the donor additives, and that there may be
fine and uniform microstructure of sol–gel derived BaTiO3 complex reciprocal actions between them [2–6]. So it is
ceramics with da ≤3 μm, dmax/da ≤1.20, and Dr ≥98% was usually difficult to explain their functions in the system.
obtained. The ever-increasing needs to discover ceramic formulation
with a good microstructure requires considerable precise
Keywords Microstructure . Grain size . Sintering . BaTiO3 mathematical models in general. It would be intractable for
us to develop a reliable modeling for the ceramics without
enough knowledge before doing the job by using conven-
1 Introduction tional method.
Due to its remarkable information processing character-
Owing to its high dielectric constant at room temperature, istics, the artificial neural network (ANN) has been
stoichiometric submicronised BaTiO3 powders are being extensively utilized in solving a diverse area of science
widely used for the preparation of dense ferroelectric and and engineering problems. As reviewed by Zupan and
Gasteiger [7], ANN seems to be of the most interest, in
H. Fan (*) : L. Liu particular, when a rigid theoretical basis or mathematical
State Key Laboratory of Solidification Processing, relationship is not available in advance. Its application can
School of Materials Science and Engineering, improve, shorten, or bring new insight into old ways of
Northwestern Polytechnical University,
Xi’an 710072, China
handling experimental data. In the past decade, there have
e-mail: hqfan3@163.com been numerous successful applications of ANN techniques
e-mail: hqfan@nwpu.edu.cn in diversified areas of science and engineering including
292 J Electroceram (2009) 22:291–296

pattern recognition and classification, voice and image and H2O were employed as solvents. The preparation of the
processing, prediction [8], digital communications [9], and precursor was a simple process. Firstly, Ba(OH)2·8H2O was
nonlinear system identification and control [10]. However, dissolved in the CH3OCH2CH2OH. Then, Ti(C4H9O)4 was
only a few attempts have been conducted ANN on diluted in CH3OH in a mole ratio of Ti(C4H9O)4:CH3OH=
electronic ceramics. 1:24. At last, Ba(OH)2·8H2O solution was added into Ti
In this work, we used ANN technique to model the (C4H9O)4 solution according to the stoichiometry of Ba:Ti=
sintering behavior of a sol–gel derived BaTiO3-based 1:1. We added well calcined and ground 4Al2O3–9SiO2–
ceramics. With the aid of ANN model, the optimized 3TiO2 powders and then dropped small amount of distilled
design of BaTiO3 ceramics is predicted. The fine micro- water after adding nitrate of La3+:Mn4+:Ce4+ =1:1:2 under
structure of BaTiO3-based ceramics is obtained by using the vigorous stirring. The precursor was kept in air at the room
optimal parameters. temperature for 4 h, yielding a transparent gel. The dried
gel, obtained by baking the fresh gel at 120 °C, was
calcined at 950 °C for 2 h and milled with carnelian ball for
2 Experimental 4 h to form doped BaTiO3 nano-powder in avenge size of
40 nm (Scherrer formula). The calcined powder was die-
2.1 Homogenous experimental design pressed to green disks in diameter of 10 mm and in
thickness of 3 mm at 120 MPa after the addition of a
Doped BaTiO3 nano-powders were prepared by a sol–gel binder. The binder was burned out at 500 °C for 1 h. The
method as shown in Fig. 1. Chemical grade barium green disks were put in a covered Al2O3 crucible and were
hydroxide (Ba(OH)2·8H2O, 99%, Beijing Chemical Facto- sintered with an electric furnace at 1150∼1300 °C for 30∼
ry, China), tetra-butyl titanate (Ti(C4H9O)4, 98%, Beijing 180 min in an air atmosphere. The apparent density of the
Chemical Factory, China), additives lanthanum oxide sintered samples was measured by the Archimedes method
(La2O3, 99.5%, Tianjing Chemical Factory, China), man- of immersion in water. The microstructure of the sintered
ganese oxide (MnO2, 99.5%, Beijing Chemical Factory, samples was examined on the polished and thermally
China) and cerium oxide (CeO2, 99.8%, Tianjing Chemical etched surfaces by using a scanning electron microscope
Factory, China), and sintering-aids (AST) [11] aluminum (SEM; JSM 6460, JEOL Inc., Tokyo, Japan). The grain
oxide (Al2O3, 99.8%, Luoyang Chemical Factory, China), sizes were determined by the line-intercept method. For
silicon oxide (SiO2, 99.5%, Tianjing Chemical Factory, knowledge acquisition the homogenous experimental de-
China) and titanium oxide (TiO2, 98%, Beijing Chemical sign offers schemes for scientific experimental design
Factory, China) were used as the starting materials. 2- within the range of interest. Sixteen formulations gained
Methoxyethanol (CH3OCH2CH2OH), methanol (CH3OH) from homogenous experimental design [12] act as the
training data set.

2.2 Knowledge acquisition

Because BaTiO3 content is an independent variable among

the reagents and the size of original particle is uniform at
about 40 nm, the average grain size (da), the degree of grain
uniformity (dmax/da), and the relative density (Dr) of doped
BaTiO3 ceramics are dependent on mainly sintering temper-
atures, sintering times, donor additives (La2O3, MnO2,
Ce2O3) and sintering aids (Al2O3–SiO2–TiO2, AST).
Hence, the ANN model is composed of four neurons in
the input layer and three neurons in the output layer. To
determine an appropriate number of hidden neurons is an
important factor determining the network's performance.
From trials using different numbers of hidden neurons 5–
10, the minimum in root mean square errors was obtained
for six hidden neurons. The architecture of the ANN model
is shown in Fig. 2. The initial connection weights are set to
be random between 0.3 and −0.3. Within the neural
Fig. 1 Flow chart of BaTiO3 nano-powders synthesized by a sol–gel network the learning rate η, the momentum μ and the
process convergence error are set 0.15, 0.075 and 0.01, respectively.
J Electroceram (2009) 22:291–296 293

that the calculated Dr is in good accordance with the

observed Dr. These results also indicate that ANN can help
us to accurately predict the influence of the sintering
process on ceramic microstructures. It should be noted that
since decreasing the convergence error can increase the
accuracy of the ANN model, however, too small conver-
gence error leads to an over-fitting of the original data
points. On the other hand, if the input data are not included
in the range of training set, the prediction errors for model
might increase dramatically, i.e. the observed values depart
from the calculated values. Therefore, for a more convinc-
ingly inspection of the accuracy of the ANN model, all
input data of simulated samples are not only within the
Fig. 2 Architecture of a BP network for the optimizing design of original additives and process levels, but also different from
BaTiO3 ceramics the training samples.
By using a conventional method, it is necessary to do
some extra work in order to study the relationship between
The neuron function is Sigmoid. The maximum iteration is a processing parameter and microstructure. For example, if
50000. Table 1 gives the partial training data set. The input we want to study the influence of sintering time on the
data and the output data were normalized to give values relative density of the ceramics, we have to fix the sintering
between around 0.01 and 1.00. temperature, the amount of donor additives and AST then
measure the density of the samples dwelling different
sintering times. The original experimental data can not be
3 Results and discussion utilized. However, we can express the framework of the
model in an intuitive way without any extra experiment by
ANN BP model was trained with the experiment data. On extracting the information of the ANN model registered.
the basis of ANN model, we made use of the method of n- Here the graphical analysis capability of the ANN model is
dimensional complex optimization with constraints to also illustrated. The relationship between sintering temper-
calculate the registered BaTiO3 formulation. The finished atures and grain sizes, as well as the degree of grain
ANN model is delineated by the connection weights uniformity is shown in Fig. 3(b). The amount of donor
between the input and hidden layers and the connection additives is 0.4 mol%, AST is 0.1 mol% and the sintering
weights between the hidden and output layers. We saved time is 120 min. The average grains size changes linearly
these connection weights, then for a more convincing with the sintering temperature increasing. It known that the
inspection of the accuracy of the ANN model, some new BaTiO3–AST eutectic temperature is 1240 °C [13]. The
samples within the original experimental levels but different motion of the grain boundary may be controlled mainly by
from the training samples were also tested. The relationship the pores when the sintering temperature is below 1240 °C.
between observed relative density and calculated relative When the sintering temperature is over 1240 °C, although
density for the latter samples is shown in Fig. 3(a), where the liquid-phase BaTiO3–AST appears, it does not enwrap
the two coordinates of a point represent the observed Dr grains but mainly exists at the intersection of three grains
and the calculated Dr of a sample respectively. It is found due to only a little AST sintering-aid. Therefore, the liquid

Table 1 Experimental results of BT ceramic samples.

Samplea 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

Dr (%)b 86.2 95.9 97.6 87.1 86.9 93.7 83.0 97.5 88.2 95.7 94.6 98.3 94.1 91.3 89.2 96.1
da (μm)c 2.0 2.9 3.1 2.6 2.3 3.2 1.9 3.1 2.4 2.5 4.0 3.4 4.5 2.4 2.2 2.2
dmax/dad 1.16 1.73 2.47 2.83 1.35 1.72 1.07 2.61 1.01 1.24 3.53 3.81 1.51 3.01 1.04 1.12
a
The amount of donor additives of La2O3–MnO2–Ce2O3 ranges between 0 and 0.6 mol%, the sintering aids of 4Al2O3–9SiO2–3TiO2 (AST)
ranges between 0 and 1 mol%, the sintering time ranges between 30 and 180 min, the sintering temperature ranges between 1150 and 1300 °C.
b
The relative density, measured by the Archimedes method.
c
The average grain size, determined by the line-intercept method on SEM micrographs.
d
The degree of grain uniformity, given by the ratio of the maximal grain size to the average grain size.
294 J Electroceram (2009) 22:291–296

Fig. 3 (a) Predicted correlation between the calculated and observed Dr by ANN model, (b) relationship between the sintering temperature and
da, dmax/da, (c) relationship between the sintering time and relative density (Dr), (d) relationship between the AST and dmax/da

phase does not enhance the densification greatly. The relative density of 98.3% is obtained at 1300 °C for
sintering temperature dependence of grain size distribution 180 min. The sintering time dependence of the relative
width (dmax/da) is nonlinear when the sintering temperature density displays power law behavior when the dwelling
is below 1240 °C, but linear at higher temperature. The time is below 90 min; however, it nearly changed linearly
sintering behavior is against that of powder synthesized by with the increase of sintering time at 1280 °C when the
a traditional method. It is suggested that many particles are sintering time is above 90 min. Since some particles are
aggregated. As the sintering temperature increases, some aggregated, the densification takes place firstly in the
grains become larger at the expense of other grains. By region of hard aggregation. Therefore, the relative density
observing the process of grain growth, we also find that increases quickly as the sintering time increases at 1280 °C
Ostwald ripening happens, and some irregular grains with a short dwelling time. As mentioned above, the liquid
appear. In any case, the presence of liquid-phase blocks phase diffusion and grain boundary diffusion generally
the motion of grain boundaries and restricts the rate of grain have an important influence on densification of bulk. It is
growth so that the grain size does not change acutely at well known that the liquid eutectic temperature for the
higher sintering temperature. TiO2–BaTiO3 system is 1320 °C for powder synthesized by
The relationship between the sintering time and the a traditional method [14, 15]. The liquid eutectic temper-
relative density is shown in Fig. 3(c). The sample, the ature for the TiO2–BaTiO3 system may be about or less
amount of donor additives is 0.56 mol%, AST is 0 mol%, is than 1300 °C because the original powders are nanometer
sintered at 1280 °C, and the other, the amount of donor powders. When the green disks are sintered at 1300 °C,
additives is 0.56 mol%, AST is 0.1 mol%, is sintered at much liquid phase may appear. The liquid phase distributes
1300 °C. It can be seen that bulk shrinkage carries through well on the grain boundaries and enwraps grains so as to
quickly when the sintering time increases from 30 to enhance the densification of BaTiO3 ceramics. The sinter-
90 min at 1300 °C, but the relative density increases slowly ing intermediate stage finishes when the sintering time is
from 90 to 150 min. Finally, the BaTiO3 ceramics in a above 90 min. But some isolated pores are left behind, and
J Electroceram (2009) 22:291–296 295

then the densification of the bulk carries through slowly by ceramics with da =3 μm, dmax/da =1.20, and Dr =98% was
lattice diffusion and grain boundaries diffusion. obtained using this inputting values. The micrographs of
The relationship between AST and dmax/da is shown in BaTiO3 ceramics are shown in Fig. 4. Figure 4(a) and (b)
Fig.3(d). The AST addition acts not only as the liquid-phase show the microstructure of BaTiO3 ceramics sintered before
formers to enhance the sintering performance but also as optimized process. Figure 4(c) and (d) show the micro-
useful contributors to the electron traps at the grain structure of BaTiO3 ceramics sintered after optimized
boundary [16]. In this work, the AST plays the role of a process. The grains are more uniform than that of samples
sintering aid to enhance the densification and prevent the sintered before optimized procedure. It is confirmed that
abnormal grain growth by forming liquid phase. It can be ANN can accurately predict the trend and rule in the change
seen that the dependence of the average grain size on the of parameters with the concentration of the additives and
AST is linear below 0.4 mol%. Meanwhile, the grains the sintering processing.
become fine as the AST increase. But the influence is
limited when the AST is above 0.4 mol%. The relationship
between AST and the degree of grain uniformity is linear 4 Conclusions
below 0.3 mol%, and the effect of the AST on grain
uniformity is pretty strong, but it is ignorable in a wild ANN model is a relatively new computational tools and
scope of AST above 0.3 mol%. methodology associated with nonlinear regression tech-
Based on above experimental and analyzing results, we nique. Application of a three-layer BP network modeling of
inputted again many groups of combinatorial data into input the homogenous experimental results, complex influence
layer of trained ANN, until the perfect data were outputted. rules of the sintering process and additives on the
Finally, a group of perfect values with da =2.8 μm, dmax/da = microstructure of BaTiO3-based ceramics are shown clearly
1.16 and Dr =99.7% was output while the amount of donor through several groups of experiments. The observed
additives is 0.6 mol%, AST is 0.4 mol%, the sintering relative density is in accordance with the predicted relative
temperature is 1300 °C, and the sintering time is 105 min density. Both the average grain size and the degree of
was inputted. A fine microstructure of doped BaTiO3 uniformity were controlled effectively by adjusting sinter-

Fig. 4 SEM micrographs

of doped BaTiO3 ceramics.
(a) Sample 4 and (b) sample 10
sintered before optimized
procedure, (c) sample 4 and (d)
sample 10 sintered after
optimized procedure
296 J Electroceram (2009) 22:291–296

ing process and additives. Based on input data and output 4. K.S. Mazdiyasni, L.M. Brown, J. Am. Ceram. Soc 54, 539–543
(1971)
data, the sintering behavior of BaTiO3 nano-powder is
5. D.A. Payne, H.U. Anderson, J. Am. Ceram Soc 50, 491–503
explained well, and then BaTiO3-based ceramics with (1967)
above 98% theoretical density and uniform grains in a size 6. A.C. Caballero, J.C. Fernandez, C. Moure, P. Duran, J. Eur.
of ∼3 μm were successfully sintered by using the group of Ceram. Soc 17, 513–523 (1997)
7. J. Zupan, J. Gasteiger, Neural Networks for Chemists: An
optimal parameters.
Introduction (VCH, Weinheim, Germany, 1993)
8. S. Haykin, Neural Networks (Maxwell Macmillan, Ontario,
Acknowledgements This work has been supported by the National Canada, 1994)
Nature Science Foundation (50672075) and the Xi'an S&T Research 9. J.C. Patra, R.N. Pal, R. Baliarsingh, G. Panda, IEEE Trans. Syst.
Foundation (GG05015, GG06023, XA-AM-200608), the EYTP and Man Cybern., Part B, Cybern 29, 262–266 (1999)
NCET and RFDP (20050699011) Program of MOE, Science Creative 10. J.C. Patra, R.N. Pal, B.N. Chatterji, G. Panda, IEEE Trans. Syst.
Foundation and Doctorate Foundation (CX200704) of NPU, Aeronautic Man Cybern., Part B, Cybern 29, 254–256 (1999)
Science Foundation (2006ZF53068) of China. The author, Dr. Huiqing 11. Y. Matsuo, M. Fujimura, H. Sasaki, K. Nagase, S. Hayakawa, Am.
Fan, gratefully acknowledges the support of K. C. Wong Education Ceram. Soc. Bull 47, 292–297 (1968)
Foundation, Hong Kong. 12. K.T. Fang, Uniform Design and Uniform Design Tables (Science
Press, Beijing, China, 1994)
13. M. Wegmann, F. Clemens, A. Hendry, T. Graule, Ceram. Int 32,
147–156 (2005)
References
14. K.W. Kirby, B.A. Wechsler, J. Am. Ceram. Soc 74(8), 1841–1847
(1991)
1. H. Kniepkamp, W. Heywang, Z. Angew. Phys 6, 385–390 (1954) 15. Y.M. Chiang, D.P. Birnie, W.D. Kingery, Physical Ceramics
2. A. Yamaji, Y. Enomoto, K. Kinoshita, T. Mrakami, J. Am. Ceram. (Wiley, New York, USA, 1997)
Soc 60, 97–101 (1977) 16. C.H. Lai, C.T. Weng, T.Y. Tseng, Mater. Chem. Phys 40, 168–72
3. M. Kahn, J. Am. Ceram. Soc 54, 452–454 (1971) (1995)