Journal of Membrane Science 251 (2005) 137–144

Modelling of the adsorption of bovine serum albumin on porous

polyethylene membrane by back-propagation artificial
neural network
Rong-Qiang Fu, Tong-Wen Xu∗ , Zhong-Xiao Pan
Department of Chemistry, University of Science and Technology of China, Hefei 230026, PR China

Received 15 August 2004; received in revised form 4 November 2004; accepted 9 November 2004
Available online 15 December 2004

Abstract
Herein, a back-propagation artificial neural network (BP-ANN), a fit and predictive tool and suitable for bridging the inputs and outputs
of a non-linear problem, is used to model the adsorption of bovine serum albumin (BSA) on porous polyethylene (PE) membrane. Based
on the adsorption data from FTIR-mapping, the parameters of the neural network with a hidden layer are determined by a trial-and-error
method. There is a good agreement between the predicted results by BP-ANN and the experimental data, and the interpolative predictions by
BP-ANN are more precise than those by convectional diffusion equation. Though BP-ANN cannot provide detail information concerning the
mechanism like a conventional diffusion equation (which can permit an evaluation of diffusion coefficient or mass transfer coefficient), it is
a powerful predictive tool as well as a useful compensation to conventional diffusion equation when dealing with the problems of which the
mechanism is not entirely understood or very complex.
© 2004 Elsevier B.V. All rights reserved.

Keywords: Adsorption; Modelling; Membrane; Bovine serum albumin; Artificial neural network

1. Introduction measurements, an improved physical model for adsorption of

Protein adsorption to solid surface, particularly to the
membrane filters and the subsequent fouling of devices, con-
tinues to be a major obstacle to the application of the mem-
brane technologies. To understand the adsorption mechanism
and thus enhance the membrane’s performance, a substantial
body of literatures has been developed on the subject of the
interaction of serum proteins with synthetic membrane sur-
faces, such as porous polyethylene (PE), PTFE, CA, PVDF,
PS, etc. [1–5]. It is now generally accepted that adsorption of
globular proteins on most hydrophobic surfaces, such as poly-
meric microfiltration membranes PE, PS and PVDF is nearly
irreversible in a way of side-on at low surface coverage and
partly irreversible in a way of end-on at high surface coverage
[3,5–9]. Recently, based on the adsorptional and desorptional
∗ Corresponding author. Tel.: +86 551 3601587; fax: +86 551 3601592.
E-mail address: twxu@ustc.edu.cn (T.-W. Xu).
bovine serum albumin (BSA) on porous polyethylene mem-
brane is proposed by the authors [10], and based on FTIR-
mapping measurements the diffusion coefficient and interfa-
cial mass transfer coefficient for BSA on PE membrane are
determined based on convectional diffusion equation [11].
It is well documented that the adsorption of serum protein
on polymeric membrane is affected by many variables, such
as pH value, salt type and concentration, buffer type and con-
centration, protein type and concentration, adsorption time,
etc. Generally, it is hard to correlate the adsorption data with
the variables by simple mathematical derivations. However,
an artificial neural network (ANN) provides a convenient tool
[12,13]. A neural network involves algorithms under which
information is accumulated in programmed objects that are
capable of learning through much iteration using simulated or
real data. This form of artificial intelligence can handle prob-
lems, in which relationships are less known or very complex,
especially in the field of membranes, e.g., ANNs have been

0376-7388/$ – see front matter © 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.memsci.2004.11.007
138 R.-Q. Fu et al. / Journal of Membrane Science 251 (2005) 137–144
used to model the evolution of membrane fouling during mi-
crofiltration [14] and ultrafiltration [15] processes under con-
stant and variable conditions, and they have also been used to
predict permeate flux and rejection for reverse osmosis (RO)
[15,16] and ultrafiltration [17,18] processes.
In addition, ANNs have been used to compress IR spec-
tra [19], recognize UV spectra [20] and create a quantita-
tive structure–property relationship of hydrocarbons [21,22]
in our group. To continue the research in ANN and com-
bine ANN with membrane technology, in this paper ANN
is used to model the adsorption dynamics (adsorption data
at different membrane depth and different adsorption time)
of serum protein on polymeric membrane. Even though the
adsortion process can be characterized with conventional dif-
fusion equation, it will be found that modelling the adsorption
with ANN is a useful compensation to conventional diffusion
equation. In fact, finding the variety in modelling the data has
been the one of main tasks for chemical engineers.
2. Artificial neural network
An artificial neural network is a computer model derived
from a simplified concept of the brain [12]. It is a parallel
distributed processing system composed of nodes (neurons)
and connections (weights), and is based on the principle that a
highly interconnected system of simple processing elements
can learn complex interrelationships between independent
and dependent variables. Neurons are single processing el-
ements, which connected to neurons in the next layer and
therefore forming different types of ANNs. The most popular
ANN is the back-propagation ANN (BP-ANN) [23], which
belongs to the class of supervised learning techniques be-
cause the method consists in comparing the response of the
output units to the desired responses via an iterative process
in which an error term is calculated and used to readjust the
weights in the network in order to obtain network responses
close to the desired responses. Its most important feature is
the ability to learn from examples and to generalize, since
the learnt information is stored across the network weights.
Thus, it is not surprising that BP-ANN has gained momen-
tum in numerous industrial and scientific areas. It is estimated
that more than 90% of the neural network applications in use
today employ BP-ANN or its some variants [24]. Though
BP-ANN have been used widely as a modelling method for
many years and suitable neural networks software packages
are available, even now BP-ANN has not spread sufficiently
widely outside the specialists in mathematical methods. Thus,
here the principles of BP-ANN are stated as the follows.
As shown in Fig. 1, a BP-ANN topologically presents three
types of layers: one input layer, one (or more) hidden layer(s)
and one output layer. The universal approximation theorem
suggests that a network with a single hidden layer with a suf-
ficiently large number of neurons can in principle relate any
given set of inputs to a set of outputs to an arbitrary degree of
accuracy [25,26]. As a result, the neural network used in this
Fig. 1. The topological structure of a typical BP-ANN.
paper is equipped with a single hidden layer. The action of
the neurons in hidden and output layers is summing weighted
inputs and produces an output signal through a transfer func-
tion (activation function). Although the transfer function can
take any form and may be linear or non-linear, the most com-
monly used, which is also used in this present study, is the
sigmoid function whose general form is given as:
1
f (x) = (1)
1 + e−x
And for the neurons in input layer the input is directed to the
output. As is the usual practice [13], an additional neuron,
called the bias neuron, which do not take any input and they
emit a constant output value of 1 to the neurons in the next
layer, is added to the input and hidden layers in order to meet
the universal approximation property of the network (relating
inputs to outputs).
Before starting the training (learning) process, all the
weights associated with the connections between the neurons
must be initialized to small random numbers (e.g., herein
between −0.1 and 0.1), which ensures that the network is
not saturated by large values of the weight, and prevents
some training pathologies. And in order to obtain conver-
gence within a reasonable number of cycles, the input and
output data should be usually normalized. The most com-
monly employed is the linear normalization method as fol-
lows:
xi − xmin
xi = a +b (2)
xmax − xmin
where xi is the normalized value, xi the original data, xmax
and xmin , respectively, the maximum and minimum values,
and a and b are the positive constants allowing to fix the
limits of the interval for the scaled values. In this paper, a and
b are set as 0.8 and 0.1, thus the scaled data ranging between
0.1 and 0.9. During the training phase, each individual data
pair of the training set is presented to the network, which
generates a calculated output. An error is calculated from
 R.-Q. Fu et al. / Journal of Membrane Science 251 (2005) 137–144
the calculated and target outputs, and weights are adjusted
by back-propagating the error from the output layer to the
input layer. The adjustment of the weights can be made
immediately after the error is detected when inputting each
individual data pair (immediate adjustment), or be made after
accumulating the individual errors for all data pairs (deferred
adjustment). Herein, the deferred adjustment is used, and
the adjustment between the (n)th and the (n + 1)th iterative
is [13]:
1

wji (n + 1) = η δk,j outk,i + αwji (n) (3)
K (both from Nicolet), respectively. The resolution of IR is
k
where wji is the weight adjustment between the neuron j
in the considering layer and the neuron i in the former layer,
outk,i the calculated output of the neuron i for data pair k,
δk,j the δ term of the neuron j for data pair k calculated from
the calculated and target outputs, η the learning rate, α the
momentum factor for preventing oscillations and K is the
number of data pairs.
Thus, the unknown parameters η, α and neuron number in
the hidden layer should be determined in advance. Herein, the
trial-and-error method and the split-sample validation proce-
dure are used, that is, the sample data pairs are divided into
training set and validating set, and the parameters and itera-
tive number must be taken to make the error of the validating
set minimize. Herein, the error is expressed as root mean
squares (rms) error.
3. Experimental descriptions
3.1. Materials
Bovine serum albumin (Lot 113H0247, water con-
tent = 1.7%, molecular weight = 69,000, isoelectric point
(IEP) = 4.9) is obtained from Sigma Chemical Co. and used
without further treatments. Polyethylene membrane (H 2200)
is provided from Asahi Kasei Co. with a thickness 200 ␮m,
porosity 69%, maximum pore diameter 0.28 ␮m and specific
area about 18 m2 /g evaluated by BET adsorption. Before ex-
periments, the membrane is cut into square shape with side
length about 50 mm and soaked in ethanol for a few min-
utes and washed with RO water and then equilibrated them
with the buffer solution complying to the experiments. Other
chemical reagents are used as received.
3.2. Adsorption and FTIR-mapping
The adsorption experiments are conducted by immers-
ing the PE sample in a BSA solution with concentration
2000 ppm and pH 5.05 (close to the isoelectric point), which
is placed in an incubate box (M-230F Taitec) at 25 ◦ C with a
vibrating speed about 50 rpm. After a given time, the mem-
branes are taken out of the solution, freeze-dried with liquid
nitrogen and cut into lines with a special knife in the direc-
139
tion perpendicular to membrane surface. The breadth of the
line sample is equal to the thickness of original membrane
sample (200 ␮m) and the thickness of line sample (distance
between the knife) is about 50 ␮m. The IR measurements
are conducted with a Nic-Plan microscope FTIR instrument
(with a Magna-IR System 560 and a Nicolet 0.25 MCT-A
IR detector, Nicolet Instrument Co., Madison, WI). The IR
system is equipped with a micro X–Y stage (INA/KLANEW,
Tokyo, Japan) for two-dimension mapping of the line sam-
ple. The IR data acquisition and microscope mapping are
controlled by a PC using the software of Omnic and Atlus
4 cm−1 and 100 IR scans are made to obtain an IR spec-
trum at each point of the sample. All the spectra are col-
lected in the range of 1000–2000 cm−1 . Each line mapping
contained about 11–13 points across the whole membrane
thickness, and included the measurement of a background IR
spectrum at a point of KBr sample holder near one edge of the
sample.
3.3. Determination of adsorption data
Herein, the PE membrane has no infrared absorbance
in the Amide I bands (nearby 1650 cm−1 ) and Amide II
bands (nearby 1550 cm−1 ), and the adsorption concentration
of BSA in membrane phase can be quantitatively analyzed
by Amide II area because the absorbance intensity in the
area has been reported to be proportional to the amount of
protein adsorbed [7,27] and is believed not to be very sen-
sitive to the conformation of the protein [5]. Our prelimi-
nary experimental results also show that Amide II area is
linearly regressed with adsorption concentration with a re-
gression factor about 0.992 [10]. As an example, the FTIR-
mapping at the adsorption time 50 min is shown in Fig.
2. The peak around 1462 cm−1 corresponds to –CH2 – of
PE membranes, while the peaks 1650 and 1550 cm−1 to
the Amide I and II bands of adsorbed protein molecules,
respectively. It is clearly observed that, both the intensity
of Amide I and Amide II peaks symmetrically decrease
with the distance from the membrane surface outside and
with a minimum value at the central point of the mem-
brane. Thus, the concentration of the adsorption protein is
not only time-related but also depth-related. However, it is
hard to read Amide II peak area from FTIR-mapping. To do
this, the mapping is further processed and over-stacked in
Fig. 2. The over-stack spectrums for Amide II area determination at different
membrane depth.
140 R.-Q. Fu et al. / Journal of Membrane Science 251 (2005) 137–144
Fig. 3. FTIR-mapping at adsorption time 50 min.
one plane by Omnic software provided by Nicolet Instru-
ment Co. Fig. 3 shows the over-stacked spectrum for ad-
sorption sample of 50 min (corresponding to Fig. 2). Then,
the data (Amide II area) is easy to be read from the over-
stacked spectrum by Omnic software for different membrane
depth.
Our pre-experiments show that the time needed for satura-
tion adsorption is about 180 min [10] and great changes occur
at the initial time period. So, adsorption time of the samples
for FTIR-mapping is not more than 50 min, and total 36 data
pairs (the adsorption time ranges from 10 to 50 min with step
10 min) are collected.
3.4. Pretreatment of adsorption data
Since the adsorption of BSA on membrane proceeds from
the two sides symmetrically, we only need to consider the
adsorption in membrane from the surface to the middle point.
For easy processing, herein the adsorption concentration C
and membrane depth x (the middle position is set as 0) are
nondimensionalized by [11]
C − Cg
θ= (4)
C0 − C g
and
x
r= (5)
d the error in the validating set begins to increase). As shown
where θ and r are especially the nondimensionalized con-
centration and depth, Cg the saturated concentration, C0 the
initial concentration and d is the half of membrane thickness.
Note that θ is defined so that it is unity in the beginning and
falls down with the time. Herein, the adsorption isotherm at
pH = 5.05 is fitted to a Langmuir-type equation to give the
saturated amount about 3.88 mg/m2 , C0 is equal to zero and
d is equal to 100 ␮m.
4. Results and discussions
4.1. The optimization of neural network parameters
Here, the adsorption of BSA on porous polyethylene mem-
brane is modelled by a back-propagation artificial neural net-
work with two inputs (nondimensionalized depth and adsorp-
tion time) and one output (nondimensionalized concentra-
tion). A trial-and-error method is used in the optimization
of neural network topological structure. The data pairs are
divided into training set and validating set, i.e., 12 represen-
tative data pairs are randomly selected as validating set and
the remaining 24 data pairs are used as the training set. It
must be concerned that both validating set and training set
contain some data at all adsorption times. According to some
preliminary calculations, the change of momentum factor α
has a little influence on the validating error when it is set as
from 0.80 to 1.00, but the best validating and training results
are obtained as α is set as 0.90. However, the learning rate η
has apparent impacts on the validating error, so several tests
are carried out. At last η is determined at 0.70 to train neural
network. Once the learning rate η and momentum factor α
are determined, the optimal hidden node number and iterative
number could be easily found. Fig. 4 is the dependencies of
validating error and the optimal iterative number on hidden
node number. Herein, the optimal iterative number means
the minimal validating error (i.e., training is stopped when
in Fig. 5 (as an example), the training error persistently drops
during the training, while the validating error firstly drops
then increases with iterative number. Thus, the optimal it-
erative number can be determined from the validating error
curve. It can be found from Fig. 4 that the validating error
tends to be a flat and there is little change when hidden node
number is larger than 5, however the optimal iterative num-
ber increases sharply with hidden node number. Thus, as a
 R.-Q. Fu et al. / Journal of Membrane Science 251 (2005) 137–144
Fig. 4. The dependencies of validating error and the optimal iterative number
on hidden node number.
compromise of calculating rate and calculating precision, the
hidden node number is set as 6 and iterative number is set as
17,280 (the corresponding optimal value). Fig. 6 shows the
correlation plot of calculated concentrations by BP-ANN at
the final parameters versus experimental concentrations. The
validating set has the correlation coefficient R = 0.9888 and
rms = 0.0082, and the training set has the correlation coeffi-
cient R = 0.9779 and rms = 0.0098. Thus, the prediction by
BP-ANN with these parameters meets with the application
need.
4.2. Modelling by BP-ANN and comparing with
convectional diffusion equation
Figs. 7–10 show the experimental data and the calculated
data by BP-ANN at the above determined parameters when
adsorption time is 10, 20, 30 and 40 min, respectively. It can
be found that there is a good agreement between calculated
data and experimental data, which proves that BP-ANN is a
powerful fit and predictive tool. However, the calculated data
Fig. 5. Dropping curves of validating error and training error during iteration
(hidden node number is 6).
141
Fig. 6. The correlation plot of calculated concentration by BP-ANN vs.
experimental concentration. (
) N = 12, R = 0.9888, rms = 0.0082; ()
N = 24, R = 0.9779, rms = 0.0103.
at adsorption time of 50 min have some deviations from the
experimental data (as shown in Fig. 11). As mentioned above,
50 min is the maximum adsorption time. The above results
prove that ANN is good in interpolation but their extrapola-
tion capabilities (including prediction in limiting inputs) are
limited [18]. Thus, it is not surprising that there are some
deviations between calculated data and experimental data at
50 min.
Moreover, it can be observed that the shape of calculated
curves by BP-ANN varies with the adsorption time. For ex-
ample, at the adsorption time of 20 min, the curve is convex
and the tangent decreases with the depth (it should be noticed
that the tangent is minus and the absolute value of the tangent
increases with the depth), which implies that the concentra-
tion gradient (dC/dr) at membrane surface is much larger
than that at membrane middle point. However, at the adsorp-
tion time of 50 min, the curve is nearly linear, which implies
Fig. 7. The calculated data by BP-ANN and the calculated data by convec-
tional diffusion equation when adsorption time is 10 min.
142 R.-Q. Fu et al. / Journal of Membrane Science 251 (2005) 137–144
Fig. 8. The calculated data by BP-ANN and the calculated data by convec-
tional diffusion equation when adsorption time is 20 min.
that there is little difference in the concentration gradient be-
tween membrane surface and its middle point. Furthermore,
the concentration gradient at membrane surface is larger at
low adsorption time, however the concentration gradient at
membrane middle point is larger at high adsorption time.
In order to evaluate the stability of the network, the cal-
culated data by BP-ANN at adsorption time of 20, 25 and
30 min are shown in Fig. 12. As mentioned above, the time
values of experimental data range from 10 to 50 min with step
10 min, that is, 25 min is not included in the time values of
training data. However, it can be found from Fig. 12 that the
predict curve at 25 min runs between the curve at 20 min and
the curve at 30 min, which proves that the network is stable
and predict data is credible.
In our previous paper [11], the experimental data are also
fitted based on the convectional diffusion equation
∂θ(r, t) D ∂2 θ(r, t)
= 2 (6)
∂t d ∂r 2
Fig. 9. The calculated data by BP-ANN and the calculated data by convec-
tional diffusion equation when adsorption time is 30 min.
Fig. 10. The calculated data by BP-ANN and the calculated data by convec-
tional diffusion equation when adsorption time is 40 min.
and the calculated data by this diffusion equation are also
shown in Figs. 7–11. It is obvious that BP-ANN is a better
tool for interpolative prediction (not prediction in limiting
inputs). As well known, there are very complex interactions
between serum proteins and synthetic membranes [1–5]. The
convectional diffusion equation based on Fick’s second law
cannot fully cover the behaviors of serum proteins in mem-
branes. Thus, it is not surprised that there are some deviations
between the calculated data by diffusion equation and experi-
mental data. However, ANN, a highly interconnected system
of many processing elements, can handle very complex or
less known interrelationships, and is irrelevant to the mecha-
nism. Thus, ANN is a good tool for bridging the experimental
data with the experimental conditions when dealing with the
adsorption of serum proteins on synthetic membranes.
However, as above mentioned, ANNs have poor extrap-
olation capability, and the predictive results by BP-ANN at
adsorption time 50 min have more deviations than those by
convectional diffusion equation. At the same time, ANN can-
Fig. 11. The calculated data by BP-ANN and the calculated data by convec-
tional diffusion equation when adsorption time is 50 min.
 R.-Q. Fu et al. / Journal of Membrane Science 251 (2005) 137–144
Fig. 12. The calculated data by BP-ANN when adsorption time t are 20, 25
and 30 min.
not provide the descriptions of diffusion details just like the
convectional diffusion equation which can permit an eval-
uation of diffusion coefficient or mass transfer coefficient.
Thus, ANN cannot fully replace conventional diffusion equa-
tion when dealing with the adsorption of serum proteins on
synthetic membranes. Maybe, it is more proper to considere
that ANN is a useful compensation to conventional diffusion
equation.
It should also be noted that our method presented here is
very general, though it is exemplified with the adsorption of
BSA in PE membrane. It can be used to all the proteins ad-
sorption on/diffusion in alkene polymeric membranes, such
as PP, PE, PVDF, PTFE, etc., which have no absorbance at
AII region or other polymeric membranes, such as PS, CA,
SPPO, which have absorbance at AII region but can be sub-
tracted from the adsorption peaks [2,7].
5. Conclusions
BP-ANN is successfully employed to model the adsorp-
tion of BSA in membrane. The interpolative predictions by
BP-ANN in this paper are very close to the experimental data,
and are better than those obtained by convectional diffusion
equation. Though it cannot provide detail information about
diffusion like a conventional diffusion equation, BP-ANN is
a powerful fit and predictive tool as well as a useful com-
pensation to conventional diffusion equation when dealing
with the problems in which the mechanism is not entirely
understood or very complex.
Acknowledgements
This research is supported by National Basic Research
Program of China (973 program, no. 2003CB615700),
the Natural Science Foundation of China (nos. 20106015,
143
20376079) and the Special Foundation for Doctoral Disci-
pline of Ministry of Education of China (no. 20030358061).
References
[1] W.R. Bowen, Q. Gan, Properties of microfiltration membrane: adsorp-
tion of BSA at polyvinylidene fluoride membranes, J. Colloid Interface
Sci. 144 (1991) 254–262.
[2] P.M. Bummer, An FTIR study of the structure of human serum albumin
adsorbed to polysulfone, Int. J. Pharma. 132 (1996) 143–151.
[3] W.J. Dillman, I.F. Miller, On the adsorption of serum proteins on poly-
mer membrane surfaces, J. Colloid Interface Sci. 44 (1973) 221–241.
[4] K. Kenjirou, Study on the correlation of protein with porous polymeric
surface, Master Dissertation, Tokyo University, 1991.
[5] K. Kontturi, M. Vuoristo, Adsorption of globular proteins on polymeric
microfiltration membranes, Desalination 104 (1996) 99–105.
[6] C.E. Giacomelli, M.G.E.G. Bremer, W. Norde, ATR-FTIR study of
IgG adsorbed on different silica surfaces, J. Colloid Interface Sci. 220
(1999) 13–23.
[7] D. Lu, K. Park, Effect of surface hydrophobicity on the conformational
changes of adsorbed fibrinogen, J. Colloid Interface Sci. 144 (1991)
271–281.
[8] S. Staunton, H. Quiquampoix, Adsorption and conformation of bovine
serum album on montmorillonite: modification of the balance between
hydrophobic and electrostatic interactions by protein methylation and
pH variation, J. Colloid Interface Sci. 166 (1994) 89–94.
[9] H. Urano, S. Fukuzaki, Conformation of adsorbed bovine serum albu-
min governing its desorption behaviour at alimina–water interfaces, J.
Biosci. Bioeng. 90 (2000) 105–111.
[10] T.W. Xu, R.Q. Fu, L.F. Yan, A new insight into the adsorption of bovine
serum albumin onto porous polyethylene membrane by zeta potential
measurements, FTIR analysis and AFM observations, J. Colloid Inter-
face Sci. 262 (2003) 342–350.
[11] T.W. Xu, R.Q. Fu, A simultaneous determination of effective diffusion
coefficient and interfacial mass transfer coefficient of bovine serum
albumin (BSA) adsorption into porous polyethylene membrane by mi-
croscope FTIR mapping study, Chem. Eng. Sci. 59 (2004) 4569–4574.
[12] J. Devillers, Neural Network in QSAR and Drug Design, Academic
Press, San Diego, CA, 1996.
[13] J. Zupan, J. Gasteiger, Neural Networks in Chemistry and Drug Design,
second ed., Wiley–VCH Verlag Gmbh, Weinheim, Germany, 1999.
[14] M. Dornier, M. Decloux, G. Trystram, et al., Dynamics modeling of
crossflow microfiltration using neural networks, J. Membr. Sci. 98
(1995) 263–273.
[15] H. Niemi, A. Bulsari, S. Palosaari, Simulation of membrane separation
by neural networks, J. Membr. Sci. 102 (1995) 185–191.
[16] M. Jafar, A. Zilouchian, Adaptive receptive fields for radial basis func-
tions, Desalination 135 (2001) 83–91.
[17] W.R. Bowen, M.G. Jones, H.N.S. Yousef, Prediction of the rate of cross-
flow membrane ultrafiltration of colloid—a neural network approach,
Chem. Eng. Sci. 53 (1998) 3793–3802.
[18] W.R. Bowen, M.G. Jones, H.N.S. Yousef, Dynamic ultrafiltration of
proteins—a neural network apporch, J. Membr. Sci. 146 (1998) 225–
235.
[19] W. Liu, J.P. Li, J.H. Xiong, Z.X. Pan, M.S. Zhang, The compression of
IR spectra by using wavelet neural network, Chin. Sci. Bull. 42 (1997)
822–825.
[20] C.S. Yin, Y. Shen, S.S. Liu, Q.S. Yin, W.M. Guo, Z.X. Pan, Simulta-
neous quantitative UV spectrophotometric determination of multicom-
ponents of amino acids using linear neural network, Comp. Chem. 25
(2001) 239–243.
[21] C.S. Yin, W.M. Guo, T. Lin, S.S. Liu, R.Q. Fu, Z.X. Pan, L.S. Wang,
Application of wavelet neural network to the prediction of gas chro-
matographic retention indices of alkanes, J. Chin. Chem. Soc. 48 (2001)
739–749.
144 R.-Q. Fu et al. / Journal of Membrane Science 251 (2005) 137–144
[22] C.S. Yin, W.M. Guo, W. Liu, W. Zhao, Z.X. Pan, Estimation and pre-
diction of gas chromatography retention indices of hydrocarbons in
straight-run gasoline by using artificial neural network and structural
coding method, Chem. Res. Chin. Univ. 17 (2001) 31–40.
[23] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning representations
of by back-propagation errors, Nature 323 (1986) 533–536.
[24] J. Zupan, J. Gasteiger, Neural networks: a new method for solving
chemical problems or just a passing phase?, Anal. Chim. Acta 248
(1991) 1–30.
[25] K.I. Funahashi, On the approximate realization of continuous
mapping by neural networks, Neural Netw. 2 (1989) 183–
192.
[26] Y. Takahashi, Generalization and approximating capabilities of
multilayer networks, Neural Comp. 5 (1993) 132–139 (Taka-
hashi).
[27] W.K. Surewicz, H.H. Mantsch, New insight into protein secondary
structure from resolution-enhanced infrared spectra, Biochim. Bio-
phys. Acta 952 (1988) 115–130.