ARTIFICIAL NEURAL NETWORK

MODELING OF ATMOSPHERIC
CORROSION IN THE MICAT PROJECT
Salvador Pintos Nestor V. Queipo
Applied Computing Institute
College of Engineering
University of Zulia
Maracaibo, Zulia 4005
Venezuela
Oladis Troconis de Rincon Alvaro Rincon
Center for Corrosion Studies
College of Engineering
University of Zulia
Maracaibo, Zulia 4005
Venezuela
Manuel Morcillo
National Center for Metallurgic Research
Gregorio del Amo Avenue, Madrid 28040
Spain
May 30, 1998

1
 ABSTRACT
This paper presents an Arti cial Neural Network(ANN)-based solution methodology for
modeling atmospheric corrosion processes from observed experimental values, and an
ANN model developed using the cited methodology for the prediction of the corrosion
rate of carbon steel in the context of the Iberoamerican Corrosion Map (MICAT) Project,
which includes seventy two (72) test sites in fourteen (14) countries throughout Iberoa-
merica. The ANN model exhibited superior performance in terms of goodness of t (sum
of square errors) and residual distributions when compared against a classical regression
model also developed in the context of this study, and is expected to provide reasonable
corrosion rates for a variety of climatological and pollution conditions. Furthermore, the
proposed methodology holds promise to be an e ective and ecient tool for the cons-
truction of analytical models associated with corrosion processes of other metals in the
context of the MICAT project, and, in general, in the modeling of corrosion phenomena
from experimental data.

KEYWORDS: A. steel, B. modeling studies, C. atmospheric corrosion

1 INTRODUCTION
The Iberoamerican Atmospheric Corrosion Map Project (MICAT-Mapa Iberoamericano
de Corrosion Atmosferica) was established as a resolution of the Board of Directors of the V
Centennial Science and Technology Program for Development (CYTED) during a meeting
in La Habana, Cuba in 1988. The project, originally conceived at the Iberoamerican
Corrosion Congress in Maracaibo, Venezuela in 1986, includes seventy two (72) test sites
2
exposed to the atmosphere in fourteen (14) countries throughout Iberoamerica, namely:
Argentina, Brasil, Chile, Colombia, Costa Rica, Cuba, Ecuador, Portugal, Peru, Mexico,
Venezuela, Panama, Espa~na, y Uruguay. Figure 1, displays the network of test sites.
Along the lines of the ISOCORRAG, and ICP/UNECE projects, the MICAT project has
three main objectives: i) construct the corrosion map for Iberoamerica, ii) provide a better
understanding of the atmospheric corrosion phenomena, and iii) identify mathematical
models that could predict the corrosion rate of metals in the atmosphere as a function
of metereological and pollution variables, for Iberoamerica. Regarding objective iii), a
signi cant amount of work has been done in order to establish analytical expressions or
models for the behavior of metals in the atmosphere as a function of easily determined
variables (instead of long lasting experiments). The task at hand has shown to be dicult,
mainly because of the complexity (non-linearities) associated with the physicochemical
processes responsible for the atmospheric corrosion phenomena.
Most of the predictive models used to date, are regression models that t the data available
such that their mean square error is minimized. However, these models have been shown
to be e ective in very restrictive areas and to be limited to capture the nonlinear nature
of the corrosion process. Hence, the continuous search for mathematical models that could
predict the atmospheric corrosion rate for rather general climatic and contamination level
conditions1 3 .
Arti cial Neural Network (ANN) Modeling, in particular the one based on a class of
ANN called multilayer perceptron, has recently emerged as a promising area in corrosion
research4 , because of the potential of the ANN to predict any complex process with arbi-

3
trary precision provided its architecture and a set of parameters are properly set5 . This
study presents the development of an ANN based model for the prediction of the corrosion
rate of carbon steel (Fe) in the context of the MICAT project (hence, considering a broad
spectrum of climatological and pollution conditions), and evaluate its performance (mean
square error and residual distribution) against a classical regression model.
The remainder of the paper is structured as follows. Section 2, provides a detailed descrip-
tion of the problem under consideration, while Section 3, presents the adopted solution
methodology including an introduction to ANN and the ANN modeling process. The
analysis and discussion of the results obtained using the developed ANN model, and its
relative performance when compared against a regression model is the subject of Section
4. Conclusions and recommendations of areas worthy of future research e orts in the
modeling of corrosion processes using ANN are also included.

2 PROBLEM DEFINITION
The problem of interest here corresponds to the identi cation of a model for the prediction
of atmospheric corrosion of a given metal in terms of metereochemical variables in the
context of the MICAT project. Speci cally, the model is expected to provide a prediction
of the corrosion rate of carbon steel (Fe in m=year) as a function of the time of wetness
(TOW ), chloride deposition rate (Cl in mg Cl =m2 :d), sulfate deposition rate (SO2 in
mg Cl =m2:d), , relative humidity (RH in %), precipitation (P in mm), and temperature
(T in oC ). Furthermore, the model must be constructed considering the actual values of
the corrosion rate of the carbon steel and the corresponding metereochemical variables at
test stations throughout Iberoamerica (Tables 1 through 5).
4
Mathematically, the problem under investigation is one of function estimation and may
be stated as follows, given a set of n observations of metereochemical variables and the
corresponding observed corrosion rate values, denoted by ~xi , f^i , with i = 1::n:
nd f (~x)
such that Pni=1(f (x~i) f^i)2 is minimized

wherein, f (~x) is a function (model) providing the corrosion rate of carbon steel (Fe), and
~x is a vector of metereochemical variables (TOW , Cl , SO2, RH , P , T ).

3 SOLUTION METHODOLOGY
The proposed solution methodology is based on the use of Arti cial Neural Networks
(ANN); speci cally the so called multilayer perceptron model. This model is an universal
approximator that is able to predict any sampled process with arbitrary precision provided
its architecture and a set of parameters are properly established. This section, after
presenting an introduction to ANN, discusses the ANN-based modeling process carried
out in this study, and the construction, for comparison purposes, of a linear regression
model along the lines of those reported in the literature for the MICAT project.
3.1 Introduction to Arti cial Neural Networks
These biologically inspired models have found applications in many di erent areas such as,
classi cation of data, pattern recognition, clustering analysis, and function approximation,
among others. This section provides an introduction to an ANN model called multilayer
(three layer) perceptron used in function approximation tasks such as the one faced in this

5
study. A more detailed discussion on other ANN models and applications may be found
in the works by Rumelhart and McClelland6.
The introduction to the ANN model under consideration (Figure 3), is given in terms of:
basic de nitions (arti cial neurons, links, layers, weights, threshold, activation fuction),
architecture, and learning process.

Basic elements. The basic processing elements are the so called arti cial neurons.
With reference to Figure 2, an arti cial neuron basically evaluates an activation function
taking as argument a weighted sum of its inputs. A commonly used activation function
are the logistic functions, an example of a logistic function is given by Equation 1; note
that the output is then limited to values in the interval [0; 1] and in general, the output
data must be subject to a normalization process.

g(x) = 1 +1e x (1)

These neurons are arranged in layers (Figure 3), and there are three di erent layers,
namely, input, hidden, and output layers. There are some "special" neurons, the ones at
the input layer and the bias neurons, that have as output their corresponding inputs.
The bias neurons are introduced to facilitate the consideration of the threshold value
associated with each neuron. Furthermore, the neurons are connected among them
through channels called links, with the strength of the connections speci ed as real
numbers called weights. The links usually connect neurons in consecutive layers from left
to right.

6
The output of the ANN model can be written as shown in Equation 2.

Op(Ip) = G[W  (G[V  Ip])] (2)

wherein, Ip denotes the input values, V and W the weight matrices between the input
layer and hidden layer, and the hidden layer and output layer, respectively. The symbol
G, represents the application of the logistic function to each of the elements of its
argument.

Architecture. It can be shown that a three layer perceptron (the one used in this
investigation) is, for function approximation purposes, as capable as ANN with higher
number of hidden layers, and can estimate a function with arbitrary precision provided
the number of neurons in the hidden layers and the weights are properly set5 . As a
result, our attention will be restricted to three-layer perceptrons.
The number of neurons in the input and output layers, are established by the number of
input, and output variable(s) of the function(s) to be estimated. The number of neurons
in the hidden layer can be settled through di erent criteria; for example, limiting the
number of parameters of the model (weights) to be a fraction of the total number of data
points available to the learning process (to be discussed next).

Learning. This process makes reference to the identi cation of a set of weights that, for
a given architecture, minimize the sum of the square of the model errors (Equation 3);
the errors represent the di erence between the ANN model output values and the
observed function values.
X
n
min Error(V; W ) = kOpi f^ik2 (3)
i=1

7
The nonlinear nature of the aforementioned error function, makes it nonconvex and
prone to have local optima. The most used ANN learning is a gradient based
optimization procedure called Backpropagation6, which essentially, modi es the weights
(from a starting set of weights) by moving in the direction contrary to the error function
gradient, thus ensuring a lower error value (Equation 3). The cited algorithm, as any
other gradient-based optimization procedure, is limited by the fact that it is sensitive to
the set of initial weights, and it may get trapped in local optima.

3.2 ANN modeling

The neural network modeling process used in this study, may be described in four
stages: preprocessing of the original data set (identi cation of outliers), partitioning of
the preprocessed data set into learning, validation and test sets, ANN model architecture
setting, learning and testing, and implementation.

Preprocessing of the original data set. Under the assumption of the continuous
nature of the corrosion process, two test sites with similar metereochemical values should
have similar corrosion rates. In order to detect data values violating this premise, a
cluster analysis was conducted on the metereochemical variable values of the original
data set. In general, observations within a cluster, with corrosion rates atypical for a
given cluster, were eliminated. In addition, since the input variables were so di erent in
ranges, in order to facilitate the learning process (to be explained shortly) the input
variables were scaled using the following transformation: x^ = x   , with x, and x^
denoting, an input variable and scaled input variable, respectively; and the symbols 

8
and , representing the mean and standard deviation among the observed values for the
metereochemical variables (Table 6).

Partitioning the data set. The resulting data from the preprocessing process
conducted in the previous step, is divided into three disjoint data sets: learning,
validation and test data sets. The learning data set is the one used to train (specify the
weights) of the ANN. The validation data set is the one used in conjunction with the
learning data set to identify when to stop the learning process so that the resulting ANN,
exhibits good generalization properties. The test data set allows the assesment of the
prediction capabilities of the ANN model. The ANN is evaluated using, as performance
criteria, mean square error and residual distributions over learning and test data sets.
The data set was sorted by country and test site and the learning, validation, and test
data were constructed using an strati ed sampling procedure. The selected sampling
procedure aims to data sets that include observations of all countries and test sites. The
number of observations per data set were assigned to be one hundred and ve (105),
sixteen (16) and nine (9), for the learning, validation and test sets, respectively.

ANN model architecture setting. Since the nodes for the input and output layer are
set by the number of metereochemical variables (six) and the number of variables to be
predicted (one), only the number of hidden neurons needs to be establish before the
ANN model architecture is completed. In this work the number of hidden neurons was
established so that the number of parameters (weights) during the learning process is a
fraction (approximately 40%) of the number of observations available in the learning data
set. Hence, avoiding the possibility of the ANN model over tting the data.
9
Learning and Testing
The learning process was conducted using as optimization procedure, the standard
Backpropagation algorithm, with weight updates each time the complete learning data
set was considered. The initial set of weights, were obtained using Montecarlo
optimization, hence, reducing the possibility of getting trapped in local optima. For a
given set of initial weights, the learning process was stopped when the mean square error
in the validation data set started to grow for higher number of iterations in the
optimization procedure. The "best" set of weights was selected as the one with the lowest
possible mean square error and relative good performance on the validation data test.
After the conclusion of the ANN learning process, the resulting model is evaluated in
terms of its mean square error and residual distributions over the test data set, and its
relative performance when compared against a regression model , with the same structure
of those reported in the literature7, created using the learning and validation data sets.

Implementation
The solution methodology is implemented using a combination of both commercial and
academic software. The Statistical Analysis System (SAS)8 was used, for the cluster
analysis (PROC CLUSTER) in the preprocessing of the original data set, for the
strati ed sampling conducted in the partitioning of the data set, and in the construction
of the regression model (PROC REG) in the testing phase of the ANN modeling process.
The Arti cial Neural Network models were created with the assistance of the Stuttgart
Neural Network Simulator (SNNS)9 .

10
 4 RESULTS AND DISCUSSION
This section presents an Arti cial Neural Network (ANN) model for the prediction of
the corrosion rate of carbon steel (Fe) as a function of relevant metereochemical
variables (TOW , Cl , SO2, RH , P , and T ) in the context of the MICAT project, and
discusses the performance exhibited by the ANN model in terms of goodness of t (mean
square error), and residual distributions for training and testing data sets. In addition,
the corresponding results for a linear regression model, also developed in the context of
this study, are included for comparison purposes.
The ANN model obtained following the methodology described in Section III is
illustrated in Figure 4, and its output can be expressed in matrix form as speci ed by
Equation 2, with V [H  (I + 1)] and W [O  (H + 1)] given by the following weight
matrices.
V= 0 1
B 0:0745 0:9768 0:9852 0:4833 0:7987 1:0064 3:0598 C
B
B 3:3410 2:4293 1:8049 2:3564 3:1194 4:6681 3:2227 C
C
B
B 0:4417 1:9811 1:5018 0:7849 0:0004 2:4218 2:0833 C
C
B
@ 0:4767 1:9384 1:2480 0:1868 1:1675 2:6001 1:8650 C
A
2:4737 1:1378 0:1231 1:6081 0:6292 0:1397 3:2680
W=
 
0:3466 3:1603 0:0620 1:4238 2:9312 3:1031

The ANN model captures well-known nonlinear interactions between the corrosion rate
and some of the most signi cant metereochemical variables such as, TOW , SO2 and
Cl , as it is illustrated in Figures 5 through 8. In general, each gure depicts the
corrosion behavior as a function of one or more of the metereochemical variables; the rest

11
of the variables are considered either constant (at their mean values) or, if they are
highly correlated with the variable(s) under consideration, at the values speci ed by
simple linear models. The climatological variables, are highly correlated as shown in
Table 7, which displays the correlation value and the probability of a null correlation for
the di erent pairs of climatological variables.
Figures 5 and 6 shows the behavior of the corrosion rate with respect to chloride
deposition rate (Cl ), and sulfate deposition rate (SO2), respectively. In both cases, as
expected, the corrosion rate increases with higher values of Cl and SO2, respectively,
with the former exhibiting a higher nonlinear interaction with the variable of interest
than the latter.
Furthermore, Figure 7, depicts the ANN model response surface as a function of Cl and
time of wetness TOW , having a constant SO2 (mean value), and simple linear regression
models for the rest of the variables; speci cally, T = 12:45 + 15:59  TOW ,
RH = 55:73 + 35:88  TOW , and P = 241:00 + 1813:00  TOW . Similarly, the ANN
model response surface as a function of TOW and SO2 is shown in Figure 8, with Cl
assumed constant (mean value) and the aforementioned linear regression models
establishing the values for the variables T , RH , and P . In both instances the gures
show signi cant nonlinear interactions among the considered variables.
The ANN model shows superior tting (learning and validation data sets) capabilities
exhibiting a mean square error (MSE) and (R2) coecient equal to 140.45 and 0.90,
respectively, which represents a 65% (27%) lower (higher) value than those provided by
a quadratic regression model constructed as part of this study (similar to the one

12
reported by Morcillo7, with a 95% degree of signi cance level regarding the linear
coecients. The linear regression model obtained is shown in Equation 4.

Fe = b0 + Cl (b1 + b2  P + b3  RH ) + b4  TOW  SO2 (4)

with, b0 = 6:8124, b1 = 1:6907, b2 = 0:0004, b3 = 0:0242, and b4 = 2:2817.

In addition, the residuals of the ANN model are smaller and more symmetrically
distributed when compared with its regression model counterpart (Figure 9).
The prediction performance of the ANN model over a set of data not used during the
training process was also superior to the one exhibited by the regression analysis model.
The MSE over the validation data provided by the ANN model was approximately four
(4) times smaller that the one provided by the linear regression model (220 versus 1082).
Within this data set, the minimum and maximum error for the ANN model (Regression
Model) were -11.9 and 32.9 (-9.5 and 88.7), respectively; note again the more symmetric
distribution of the errors for the ANN model.

5 CONCLUSIONS
1. This paper presented an Arti cial Neural Network-based methodology for the
modeling of atmospheric corrosion. The methodology involves the following steps:
preprocessing and partitioning of the experimental data set, ANN model
architecture setting, construction and testing of the ANN model, and
implementation.

2. Using the proposed solution methodology, an ANN model was constructed and
evaluated using experimental data associated with the corrosion rate of carbon
13
 steel (Fe) as a function of the time of wetness (TOW ), chloride deposition rate
(Cl ), sulfate deposition rate (SO2), relative humidity (RH ), precipitation (P ),
and temperature (T ), in the context of the MICAT project. The ANN model: i)
reproduced some well-known nonlinear interactions among the variables of interest,
and, ii) provided excelent results regarding goodness of t (MSE) and residual
distributions on training and testing data sets.

3. Speci cally, regarding item ii) the application of the ANN model on the training
and validation (testing) data sets resulted in a MSE equal to 140 (220), which, was
63% (80%) lower than those obtained using a linear regression model developed
from the same data. In addition, the residual distributions reported by the ANN
model when using training and testing data sets, were signi cantly smaller and
more symmetric than the corresponding to the aforementioned regression model.

4. The ANN model developed as part of the study holds promise to be useful in the
prediction of carbon steel corrosion rates under a wide spectrum of climatological
and pollution conditions in Iberoamerica, and, the proposed solution methodology
is expected to be an e ective and ecient tool for the construction of models
associated with atmospheric corrosion processes for other metals in the context of
the MICAT project, and, in general, in the modeling of corrosion processes from
experimental data.

14
Acknowledgments
The authors would like to thank the Consejo de Desarrollo Cient co y Humanstico, of the
University of Zulia, and the Consejo Nacional de Investigaciones Cient cas, Venezuela,
for prividing nancial support to conduct this work. Also special thanks are given to the
MICAT Iberoamerican Group for sharing their experimental data.

References
[1] Legault, R., and Pearson, P., "Atmospheric corrosion in marine environments,"
Corrosion, Vol. 34, No. 12, p. 433, 1978.
[2] Sanchez, A., Perez, C., Merino, P., and Espada, L., "Modelo teorico para la
corrosion atmosferica del Cinc," 4th Iberoamerican Congress of Corrosion and
Protection,, Vol. I, p. 61, Argentina, 1992.
[3] Feliu, S., Morcillo, M., and Feliu, J., "The prediction of atmospheric corrosion from
metereological and pollution parameters," Corrosion Science, Vol. 34, No. 3, p. 403,
1993.
[4] Helliwell, I., Turega, M., Cottis, R., "Neural networks for corrosion data reduction,"
em Corrosion 96, p. 739, Denver, USA, 1996.
[5] Hecht-Nielsen, R., Neurocomputing, Addison Wesley, 1989.
[6] Rumelhart, D., and McClelland, J., Parallel Distribution Processing: Explorations
in the Microstructure of Cognition (Volumes 1 & 2), The MIT Press, 1986.
[7] Morcillo, M., "Atmospheric corrosion in Iberoamerica: The MICAT Project",
Atmospheric Corrosion, ASTM STP 1239, W.W. Kirk and Herbert H. Lawson,
Eds., American Society for Testing and Materials, Philadelphia, 1994.
[8] SAS Institute Inc., "SAS Quality Control Reference Manual", Ver. 6.0, First
edition, Cary: NC, USA, 1990.
[9] University of Stuttgart, "Stuttgart Neural Network Simulator (SNNS) User
Manual," Ver. 4.1, Report No. 6/95, available through anonymous ftp at
ftp.informatik.uni-stuttgart.de in the directory /pub/SNNS, Germany, 1995.

15
 LIST OF FIGURES
1. Network of test sites in the MICAT project
2. Illustration of the information processing ability of an arti cial neuron
3. Schematic of the three-layer perceptron ANN model
4. Arti cial Neural Network model
5. Corrosion rate of carbon steel (Fe) versus chloride deposition rate Cl
6. Corrosion rate of carbon steel (Fe) vs. sulfate deposition rate SO2
7. Corrosion rate of carbon steel (Fe) as a function of chloride deposition rate Cl
and time of wetness (TOW )
8. Corrosion rate of carbon steel (Fe) as a function of sulfate deposition rate SO2 and
time of wetness (TOW )
9. Residual distributions for the ANN and regression analysis models

16
17
18
19
20
21
22
23
24
 LIST OF TABLES
1. Learning data set (Part I) with the pre xes A,B,CO,CR,CU, and CH, denoting
test sites in Argentina, Brazil, Colombia, Costa Rica, Cuba and Chile, respectively
2. Learning data set (Part II) with the pre xes EC, E, M, and PA, denoting test sites
in Ecuador, Spain, Mexico, and Panama, respectively
3. Learning data set (Part III) with the pre xes PA, PE, PO, U, and V, denoting test
sites in Panama, Peru, Portugal, Uruguay and Venezuela, respectively
4. Validation data set
5. Test data set
6. Fundamental statistical information associated with the data set used for the
construction of the ANN model
7. Correlation values (top) and probability of a null correlation (bottom) among the
climatological variables

25
Test site T RH T OW P SO2 Cl Fe
A011 14.1 79.2 .681 817 5.0 30.0 24.9
A012 13.9 78.8 .707 805 5.0 40.2 54.8
A013 14.5 80.3 .736 1226 5.0 70.2 66.0
A022 17.1 71.5 .482 983 5.0 10.0 16.1
A023 17.0 73.6 .494 1420 5.0 9.0 12.4
A031 20.6 75.8 .666 2158 5.0 1.5 5.7
A033 22.1 74.7 .633 1720 5.0 1.5 5.5
A042 20.0 49.3 .097 111 5.0 1.5 4.5
A043 18.3 50.8 .099 93 5.0 1.5 5.6
A051 -2.0 83.8 .307 114 5.0 30.2 37.4
A061 17.0 77.5 .593 1178 6.2 1.5 25.3
B011 21.5 73.5 .482 847 0.8 8.9 8.6
B012 20.9 74.9 .482 1167 1.3 7.4 11.5
B013 21.2 74.6 .482 996 1.7 1.6 13.1
B021 23.8 89.3 .482 1122 23.8 8.6 52.5
B022 22.9 91.1 .482 1471 20.7 6.7 47.2
B031 24.8 77.2 .582 605 9.5 359.8 159.8
B032 24.5 78.6 .582 985 5.2 174.7 194.6
B041 22.7 72.7 .579 960 40.4 4.4 98.7
B042 22.5 71.0 .579 876 59.1 10.7 161.2
B061 19.8 74.9 .648 1409 28.0 1.5 14.6
B062 19.6 76.0 .648 1910 28.0 1.5 23.4
B063 19.4 75.2 .648 1034 28.0 1.5 24.2
B071 20.1 79.1 .598 1305 46.5 13.9 127.1
B073 20.2 83.6 .598 1229 26.6 13.7 73.1
B081 26.1 87.7 .682 2395 5.0 1.5 19.4
B101 20.4 72.0 .442 1440 5.0 1.5 12.9
CO12 27.6 87.0 .966 940 14.2 36.4 30.6
CO13 28.2 87.1 .976 940 8.9 69.5 54.0
CO22 11.4 89.5 .800 1800 5.0 1.5 17.7
CO23 14.2 73.4 .800 1800 5.0 1.5 19.6
CR31 22.9 88.3 .838 3677 8.2 20.5 69.3
CR41 18.9 83.4 .695 1845 4.2 15.5 16.6
CU11 25.2 79.5 .468 1591 37.1 15.8 36.0
CU12 25.4 79.4 .468 1303 36.5 10.9 26.4
CU13 25.3 79.4 .468 1447 19.8 10.9 29.0
CU31 23.9 81.0 .571 1488 16.0 12.3 32.5
CU33 23.9 81.0 .571 1488 15.3 4.2 34.1
CH11 14.2 71.0 .130 355 21.0 4.4 29.3
CH12 14.2 68.3 .372 367 17.9 4.2 44.1
CH41 12.2 82.0 .686 1294 63.9 14.3 167.2
CH51 -2.3 85.0 .264 114 4.5 14.2 24.1

26
Test site T RH T OW P SO2 Cl Fe
EC12 26.9 82.1 .661 635 2.7 1.3 22.6
EC21 12.9 66.0 .409 554 1.0 0.4 7.7
EC31 23.0 81.0 .776 631 3.0 58.4 66.1
E011 12.0 68.8 .384 652 16.2 1.5 19.3
E013 11.1 62.7 .241 334 16.2 1.5 19.7
E021 13.0 62.8 .416 359 3.8 3.9 12.6
E022 13.0 62.8 .416 359 8.9 3.9 10.8
E023 13.0 62.8 .416 359 6.3 3.9 10.6
E031 16.8 65.1 .170 443 18.6 3.8 17.9
E033 15.7 65.5 .156 655 14.8 5.1 18.1
E041 18.1 65.2 .390 554 8.3 1.5 20.3
E042 17.0 62.8 .302 521 5.7 1.5 19.4
E043 17.2 61.9 .316 374 1.9 1.5 21.0
E051 16.3 59.3 .151 416 10.3 1.5 12.3
E053 15.6 57.5 .274 266 2.8 1.5 6.4
E061 16.9 71.5 .369 1828 38.1 11.6 27.0
E062 15.8 68.1 .314 1704 34.7 10.3 29.6
E063 15.1 69.6 .249 1312 45.7 28.2 28.1
E071 14.3 69.0 .339 271 4.1 10.4 14.1
E072 11.0 77.0 .450 510 3.7 10.2 12.8
E073 10.8 74.0 .447 451 3.0 8.7 18.2
E081 8.8 72.0 .100 738 9.1 1.8 3.3
E082 6.9 72.0 .100 624 8.9 1.6 3.6
E083 7.8 72.0 .100 681 9.0 1.7 4.8
M011 16.0 62.0 .265 747 15.6 1.5 15.4
M012 15.2 64.5 .288 747 5.6 1.5 8.6
M013 15.6 63.2 .277 747 17.6 1.5 5.1
M022 21.0 56.0 .212 1724 9.9 1.5 11.4
M023 21.0 56.0 .200 1372 7.1 1.5 13.7
M041 28.0 78.0 .564 834 6.5 34.4 22.6
M042 27.0 77.0 .581 792 15.2 22.9 24.8
M043 28.0 73.3 .570 985 7.2 14.0 18.7
PA11 26.5 69.2 .558 1686 38.5 4.8 25.6
PA12 27.2 70.3 .562 1739 20.0 8.5 23.0
PA13 27.1 74.0 .588 1387 11.0 9.6 28.3
PA21 27.1 75.5 .620 3815 63.0 11.0 125.0
PA22 27.3 76.5 .658 4656 50.5 18.7 93.3
PA23 26.9 75.5 .691 3622 28.7 20.1 113.7
PA32 27.2 68.7 .568 2082 20.3 12.7 14.2
PA33 27.3 71.6 .580 2171 9.1 24.4 20.1
PA41 27.1 65.3 .531 2427 13.7 4.5 23.0

27
Test site T RH T OW P SO2 Cl Fe
PA42 26.9 69.1 .560 2210 6.9 9.8 20.9
PA43 27.4 70.7 .579 2123 3.9 11.8 23.2
PE24 19.1 86.0 .837 14 17.6 39.4 35.2
PE32 18.8 83.5 .761 13 28.9 19.8 34.0
PE41 16.4 37.0 .003 17 5.0 1.5 15.0
PE42 17.2 33.3 .003 89 5.0 1.5 16.5
PE61 25.4 84.0 .500 1523 5.0 1.5 15.7
PE62 25.8 82.9 .500 1656 5.0 1.5 12.9
PO11 15.9 62.1 .379 1129 68.2 161.8 78.9
PO12 16.9 56.4 .315 1129 67.2 82.6 106.4
PO31 17.2 71.3 .351 685 9.1 6.9 27.1
PO32 16.1 66.0 .350 685 7.2 5.1 24.5
PO33 17.4 72.4 .400 685 7.1 6.1 29.9
U011 16.8 73.5 .586 1185 1.0 2.2 8.2
U021 17.1 76.2 .479 1036 16.3 7.6 37.6
U041 -2.3 90.0 .354 473 3.2 30.2 56.0
V011 27.7 78.0 .419 809 6.6 73.1 30.1
V021 28.0 74.5 .462 311 2.7 24.3 20.9
V031 26.6 77.1 .571 362 3.9 32.0 16.6
V032 28.1 74.2 .427 344 2.6 25.7 15.1
V041 27.7 75.0 .513 983 19.0 15.8 23.0
V061 26.5 84.0 .684 807 1.4 12.8 37.3

28
Test site T RH T OW P SO2 Cl Fe
A041 18.0 50.6 .114 35 5.0 1.5 4.6
A052 -3.1 84.0 .276 114 5.0 30.2 35.9
B023 23.0 90.4 .482 1444 24.5 5.2 48.5
B072 23.2 80.0 .598 1152 33.9 14.2 61.3
CO11 27.6 87.0 .971 940 7.7 43.7 15.9
CR11 27.6 80.2 .562 1598 5.3 45.2 61.6
CU32 23.9 81.0 .571 1488 8.9 5.9 28.6
EC11 26.1 71.4 .554 936 4.2 1.5 19.5
E032 15.9 63.0 .168 705 16.6 4.3 16.1
E052 15.8 57.7 .126 239 5.4 1.5 6.8
M021 21.0 56.0 .190 1352 6.7 1.5 15.2
PA31 27.2 68.0 .559 2593 20.9 7.4 25.7
PE31 19.2 85.0 .761 13 28.9 19.8 27.5
PE52 12.2 67.0 .325 792 0.0 0.0 1.7
U031 17.3 80.1 .608 1515 0.7 5.0 14.0
V051 26.5 77.0 .570 608 1.6 23.1 53.0

29
Test site T RH T OW P SO2 Cl Fe
A032 20.9 73.8 .631 2624 5.0 1.5 5.8
A053 -2.9 84.5 .295 240 5.0 30.2 41.1
B033 24.2 77.0 .582 716 4.4 167.7 128.4
CO21 14.1 81.4 .800 1800 5.0 1.5 13.6
CR21 25.3 88.4 .763 3531 4.7 138.1 371.5
CH21 14.0 82.0 .762 463 18.7 8.4 35.5
E012 10.6 64.5 .271 495 16.2 1.5 22.5
PE51 12.2 67.0 .325 632 0.0 0.0 1.0
V021 27.1 76.6 .496 263 4.2 38.7 23.9

30
Variable Minimum Maximum Mean Std. Dev.
T -3.100 28.200 19.202 7.024
RH 33.300 91.100 73.238 10.267
TOW 0.003 0.976 0.479 0.208
P 13.000 4656.000 1093.423 831.236
SO2 0.000 68.200 13.930 14.928
Cl 0.000 359.800 19.927 42.623
Fe 1.000 371.500 36.212 46.304

31
 T RH TOW P
T 1.00 0.18 0.47 0.44
0.00 0.05 0.00 0.00
RH 0.18 1.00 0.69 0.32
0.05 0.00 0.00 0.00
T OW 0.47 0.69 1.00 0.44
0.00 0.00 0.00 0.00
P 0.44 0.32 0.45 1.00
0.00 0.00 0.00 0.00

32