See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330082095

Artificial Neural Network Models for Airport Rigid Pavement Top-Down Critical
Stress Predictions: Sensitivity Evaluation

Conference Paper · July 2019

DOI: 10.1061/9780784482476.030

CITATIONS READS

4 300

6 authors, including:

Adel Rezaei-Tarahomi Halil Ceylan

The National Academies of Sciences, Engineering, and Medicine Iowa State University
23 PUBLICATIONS 77 CITATIONS 334 PUBLICATIONS 4,792 CITATIONS

SEE PROFILE SEE PROFILE

Kasthurirangan Gopalakrishnan Sunghwan Kim

Iowa State University Iowa State University
213 PUBLICATIONS 4,166 CITATIONS 178 PUBLICATIONS 2,384 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Adel Rezaei-Tarahomi on 26 July 2019.

The user has requested enhancement of the downloaded file.

 Artificial Neural Network Models for Airport Rigid Pavement Top-
Down Critical Stress Predictions: Sensitivity Evaluation

Adel Rezaei-Tarahomi1, Halil Ceylan2, Kasthurirangan Gopalakrishnan3, Sunghwan

Kim4, Orhan Kaya5 and David R. Brill6
1
Graduate Research Student, Iowa State University, Email: adelrt@iastate.edu
2
Professor, Iowa State University; Phone: 515-294-8051; Fax: 515-294-8216; email:
hceylan@iastate.edu (Corresponding Author)
3
Research Scientist, Iowa State University; email: sunghwan@iastate.edu
4
Research Associate Professor; Iowa State University; email: rangan@iastate.edu
5
Graduate Research Student, Iowa State University; email: okaya@iastate.edu
6
FAA Airport Technology R&D Branch, ANG-E262, William J. Hughes Technical
Center, Atlantic City, NJ; email: david.brill@faa.gov

ABSTRACT

To consider top-down cracking failure in current airport rigid pavement design

practices, a multiple-slab pavement structure’s three-dimensional finite element (3D
FE) model should be analyzed for determining critical responses associated with such
failures, and artificial neural networks (ANNs) can be considered a robust and
computationally efficient alternative for 3D FE analysis. This study compares the
effects of airport rigid pavement’s most important properties on the critical tensile
stresses predicted by the ANN models relative to those for 3D-FE solutions. Sensitivity
evaluation of both the 3D-FE solutions and the ANN model predictions for two new
large and heavy aircraft (B777-300 ER and B787-8) have been conducted and are
discussed in this study. The normalized sensitivity index (NSI) has been utilized to
quantify the levels of sensitivity of the critical tensile stresses to changes in PCC slab
thickness, base layer thickness, PCC slab modulus, subgrade elastic modulus,
temperature gradients, and thermal coefficient. The results demonstrate that the
developed ANN model is able to determine top-down critical tensile stress sensitivity
similarly to the 3D-FE model.

INTRODUCTION

Surrogate response models predicting the critical stress associated with top-down
cracking in rigid airfield pavements were previously developed by the authors as part
of a research study sponsored by the Federal Aviation Administrative (FAA) (Kaya, et
al., 2017; Kaya, et al., 2018; Rezaei-Tarahomi, et al., 2017a; Rezaei-Tarahomi, et al.,
2017b; Rezaei-Tarahomi, et al., 2017c; Rezaei-Tarahomi, et al., 2018). The models
were trained using artificial neural network (ANN) algorithms for each type of aircraft
under a variety of loading circumstances. Using ANN models as an alternative to the
three-dimensional finite element (3D FE) pavement response model (NIKE3D-FAA)
employed in the current FAA design process helps with rapid analysis and reduces

1
computer run time for multiple-slab simulation to a matter of seconds. Moreover,
incorporation of ANN surrogate response models into the pavement design process
could also significantly reduce the iteration time for calculation of critical responses
for each type of aircraft in mixed aircraft traffic loading and simultaneously help
implement the whole design process more efficiently. Using rapid ANN response
models can also help expand the current design method beyond the currently-used one-
slab model for limited loading circumstances. Before incorporating these alternative
models into the current FAA design process, they must be tested and validated.
As a follow-up to ANN response model development, this study was conducted
to validate the developed ANN models using sensitivity analysis, by both NIKE3D-
FAA solutions and those predicted by ANN models, of the critical tensile stresses on
top of rigid airfield pavements. Moreover this validation approach helped improving
the models and using the most accurate ones in developing ANN based analyzing
program called ANNFAA (Rezaei-Tarahomiet, et al., 2019). Sensitivity analysis
results of the ANN critical response prediction models developed for the B777-300 ER
and the B787-8 are discussed in this paper. To present sensitivity of the critical tensile
stresses, the normalized sensitivity index (NSI) was obtained, for both NIKE3D-FAA
solutions and the ANN response model predictions, for six input variables: PCC slab
thickness, base layer thickness, PCC slab modulus, subgrade elastic modulus,
temperature gradient and thermal coefficient. These variables have been selected since
they have been obtained as the most effective inputs for the NIKE3D-FAA computed
critical stresses on top of the slab (Rezaei-Tarahomiet, et al., 2017a).

METHODOLOGY

Sensitivity was evaluated for aircraft-only and combined temperature/aircraft loads.

Under mechanical-loading-only conditions, airplane gear loading was applied on the
pavement system, while for simultaneous temperature and mechanical loading, both
mechanical loading and initial temperature-induced curling were applied to the
pavement system. Two aircraft types, the B787-8 and the B777-300 ER, were selected
as representative of wide types of new-generation large aircraft with dual-tandem gear
with three dual gears in tandem. The method used for sensitivity analysis was the same
one-at-a-time sensitivity analysis (OAT) used in the authors’ former studies (Ceylan,
et al., 2014; Rezaei-Tarahomi, et al., 2017a; Schwartz, et al., 2011;).
Table 1 presents the architecture, the training algorithm, and the accuracy
performance of the developed ANN models. Details on ANN model development and
architectures implemented for training the models, as well as the steps involved in the
synthetic database development, have previously been presented in detail (Kaya et al.,
2017; Kaya, et al. 2018; Rezaei-Tarahomi, et al. 2017b; Rezaei-Tarahomi, et al. 2017c;
Rezaei-Tarahomi, et al., 2018). The highest-accuracy ANN model for each aircraft and
each loading condition were utilized for the sensitivity analysis testing. For example,
Table 1 shows that the most accurate sensitivity-testing results for predicting maximum
tensile stresses in X direction at top of the PCC slab concrete (σ XX-Max-Top-Tens. ) for the
B777-300 ER mechanical-loading-only case occurred for the ANN model with 17-15-
15-1 (17 input features and one output with two hidden layers with 15 neurons in each
hidden layer) architecture trained by CGF (Conjugate Gradient_Fletcher-Reeves

2
Update) algorithm, because that model has the lowest Root-Mean-Squared Error
(RMSE) and the highest coefficient of determination, R2. Table 1 also shows that the
σ XX-Max-Top-Tens.– prediction ANN model with 17-5-5-1 architecture trained by CGB
algorithm exhibited the best performance for the B787-8 mechanical-loading-only
case.

Table 1. ANN model prediction performance for mechanical-loading-only and

combined temperature and mechanical loading cases.
Training Accuracy Mechanical only Temperature and
Algorithm Criteria Mechanical
B777-300 ER B787-8 B777-300 ER B787-8
σ XX-Max-Top- σ XX-Max-Top- σ XX-Max-Top- σ XX-Max-Top-
Tens. Tens. Tens. Tens.
LM Architecture 17-25-1 17-10-1 19-5-5-1 19-5-5-1
R2 0.948 0.937 0.850 0.917
AAE (kN/m2) 22.7 21.2 60.0 37.9
RMSE (kN/m2) 38.1 30.7 83.7 51.6
CGB Architecture 17-40-1 17-5-5-1 19-35-1 19-5-5-1
R2 0.958 0.958 0.885 0.898
AAE (kN/m2) 28.5 19.4 52.9 40.9
RMSE (kN/m2) 34.4 25.0 73.3 57.3
CGF Architecture 17-15-15-1 17-30-30-1 19-20-20-1 19-10-10-1
R2 0.963 0.931 0.868 0.929
AAE (kN/m2) 26.5 23.8 60.6 35.8
RMSE (kN/m2) 32.3 32.2 78.5 47.7
CGP Architecture 17-5-5-1 17-30-30-1 19-35-35-1 19-5-5-1
R2 0.957 0.942 0.865 0.917
AAE (kN/m2) 25.8 17.7 63.1 39.1
RMSE (kN/m2) 34.7 29.4 79.3 51.6
OSS Architecture 17-10-10-1 17-30-30-1 19-5-5-1 19-5-5-1
R2 0.956 0.951 0.864 0.924
AAE (kN/m2) 24.8 22.8 46.2 37.9
RMSE (kN/m2) 34.9 27.2 79.9 49.6
Model With Min. RMSE CGF CGB CGB CGF
17-15-15-1 17-5-5-1 19-35-1 19-10-10-1
Note: R2 = coefficient of determination, AAE = absolute average error, MAE = maximum absolute error,
RMSE = root-mean-squared error, LM= Levenberg-Marquardt, CGB= Conjugate Gradient_ Powell-
Beale Restarts, CGF= Conjugate Gradient_Fletcher-Reeves Update, CGP= Conjugate Gradient_ Polak-
Ribiére Update, and OSS= One-step Secant

SENSITIVITY OF THE PREDICTED RESPONSES

TO PCC SLAB THICKNESS

Figure 1 depicts variation of σ XX-Max-Top-Tens obtained from the NIKE3D-FAA and the
ANN response model. Figure 1 (a) shows that σ XX-Max-Top-Tens decreases while the slab
thickness increases for both gear types. Figure 1 shows similar trends in the stress
response due to variation of the slab thickness and NSI is reasonably close for both
aircraft analysis and both loading condition. Figure 1 (b) shows deviation of the ANN
prediction from direct 3D FEM solution for thin PCC thickness for both type of aircraft.
It seems that for thin PCC thickness, since there is higher deflection under combined

3
loading condition, the nonlinearity of the pavement system behavior increases so that
the responses are not precisely predicted.

2,700
NSINIKE3D- B787 = 1.22 NIKE3D_B787-8
2,400 NSIANN-B787 = 1.25 ANN_B787-8

σXX-Max-Top-Tens. (kPa)
2,100 NSINIKE3D- B777 = 0.91 NIKE3D_B777
1,800 NSIANN-B777 = 0.97 ANN_B777
1,500
1,200
900
600
300
0
0 10 20 30 40 50 60 70 80
PCC Slab Thickness (cm)
(a)
2,700
2,400
σXX-Max-Top-Tens. (kPa)

2,100
1,800
1,500
1,200
900 NSINIKE3D-B787 = 0.33 NIKE3D_B787-8
NSIANN-B787 = 0.25 ANN_B787-8
600
NSINIKE3D-B777 = 0.33 NIKE3D_B777
300 NSIANN-B777 = 0.34 ANN_B777
0
0 10 20 30 40 50 60 70 80
PCC Slab Thickness (cm)
(b)

Figure 1. Variation of the σ XX-Max-Top-Tens. from NIKE3D-FAA solutions and ANN

response model predictions vs. PCC slab thickness changes under (a) mechanical
loading only and (b) simultaneous mechanical and temperature loading.

SENSITIVITY OF THE PREDICTED RESPONSES

TO BASE LAYER THICKNESS

Sensitivity evaluation of ANN model prediction to base layer thickness changes is

illustrated in Figure 2. This figure presents sensitivity analysis results for base-layer
thickness values varying from 12.5 to 75 cm with other inputs kept constant and the
B787-8 and B777-300 ER loads are applied. Overall, the sensitivities of the critical
responses to the base layer thickness are similar for both aircraft. It can also be seen
that the ANN response models can predict the effects of increasing base layer thickness
very well. Figure 2 also shows that ANN models provide very accurate predictions for
B787-8 and B777-300 ER loading, with the sensitivity and the exact value of the critical
response predictions almost the same as for the NIKE-3D FAA solutions in the X
direction.
Figure 2 (a) shows that the NSINIKE3D of the slab thickness for σ XX-Max-Top-Tens.
are 0.56 and 0.51 for the B787-8 and B777-300 ER, respectively, and that the ANN
prediction produces NSI values of 0.67 and 0.61 for B787-8 and B777-300 ER,
4
respectively. These results affirm that ANN model predictions are as sensitive as
NIKE3D-FAA solutions to base-layer thickness.

2,700
NSINIKE3D-B787 = 0.56 NIKE3D_B787-8
2,400 NSIANN-B787 = 0.67 ANN_B787-8

σXX-Max-Top-Tens. (kPa)
2,100 NSINIKE3D-B777 = 0.51 NIKE3D_B777
1,800 NSIANN-B777 = 0.61 ANN_B777
1,500
1,200
900
600
300
0
0 10 20 30 40 50 60 70 80 90
Base Layer Thickness (cm)
(a)
2,700
2,400
σXX-Max-Top-Tens. (kPa)

2,100
1,800
1,500
1,200
900 NSINIKE3D-B787 = 0.06 NIKE3D_B787-8
NSIANN-B787 = 0.04 ANN_B787-8
600
NSINIKE3D-B777 = 0.02 NIKE3D_B777
300 NSIANN-B777 = 0.00 ANN_B777
0
0 10 20 30 40 50 60 70 80 90
Base Layer Thickness (cm)
(b)

Figure 2. Variation of the σ XX-Max-Top-Tens . from NIKE3D-FAA solutions and ANN

response model predictions vs. base layer thickness changes under (a)
mechanical loading only and (b) simultaneous mechanical and temperature
loading.

Figure 2 (b) presents sensitivity evaluation of ANN model prediction with

respect to base-layer thickness changes when temperature loading and mechanical
loading are applied simultaneously. These figures show that, as for the mechanical
loading condition, ANN critical response prediction models can predict stresses when
the base layer thickness is changing with the same trend as for the FE solution.
Comparing Figure 2 (a) to Figure 2 (b) reveals that the sensitivity of the critical top
tensile stresses to the base layer thickness is much higher for the mechanical-loading-
only case than for combined temperature and mechanical loading. In the presence of
the temperature load, the thermal properties of the slab (temperature gradient and
thermal coefficient) affect the tensile stresses, while base and subbase layer properties
have less effect on the stresses.

5
SENSITIVITY OF THE PREDICTED RESPONSES
TO PCC SLAB MODULUS

The sensitivity of tensile stresses to PCC slab modulus variation is illustrated in Figure
3 (a) for the mechanical-loading-only case. As seen in this figure, response prediction
models can replicate the NIKE3D-FAA FE solution so that the sensitivity of the tensile
stresses in X direction to PCC slab modulus changes are very similar for the ANN
prediction and the FE solution. As Figure 3 (a) shows, the NSIB787-8 values are 0.27
and 0.28 for the NIKE3D-FAA solution and the ANN prediction, respectively. The
similar sensitivity index values obtained for the ANN prediction and the actual FE
solution supports the assertion that ANN response models seem promising in predicting
critical stresses and their variational behavior when the input parameters are changing
only slightly.
Figure 3 (b) shows that ANN critical response models provide similar
sensitivity as the actual solutions to the PCC slab modulus even when temperature
loading is present. As the figure shows, the NSI B787-8 values are 0.47 and 0.51 for the
NIKE3D-FAA solution and the ANN prediction, respectively.

2,700 NSINIKE3D-B787 = 0.27 NIKE3D_B787-8

2,400 NSIANN-B787 = 0.28 ANN_B787-8
σXX-Max-Top-Tens. (kPa)

2,100 NSINIKE3D-B777 = 0.44 NIKE3D_B777

1,800 NSIANN-B777 = 0.60 ANN_B777
1,500
1,200
900
600
300
0
0 10,000 20,000 30,000 40,000 50,000
PCC Slab Modulus (MPa)
(a)
2,700
2,400
σXX-Max-Top-Tens. (kPa)

2,100
1,800
1,500
1,200
900 NSINIKE3D-B787 = 0.47 NIKE3D_B787-8
NSIANN-B787 = 0.51 ANN_B787-8
600
NSINIKE3D-B777 = 0.49 NIKE3D_B777
300 NSIANN-B777 = 0.54 ANN_B777
0
0 10,000 20,000 30,000 40,000 50,000
PCC Slab Modulus (MPa)
(b)

Figure 3. Variation of the σ XX-Max-Top-Tens . from NIKE3D-FAA solutions and ANN

response model predictions vs. PCC slab modulus changes under (a) mechanical
loading only and (b) simultaneous mechanical and temperature loading.

6
SENSITIVITY OF PREDICTED RESPONSES TO SUBGRADE MODULUS

Figure 4 presents the sensitivity of the tensile stresses obtained by the NIKE3D-FAA
solution and the ANN response models to subgrade modulus variation from 570 to
1,890 MPa for B787-8 and B777-300 ER loading while other inputs are kept constant.
The figure shows that the sensitivity values of the different stress responses are very
close for the NIKE-3D results and the ANN response model prediction. Figure 4 also
shows that tensile stress is more sensitive to subgrade modulus changes under B777-
300 ER loading than under B787-8 loading. It can also be observed that the NSI values
obtained by B777-300 ER’s ANN response models are closer to those of the NIKE3D-
FAA solution than are those for the B787-8 models. As Figure 4 (a) shows, NSIB777 is
0.59 and 0.60, respectively, for the NIKE3D solution and the ANN prediction, and
NSIB787 is, respectively, 0.41 and 0.28.

2,700 NSINIKE3D-B787 = 0.41 NIKE3D_B787-8

2,400 NSIANN-B787 = 0.28 ANN_B787-8
σXX-Max-Top-Tens. (kPa)

2,100 NSINIKE3D-B777 = 0.59 NIKE3D_B777

1,800 NSIANN-B777 = 0.60 ANN_B777
1,500
1,200
900
600
300
0
0 50 100 150 200 250 300 350 400
Subgrade Modulus (MPa)
(a)
2,700
2,400
σXX-Max-Top-Tens. (kPa)

2,100
1,800
1,500
1,200
900 NSINIKE3D-B787 = 0.04 NIKE3D_B787-8
NSIANN-B787 = 0.02 ANN_B787-8
600
NSINIKE3D-B777 = 0.16 NIKE3D_B777
300 NSIANN-B777 = 0.10 ANN_B777
0
0 50 100 150 200 250 300 350 400
Subgrade Modulus (MPa)
(b)

Figure 4. Variation of the σ XX-Max-Top-Tens . from NIKE3D-FAA solutions and ANN

response model predictions vs. subgrade modulus changes under (a) mechanical
loading only and (b) simultaneous mechanical and temperature loading.

SENSITIVITY OF THE PREDICTED RESPONSES

TO TEMPERATURE GRADIENT AND THERMAL COEFFICIENT

Figure 5 shows the sensitivity evaluation of the ANN model predictions to slab
temperature gradient changes. The figure presents the sensitivity analysis results for
7
temperature gradient values varying from -6.6 to -41.6 oC/m while other inputs remain
constant and the B787-8 load and the B777-300 ER load are applied. Overall, these
figures show that the sensitivities of the critical responses to the temperature gradient
for both aircraft are very similar. It can also be seen that the ANN response models can
very accurately predict the effect of increasing temperature gradient.

2,700 NSINIKE3D-B787 = 0.48

2,400 NSIANN-B787 = 0.44

σXX-Max-Top-Tens. (kPa)
2,100 NSINIKE3D-B777 = 0.39
NSIANN-B777 = 0.37
1,800
1,500
1,200
900 NIKE3D_B787-8
600 ANN_B787-8
NIKE3D_B777
300
ANN_B777
0
0 10 20 30 40 50
Temperature Gradient (oC/m)

Figure 5. Variation of the σXX-Max-Top-Tens. obtained by NIKE3D-FAA

solutions and ANN response model predictions vs. temperature gradient
changes.

The sensitivity of the tensile stresses to thermal coefficient variation is

illustrated in Figure 6. The response prediction models can replicate the NIKE3D-FAA
FE solution. The sensitivity of predicted stress to variations in the thermal coefficient
is similar in the ANN prediction and the FE solution. Figure 6 shows NSI B787-8 values
of 0.49 and 0.50 for the NIKE3D-FAA solution and the ANN prediction, respectively.
Figure 6 demonstrates that, for thermal coefficients higher than 11.9×10-6 1/oC, while
the B787-8’s ANN response prediction and NIKE3D solution do not exactly agree, the
difference is less than 206.8 kPa. For the B777-300 ER the ANN response predictions
fit the NIKE3D solutions. The ability of the ANN critical response models to predict
maximum stresses and their sensitivities to input values can assure accurate replication
of the NIKE3D-FAA solutions for the critical stresses.

2,700
2,400
σXX-Max-Top-Tens. (kPa)

2,100
1,800
1,500
1,200
900 NSINIKE3D-B787 = 0.49 NIKE3D_B787-8
600 NSIANN-B787 = 0.50 ANN_B787-8
NSINIKE3D-B777 = 0.45 NIKE3D_B777
300 NSIANN-B777 = 0.52 ANN_B777
0
5 6 7 8 9 10 11 12 13 14 15
Thermal Coefficient (×10-6) (1/oC)

Figure 6. Variation of the σ XX-Max-Top-Tens . obtained by NIKE3D-FAA solutions

and ANN response model predictions vs. thermal coefficient changes.

8
DISCUSSION

Table 2 shows rigid pavement feature rankings based on the sensitivity of the critical
top tensile stresses to these features under mechanical-loading-only circumstances. The
pavement features ranking is almost the same for both the NIKE3D-FAA solution and
the ANN critical response prediction. The top three most sensitive parameters have the
same ranking order for the B777-300 ER’s NIKE3D solution and ANN prediction, and
the top four most sensitive parameters and their ranking order are the same for the
B787-8’s NIKE3D solution and its ANN prediction.

Table 2. NSI value showing the sensitivity of top tensile stresses to different
pavement characteristics for NIKE3D FE solutions and ANN predictions.
Mechanical Loading Only
B777-300 ER B787-8

NIKE3D ANN NIKE3D ANN

Input NSI Input NSI Input NSI Input NSI

variables variables variables variables
1 Slab 0.91 Slab 0.97 Slab 1.22 Slab 1.25
Thickness Thickness Thickness Thickness
2 Subgrade 0.59 Base 0.61 Base 0.56 Base 0.67
Modulus Thickness Thickness Thickness
3 Base 0.51 Subgrade 0.60 Subgrade 0.41 Subgrade 0.28
Thickness Modulus Modulus Modulus
4 Slab 0.44 Slab 0.60 Slab 0.27 Slab 0.28
Modulus Modulus Modulus Modulus
Simultaneous Mechanical and Temperature Loading
B777-300 ER B787-8

NIKE3D ANN NIKE3D ANN

Input NSI Input NSI Input NSI Input NSI

variables variables variables variables
1 Slab 0.49 Slab 0.54 Thermal 0.49 Slab 0.51
Modulus Modulus Coefficient Modulus
2 Thermal 0.45 Thermal 0.52 Temperature 0.48 Thermal 0.50
Coefficient Coefficient Gradient Coefficient
3 Temperature 0.39 Temperature 0.37 Slab 0.47 Temperature 0.44
Gradient Gradient Modulus Gradient
4 Slab 0.33 Slab 0.34 Slab 0.33 Slab 0.25
Thickness Thickness Thickness Thickness
5 Subgrade 0.16 Subgrade 0.10 Base 0.06 Base 0.04
Modulus Modulus Thickness Thickness
6 Base 0.02 Base 0.00 Subgrade 0.04 Subgrade 0.02
Thickness Thickness Modulus Modulus

Table 2 also provides a rigid pavement feature ranking based on the sensitivity of the
critical top tensile stresses under combined temperature and mechanical loading. The

9
table shows that the pavement feature rankings based on the NIKE3D-FAA solution
and ANN critical response prediction are almost identical. Table 2 shows that the six
parameters and their ranking order are the same for the B777-300 ER’s NIKE3D
solution as for the ANN prediction. Also, for B787-8, the six most sensitive parameters
except for slab modulus have almost the same ranking, and the NSI value obtained by
NIKE3D solution and ANN model prediction for the slab modulus are very similar
(0.47 and 0.51 respectively).
From sensitivity analysis, it can be concluded that the ANN model developed
to study different critical responses for each type of aircraft satisfactorily predict both
the actual values and the response sensitivities to pavement feature alterations. As
mentioned previously, the use of such ANN models as surrogate 3-D FE solutions
would facilitate routine design orders of magnitude faster than traditional 3-D FE
analysis. Sensitivity analysis is also a good method of testing the trained ANN models
with independent testing data sets.

CONCLUSIONS

Sensitivity of artificial neural network (ANN) response model predictions with respect
to some of the most effective rigid airfield pavement properties have been determined.
This evaluation can help shed light on the performance of ANN based multiple-slab
response models when a slight variation in pavement properties occurs. Key findings
and major conclusions of this study are as follows:

• ANN models developed for studying different critical responses for each type of
aircraft can predict the actual values and the sensitivity of the responses to
pavement feature alterations. Sensitivity analysis is also a good method for testing
trained ANN models with an independent testing data set.
• Virtually identical sensitivity of the ANN response model predictions and the
NIKE3D-FAA solutions with respect to PCC slab variations suggests that ANN
models are promising for predicting σ XX-Max-Top-Tens. even for slight changes in PCC
slab thickness for both mechanical and combined mechanical and temperature
loading. In other words, an ANN model can precisely predict top tensile stresses
when the PCC slab thickness varies even when temperature loading is applied.
• Under both mechanical-only loading and simultaneous temperature and mechanical
loading conditions, ANN critical response models provide similar sensitivity
indices to those from finite element solutions with respect to PCC slab modulus
changes.
• ANN model sensitivity analysis results suggest that their critical tensile stress
predictions are as sensitive as NIKE3D-FAA solutions with respect to thermal
coefficient and temperature gradients. In other words, the results indicate that ANN
models can detect the influence of a PCC slab’s thermal properties variations on
the critical top tensile stresses prediction, and can also lead to better prediction of
top-down cracking behavior even under complex combined temperature and
mechanical loading conditions.
• Under simultaneous temperature and mechanical loading conditions, a slab’s
thermal properties (temperature gradient and thermal coefficient) significantly

10
 affect the critical top tensile stresses, while base and subbase layer properties have
less effect on stresses.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge the Federal Aviation Administration (FAA) for
supporting this study. The contents of this paper of this paper reflect the views of the
authors, who are responsible for the facts and accuracy of the data presented within.
The contents do not necessarily reflect the official views and policies of the FAA. The
paper does not constitute a standard, specification, or regulation.

REFERENCES

Ceylan, H., Kim, S., Gopalakrishnan, K., Schwartz, C. W., and Li, R. (2014).
“Sensitivity Analysis Frameworks for Mechanistic-Empirical Pavement Design of
Continuously Reinforced Concrete Pavements.” Construction and Building
Materials 73: 498–508.
Kaya, O., Rezaei-Tarahomi, A., Gopalakrishnan, K., Ceylan, H., Kim, S., and Brill, D.
R. (2017). Alternative Approaches to Determining Robust ANN Based Models
for Predicting Critical Airport Rigid Pavement Responses. In Airfield and
Highway Pavements, 51–60.
Kaya, O., Rezaei-Tarahomi, A., Ceylan, H., Gopalakrishnan, K., Kim, S., and Brill, D.
R. (2018). “Neural Network--Based Multiple-Slab Response Models for Top-
Down Cracking Mode in Airfield Pavement Design.” Journal of Transportation
Engineering, Part B: Pavements 144 (2). American Society of Civil Engineers:
4018009.
Rezaei-Tarahomi, A., Kaya, O., Ceylan, H., Gopalakrishnan, K., Kim, S., and Brill, D.
R. (2017a). “Sensitivity Quantification of Airport Concrete Pavement Stress
Responses Associated with Top-down and Bottom-up Cracking.” International
Journal of Pavement Research and Technology 10 (5). Elsevier: 410–20.
Rezaei-Tarahomi, A., Kaya, O., Ceylan, H., Gopalakrishnan, K., Kim, S., and Brill, D.
R. (2017b). Neural Networks Prediction of Critical Responses Related to Top-
Down and Bottom-Up Cracking in Airfield Concrete Pavement. In 10th
International Conference on the Bearing Capacity of Roads, Railways and
Airfields. Athens, Greece.
Rezaei-Tarahomi, A., Kaya, O., Ceylan, H., Kim, S., Gopalakrishnan, K., and Brill, D.
R. (2017c). “Development of Rapid Three-Dimensional Finite-Element Based
Rigid Airfield Pavement Foundation Response and Moduli Prediction Models.”
Transportation Geotechnics 13. Elsevier Ltd: 81–91.
https://doi.org/10.1016/j.trgeo.2017.08.011.
Rezaei-Tarahomi, A., Kaya, O., Ceylan, H., Kim, S., and Brill, D. R. (2018). Neural
Network Algorithms for Rigid Airfield Pavement Responses. In Advances in
Materials and Pavement Performance Prediction, Doha, Qatar.
Rezaei-Tarahomi, A., Ceylan, H., Gopalakrishnan, K., Kim, S., Kaya, O. and Brill, D.
R. (2019). Development of a Top-Down Cracking Rapid Analysis Tool for Rigid

11
 Airfield Pavement Systems. In Transportation Research Board 98th Annual
Meeting, Washington D.C.
Schwartz, C. W., Li, R., Kim, S., Ceylan, H., and Gopalakrishnan, K. (2011).
“Sensitivity Evaluation of MEPDG Performance Prediction.” Report for National
Cooperative Highway Research Program Transportation Research Board.

12

View publication stats