International Journal of Pavement Engineering

ISSN: 1029-8436 (Print) 1477-268X (Online) Journal homepage: https://www.tandfonline.com/loi/gpav20

ANNFAA: artificial neural network-based tool for

the analysis of Federal Aviation Administration’s
rigid pavement systems

Adel Rezaei Tarahomi, Orhan Kaya, Halil Ceylan, Kasthurirangan

Gopalakrishnan, Sunghwan Kim & David R. Brill

To cite this article: Adel Rezaei Tarahomi, Orhan Kaya, Halil Ceylan, Kasthurirangan
Gopalakrishnan, Sunghwan Kim & David R. Brill (2020): ANNFAA: artificial neural network-based
tool for the analysis of Federal Aviation Administration’s rigid pavement systems, International
Journal of Pavement Engineering, DOI: 10.1080/10298436.2020.1748627

To link to this article: https://doi.org/10.1080/10298436.2020.1748627

Published online: 16 Apr 2020.

Submit your article to this journal

Article views: 19

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=gpav20
INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING
https://doi.org/10.1080/10298436.2020.1748627

ANNFAA: artificial neural network-based tool for the analysis of Federal Aviation
Administration’s rigid pavement systems
a
Adel Rezaei Tarahomi , Orhan Kayab, Halil Ceylanc, Kasthurirangan Gopalakrishnanc, Sunghwan Kimc and David
R. Brilld
a
Turner-Fairbank Highway Research Center, Federal Highway Administration, Mclean, VA, USA; bDepartment of Civil Engineering, Adana Alparslan
Turkes Science and Technology University, Saricam, Adana, Turkey; cDepartement of Civil, Construction, and Environmental Engineering, Iowa State
University, Ames, IA, USA; dFAA Airport Technology R&D Branch, William J. Hughes Technical Center, ANG-E262, Atlantic City, NJ, USA

ABSTRACT ARTICLE HISTORY

Three-dimensional Finite Element (3D-FE) stress computations involved in the current rigid airport Received 17 October 2019
pavement design methodology, are time consuming when considering top-down cracking failure Accepted 23 March 2020
mode. In this study, Artificial Neural Network (ANN) models are integrated into a tool called ANNFAA to
KEYWORDS
replace such 3D-FE computations. ANNFAA makes use of the best ANN models developed in MATLAB Rigid airfield pavements;
for 156 different airplanes without requiring any additional software installation or cumbersome critical tensile stresses;
learning of a new program. Within ANNFAA development, about 4,000 of 3D-FE simulations and many artificial neural networks;
ANN models have been developed for each of these airplanes. Three useful tools were also developed Top-Down cracking
using C# and MATLAB for implementing the 3D-FE analysis, post-processing the results, training the
ANN models, and determining accuracy and performance of the ANN models. ANNFAA provides an
accurate and rapid procedure for practitioners, engineers, and researchers for computing the critical
stress responses associated with top-down cracking in multiple-slab rigid airfield pavements. This
should make pavement design and analysis more practical, especially when a significantly large
number of different cases that include top-down cracking failure mode are investigated. Also, this will
help when currently used bottom-up cracking mode in the FAA standard rigid pavement design
procedures is being considered in a design.

Introduction
of many complex and non-linear pavement engineering pro-
The significant potential of artificial intelligence (AI) tech- blems and mechanisms, including some that have not yet
niques in solving resource-intensive complex problems has been well understood and formulated (Ceylan et al. 1999, Cey-
increased interest in using such methods in different engineer- lan 2002, Ceylan and Gopalakrishnan 2007, Ceylan et al. 2013,
ing areas, including civil engineering fields such as structural Ceylan et al. 2014, Ghanizadeh and Ahadi 2015, Elbagalati et al.
engineering, environmental and water resources engineering, 2017, Rezaei-Tarahomi et al. 2017a, Kaya et al. 2018). Neural
infrastructure resilience characterising in disasters, traffic networks have been found to be very reliable and versatile com-
engineering, geotechnical engineering, and pavement engineer- putational tools for determining and predicting future con-
ing (Rezaei-Tarahomi et al. 2017a, Salehi and Burgueño 2018, dition and performance of existing pavement systems related
Rezaei-Tarahomi et al. 2018, Rastegaripour et al. 2019, Gha- to maintenance and rehabilitation decision making actions
semi et al. 2019a, 2019b). During recent decades, artificial for deteriorated pavement sections, developing cost-effective
neural networks (ANN) is an AI technique that has become prediction models for estimating the future condition of a pave-
popularly used not only in research activities but also in com- ment section, predicting properties of pavement layers and
mercial applications. pavement surfaces, and understanding and characterising Port-
ANN has been addressed to a wide variety of pavement land cement concrete (PCC) and asphalt cement concrete
engineering issues associated with design, analysis, distress (ACC) design and behaviour (Gopalakrishnan et al. 2009,
evaluation, performance, maintenance, rehabilitation, and Ceylan et al. 2014).
management of both highway and airport pavements (flexible, One reason for using ANN is to try to avoid long and unpre-
rigid, and composite) (Ceylan et al. 1999, Ceylan 2002, Ceylan dictable analysis running times associated with the multiple-
and Gopalakrishnan 2007, Ceylan et al. 2013, Ceylan et al. slab model and curled slabs addressed in a 3D-finite element
2014, Elbagalati et al. 2017, Ghanizadeh and Ahadi 2015, (FE) computer program called NIKE3D_FAA (Brill 1998,
Ling et al. 2017, Rezaei-Tarahomi et al. 2017a, Kaya et al. 2000), currently employed in the FAA Rigid and Flexible Itera-
2018). Use of ANN in many research studies related to pave- tive Elastic Layered Design (FAARFIELD) software for rigid
ment engineering has shown that ANN holds promise for use pavement design. Note that FAARFIELD 1.42 is the latest ver-
in investigating, modelling, and seeking better understanding sion of standard thickness design software accompanying AC

CONTACT Adel Rezaei Tarahomi [email protected]

© 2020 Informa UK Limited, trading as Taylor & Francis Group
2 A. REZAEI TARAHOMI ET AL.

150/5320-6F Airport Pavement Design and Evaluation (Ceylan post-processing program capable of analysing all output
et al. 2013), and NIKE3D_FAA is a modification of the files, calculating maximum and minimum stress components
NIKE3D program originally developed by the Lawrence Liver- and deflections for each case, and finding critical response
more National Laboratory (LLNL) of the U.S. Department of locations from the NIKE3D-FE outcomes has also been devel-
Energy (Brill 1998, 2000). The FAA has also developed Finite oped (see Figure 3). Using this post-processing program, not
Element Analysis FAA (FEAFAA) software that makes use of only normal stresses (σXX and σYY) and principal stresses (σ1
NIKE3D_FAA as a stand-alone tool for 3D-finite element and σ2) are determined and calculated at each nodal point
(3D-FE) analysis of multiple-slab rigid airport pavements and of the FE model, but also, maximum σXX, σYY, σ1, and σ2 at
overlays. Incorporating ANN surrogate response models into slab top, as well as corresponding principal stress angle ⍰p,
the rigid pavement design process can significantly reduce are then determined. Principal stress is a combination of nor-
the iteration time for calculating critical responses of each mal stressessx andsy that also considers the shear stress (txy ),
type of aircraft in mixed aircraft traffic loading and significantly i.e. principal stress σ1 represents an extreme value of tensile
help with more efficient implementation of the whole routine stress at a given point of the structure.
design process. For more details about the mechanism of top-down cracking
The primary objective of this paper is to develop a compre- and the sensitivity of the stresses associated with top-down
hensive tool employing the surrogate computational response cracking, please refer to the sensitivity analysis study by Rezaei
models published by the authors in previous studies (Rezaei- Tarahomi et al. (2017b).
Tarahomi et al. 2017a, Kaya et al. 2018, Rezaei-Tarahomi
et al. 2018). This new tool will be suitable for implementation
in next generation of FAA thickness design software, FAAR-
Development of an ANN-based multiple-slab
FIELD 2.0, and it will also provide a close estimate of the
response models
top-down bending stress computed by NIKE3D_FAA both
for mechanical loading only and combined airplane and temp- Both one and two hidden layer architectures with numbers of
erature loading of rigid airport pavements. This will enable fas- neurons varying from 5 to 50, and five different algorithms
ter 3D-FE computations of design stresses in FAARFIELD 2.0. were used. Avoiding being caught at a local minimum, the
Figure 1 simply demonstrates the contribution of this study’s training process of each ANN model was restarted at 10 differ-
outcomes in the iterative process of the rigid airfield pavement ent initial conditions. It leads the ANN to start from a new
thickness design in the FAARFIELD. The new ANN-based point in the error surface and consequently to find the desired
FAA rigid pavement analysis tool (ANNFAA) developed in global minimum error on the entire error domain. Otherwise,
this study can be used in the design process instead of 3D-FE ANN will be trapped in a minimum error at the near neigh-
analysis can predict the critical stresses, i.e. the maximum hori- bourhood called local minimum.
zontal stresses at the slab surfaces. The damage factor (DF), the Considering various algorithms and architectures, totally
ratio of tensile strength (R) to the critical stress value, is then 1,000 ANN models were developed for each critical response
determined, followed by computation of a cumulative damage (σXX-Max-Top-Tens., σYY-Max-Top-Tens., and σ1-Max-Top-Tens.). Since
factor (CDF) for each type of aircraft in the applied aircraft two loading conditions were used, ultimately 6,000 ANN
traffic mix. The desired design thickness is achieved when the models (number of responses × loading condition× ANN
total CDF (TCDF) either reaches 1 or meets the tolerance; models developed for each response = 3 × 2× 1,000) were devel-
otherwise the PCC thickness is incremented (hi = hi + 1) and oped for each aircraft loading. To generate this number of ANN
another iteration performed. models for each of the 156 aircraft provided in FEAFAA 2.0, an
ANN model development tool was created in MATLAB.
The project team wrote a utility program in MATLAB to
3D-FE simulation batch run automation
import data, train the models, and export and find the opti-
Developing an extensive database of input-output records from mum models. Automating data import significantly reduces
FEAFAA 2.0 requires use of automation programes to reduce the chance of mistakes during manual data entry.
time required, increase accuracy, and decrease investment in Some features of the ANN model development tool are:
human interaction. For that purpose, two tools have been
developed to aid in performing batch runs of NIKE3D-FAA . Creating a database of all trained ANN models
provided in FEAFAA 2.0 software and post-processing the . Training ANN models using various algorithms
raw output files generated by NIKE3D_FAA to summarise . Including a range of hidden layer sizes
maximum and minimum responses and their locations on . Developing ANN models in both single-mode and batch
pavement systems. Table 1 shows the range of input parameters mode
considered for FEM analysis by NIKE3D-FAA. . Developing any single ANN model (e.g. for any response
The C# programming tool was utilised along with an and a specific algorithm and hidden layer size)
AutoIt scripting tool (2017) to create a powerful automation . Developing several different ANN models (e.g. for all
program scheme for executing 3D-FE simulation with responses with different algorithm and various hidden
NIKE3D-FAA. The ultimate product can automatically per- layer sizes)
form FEAFAA 2.0 software, save elements and nodal point . Calculating accuracy parameters (R2, RMSE, AAE) for each
information files, execute NIKE3D_FAA for analysis, and ANN model, to be compared with NIKE3D solutions for
save the response output files for each case (see Figure 2). A selecting the most accurate model for each response
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 3

Figure 1. Flowchart displaying the contribution of this study in FAA rigid airfield pavement design.

. Exporting the most accurate ANN model results into . Step2: Training ANN models
spreadsheets. . Step3: Optimising and exporting the data

The utility program executes three main steps: Step 1: creating workspace
The first step in developing the ANN models is to create the
. Step1: Creating a workspace for all inputs and outputs input and output matrix. Figure 4 the graphical user interface
4 A. REZAEI TARAHOMI ET AL.

Table 1. Ranges of input parameters used for finite element analysis runs. of the computer program written by the author for this pur-
Ranges pose. The source file used for creating the inputs and outputs
Inputs Min Max matrix is the spreadsheet obtained from post-processing of
PCC Slab FE analysis outputs (see Figure 5). Since there are 17 inputs
Modulus (GPa) (psi) 20.7 (3 × 106) 48.3 (7 × 106) for mechanical-loading-only cases, and 19 inputs for simul-
Thickness (cm.) (in.) 15.2 (6) 60.9 (24)
Poisson ratio 0.10 0.20 taneous mechanical and temperature loading, the type of load-
Base ing must first be specified to determine the number of inputs.
Modulus (GPa) (psi) 1.4 (2 × 105) 13.8 (2 × 106) At the first step, a matrix of all cases used for developing
Thickness (cm.) (in.) 10.0 (4) 76.2 (30)
Poisson ratio 0.15 0.25 ANN models is created and a specified percentage of all cases
Granular subbase is randomly chosen for independent testing purposes. The
Modulus (GPa) (psi) 1×10−1 (15,000) 5.2×10−1 (75,000) first step is terminated when a MATLAB file (workspace)con-
Thickness (cm.) (in.) 15.2 (6) 127 (50)
Poisson ratio 0.20 0.40 taining input and output matrices both for developing ANN
Subgrade models and independent testing has been created. Results
Modulus (GPa) (psi) 2.1×10−2 (3000) 3.4×10−1 (50,000) obtained by conducting subsequent steps are saved in the work-
Poisson ratio 0.30 0.45
Slab dimension (m.) (ft.) 4.6 (15) 9.1 (30) space created during the first step.
Slab number of elements 30
Number of slabs 9
Foundation number of elements 30
Loading angle 0 90 Step 2: training ANN models
Temperature gradient −0.3 (−2) +0.3 (2)
(°C/cm) (°F/in.) The second step for developing ANN models is to train them.
Thermal coefficient 7.4×10−6 (4.1 × 10−6) 12.9×10−6 (7.2 × 10−6) As shown in Figure 4, the 12 critical responses accommodated
(1/°C) (1/°F)
Equivalent joint stiffness 2.7 ×10−1 (1.0 ×103) 162.6 (6.0 ×105)
for developing ANN models in the program are: Critical tensile
(GPa/m) (psi/in) stresses in X and Y directions (σXX, σYY), Von Mises, deflection,
and principal stresses (σ1 and σ2) at top and bottom of the PCC

Figure 2. Graphical interface of FEAFAA batch run program.

Figure 3. Graphical interface of FEAFAA batch run post-processing program.

 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 5

Figure 4. ANN model development program.

Figure 5. Input file data spread sheet.

slab. Also five high-performance back-propagation training 2017) have been utilised in this study. The three different con-
algorithms that use standard numerical optimisation tech- jugate gradient methods that have been used for evaluating pre-
niques such as Conjugate Gradient, One-step Secant (OSS), diction accuracy improvement in this study are: Fletcher-
and Levenberg-Marquardt (LM) algorithms (Beale et al. Reeves Update (CGF), Polak-Ribiére Update (CGP), and
6 A. REZAEI TARAHOMI ET AL.

Powell-Beale Restarts (CGB) (Beale et al. 2017, Gopalakrishnan of 10,000 and 1,000, respectively, were assigned for the maxi-
2010). The various versions of conjugate gradient algorithms mum number of epochs for training and the maximum num-
are distinguished by the method through which the ratio ber of iterations for validation checking. The maximum
between the current and the previous gradient is computed in number of allowable iterations for validation check would
determining a new search direction (Beale et al. 2017). be used to stop the training if the error failed to improve
The One-step Secant method attempts to find compromise over the validation set.
between full quasi-Newton algorithms and conjugate gradient In this study, 10% of each dataset was assigned for indepen-
algorithms. It has the additional advantage that a new search direc- dent testing and 90% was utilised for model development, and
tion can be calculated without computing a matrix inverse, and it datasets were divided into training (70% of data set elements
requires only slightly more storage and computation per epoch for model development), validation (20% of data set elements
than the conjugate gradient algorithms (Beale et al. 2017). The for model development), and test sets (10% of data set elements
Levenberg-Marquardt algorithm is the fastest method for training for model development). The result was that a 2000-element
as many as several hundred weight-neural networks, but it requires data set, with 200, 1260, 360, and 180 data set elements, respect-
more storage than conjugate gradient algorithms. ively, assigned to independent testing, training, validation, and
Most practical neural networks have two or three layers testing sets. The training set is used to compute gradients or
(Hagan et al. 2014). More neurons, and therefore more par- Jacobians and to determine the weight update at each iteration.
ameters, result in more network flexibility, creating a significant The testing and validation set are indicators of what is happen-
possibility that the network will overfit the training data and fail ing to the network function ‘in-between’ the training points,
to generalise well to new situations (Hagan et al.2014). In this and their errors are monitored during the training process.
study, two and three-layer networks (one and two hidden layers When the error on the validation set goes up over several iter-
and one output layer) have been used for developing ANN ations, the training is stopped, and the weights that produced
models. As shown in Figure 4, the number of hidden layers the minimum error on the validation set are used as the final
and the numerical range of neurons in each layer can be trained network weights (Brill 1998). Also, the testing and vali-
defined in the ANN model development tool. dation are conducted during the training process to evaluate
The global approach tends to require fewer neurons in the hid- whether the models were overfitted or not. The independent
den layer, but if the number of neurons is too small, the network testing was used to measure the performance of the models
may not adequately fit the training data and produce an under- to choose the optimum model through 1,000 models developed
fitting result. Therefore, to determine the optimal number of neur- for each response. In this study, the independent testing is used
ons, the networks were trained using between 5 and 50 neurons in after a model was trained, however, the testing occurs during
a hidden layer for evaluating improvement in prediction accuracy. the training process at each epoch.
The architectures used for training ANN models for mechanical- As illustrated in Figure 4, user can define ANN model
loading-only cases are shown in Figure 6. Figure 6(a) depicts an retraining number which is an additional step made a part
architecture with one hidden layer with17 predictors and one out- of the training process to ensure that it has not fallen into
put layer, with the number of neurons (n) in the hidden layer vary- a local minimum. Since a single training run may not pro-
ing from 5 to 50. Figure 6(b) depicts an architecture with two duce optimal performance because of the possibility of
hidden layers with 17 predictors and one output layer. reaching a local minimum of the performance surface, it is
Figure 7 illustrates the architectures used for training ANN best to restart the training at several different initial con-
models for cases of simultaneous mechanical and temperature ditions and select the network that produces the best per-
loading. Figure 7(a,b) depict architecture with one hidden layer formance. Since five to ten restarts will almost always
with 19 predictors and one output layer, and architecture with produce a global optimum (Hamm et al. 2007), 10 different
two hidden layers with 19 predictors and one output layer initial conditions for each ANN model architecture have
respectively, while the number of neurons (n) in the hidden been considered in this study.
layer(s) varying from 5 to 50.
Regularisation and early stopping techniques are used for
Step 3: export the optimum ANN model
preventing overfitting and improving generalisation. In this
study, the early stopping technique has been used. Early stop- The program computes several measures of accuracy prediction
ping restricts network weights by stopping training before the (MSE, RMSE, R2 and AAE), which are started in the workspace
weights have converged to the minimum value of the squared along with the trained ANNs. The optimal model is defined as the
error. The basic underlying idea of this method is that, as model with the lowest RMSE among those considered. In step
training progresses, the network uses more and more of its three the optimum model results, the targets and predictions, and
weights until all weights have been fully used when training all prediction accuracy measures, are exported to a spreadsheet.
reaches an error surface minimum. Increasing the number
of training iterations increases the complexity of the resulting
Accuracy evaluation of ANN-based multiple-slab
network. If training is stopped before the minimum is
response model
reached, the network will effectively be using fewer par-
ameters and therefore be less likely to overfit (Hagan et al. In this section, the ANN model performance results for various
2014). As shown in Figure 4, the user can define the maximum aircraft types are presented and compared with NIKE3D-FAA
number of epochs for training and the maximum number of FE solutions and model accuracy and the robustness is dis-
allowable iterations for validation check. For this study, values cussed. Results from the ANN models related to predicting
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 7

Figure 6. ANN model architectures (a) two-layer and (b) three-layer for mechanical-loading-only cases.

critical slab-top tensile stresses most associated with top-down maximum absolute error (MAE), R2, and RMSE are the stan-
cracking are also demonstrated in this section. dard statistical metrics used to measure model performance
and their accuracy in this study. These are all useful measures
widely used as important criteria in model evaluation and
Descriptions of accuracy assessment criteria
assessment. Since any individual metric usually projects and
While several measures have for decades been used to assess emphasises only one aspect of model performance, a combi-
model performance and accuracy, there is no consensus nation of different assessment criteria are usually required to
about the most appropriate measure for model errors. AAE, completely assess model performance and accuracy.
8 A. REZAEI TARAHOMI ET AL.

Figure 7. ANN model architectures (a) two-layer and (b) three-layer for simultaneous temperature- and mechanical-loading cases.

While R2 is a relative measure of fit, RMSE is an absolute quantifying the error using the same measuring unit as the
measure of fit (Grace-Martin 2013). While R 2 is the most com- responses is often needed, and RMSE, AAE, and MAE use
monly used metric to help determine the percentage of variance the same units as the model’s prediction outputs, (psi) in this
in the target response produced by a model’s prediction, study. Among these measures, RMSE is distinct from AAE in
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 9

avoiding the use of absolute value. RMSE indicates how accu- loading), the ANN model with an architecture of 19-5-5-1
rately the model predicts the response, the most important cri- and a CGF training algorithm is the most accurate model,
terion for fit if the main purpose of the model is prediction with R2 accuracy of 0.951 and 0.925, respectively, for training
(Grace-Martin 2013). Equations 3–7 provide formulas for cal- and independent testing.
culation of metrics for model performance and accuracy evalu- The table also represents the performance of the B747-8’s
ation. most accurate σXX-Max-Top-Tens.-ANN model with an architec-
ture of 17-20-1 and a CGF training algorithm for mechanical
1
n
prediction 2
loading, and an architecture of 19-5-5-1 and a CGP training
MSE = × (yj − yjsolution ) (1)
n j=1 algorithm for simultaneous temperature and mechanical load-
√ ing. As shown, the ANN models are able to predict tensile stress
RMSE = MSE (2) with high accuracy for both training and independent testing
portions of the data, with R 2 measures of 0.978 and 0.912 for

n
prediction 2
(yjsolution − yj ) mechanical and 0.940 and 0.929 for the combined loading,
R =1−
2 j=1
(3) respectively. It can also be observed for the mechanical loading
n
2 condition that the RMSE values are 5 psi for training and 9 psi
(yjsolution −
ysolution )
j=1 for independent testing, while the range of stress varies from
about 0 to 250 psi. Under simultaneous temperature and mech-
1
n
prediction anical loading conditions, the RMSE values are 28.5 psi for
AAE = × |ysolution − yj | (4)
n j=1 j training and 31.7 psi for independent testing while the range
of stresses varies from about zero to 600 psi.
MAE = Max(|yjsolution − yj
prediction
|) (5) Table 2 describes the prediction accuracy of the A380-800’s
ANN model with the best performance for predicting σXX-Max-
where, n = Data set size; j = Case number in the data set; Top-Tens.. The most accurate model for mechanical loading was a
y solution = FEAFAA outputted critical pavement responses; 17-5-5-1 architecture and an OSS algorithm, producing accu-
y prediction = ANN models predicted critical pavement responses racy of R 2 values of 0.937 and 0.929 for training and indepen-
Since the RMSE is very popular for use in ANN model per- dent testing, respectively, and RMSE values of 12.2 psi for both
formance assessment (Gopalakrishnan et al. 2006, Gopalak- training and independent testing. For combined loading, the
rishnan et al. 2009, Chai and Draxler 2014, Ghanizadeh and best performance was achieved by a model with 19-5-5-1 archi-
Ahadi 2015) and finding the best model throughout different tecture trained by a CGB algorithm. The performance of this
algorithms and architectures, the RMSE independent testing model shows R 2 accuracy of 0.958 and 0.959 and RMSE values
value has been considered for comparing different models, of 27.7 and 29.5 psi for training and independent testing
and the best models with the smallest RMSE value with respect respectively.
to independent testing have been selected. Table 2 also displays the performance of the best σXX-Max-
Top-Tens.-prediction model, an A340-500 opt trained with a
CGP algorithm with two hidden layers, each containing 10
Accuracy evaluation results of ANN models neurons (17-10-10-1). The high accuracy achieved, R 2 = 0.970
for training and R 2 = 0.925 for independent testing for the
In the full study ANN models were developed for 156 airplanes
mechanical loading only condition, and R 2 = 0.950 for training
provided by FEAFAA, and in this paper the results for four
and R 2 = 0.943 for independent testing for the combined load-
types of heavy and new generation aircraft that exhibit high
ing condition, shows that the models are well trained and gen-
top tensile stresses are presented. Table 2 shows the training
eralised, ensuring adequate fits. The low RMSE value for all
algorithm and the architecture of the most accurate ANN
training, testing sets (lower than 9 psi for mechanical loading
model obtained for each critical response for the four types
and lower than 30 psi for combined loading) also shows the
of aircraft (B777-300 ER, B747-8, A380-800, and A340-500
negligible difference between ANN model prediction and an
opt) both under mechanical-loading-only and simultaneous
FE analysis solution.
mechanical and temperature-induced loading. This table also
shows the training and independent testing accuracy measures
Maximum top tensile stress in Y-direction
of the ANN models.
(σYY-Max-Top-Tens.)
Table 2 presents the performance of the σYY-Max-Top-Tens. pre-
Maximum top tensile stress in X-direction (σXX-Max-Top- diction ANN models for four types of aircraft, and it can be
Tens.) observed that the models are promising for adequately pre-
Table 2 shows the ANN model with an architecture of 17-15-1 dicting σYY top tensile stress in terms of R 2 and RMSE, with
and an OSS training algorithm, the most accurate model for R 2 values ranging between 0.950 and 0.977 for training and
predicting σXX-Max-Top-Tens. for the B777-300 ER with mechan- 0.905 and 0.957 for independent testing for the four aircraft
ical loading only. As can be seen in Table 2, the developed ANN under mechanical loading only. In addition, RMSE for train-
model provides high accuracy for training datasets as well as for ing and independent testing respectively ranged from 6.8 to
an independent dataset, with R 2 of 0.968 and 0.947, respect- 8.3 psi and 7.7 to 12.1 psi for mechanical loading conditions.
ively, for mechanical loading conditions. For a combined load- Furthermore, based on AAE, the difference between ANN
ing condition (i.e. simultaneous temperature and mechanical model prediction and the NIKE3D FE solution varies from
10 A. REZAEI TARAHOMI ET AL.

Table 2. ANN models’ training and independent testing results.

Top tensile stresses
Simultaneous temperature and mechanical
Algorithm, architecture, accuracy Only mechanical loading case loading case
Aircraft criteria σXX-Max a σYY-Max b σ1-Max c σXX-Max a σYY-Max b σ1-Max c
B777-300 ER Training algorithm OSSd OSSd CGBe CGFf OSSa CGPg
ANN architecture 17-15-1 17-10-1 17-15-1 19-5-5-1 19-5-5-1 19-5-5-1
Training AAE* 4.7 4.9 6.4 21 20.3 23.4
MAE* 50.0 50.9 66.2 113.1 158.1 233.3
RMSE* 6.4 6.8 9.1 27.7 27.9 32.7
R2 0.968 0.963 0.946 0.951 0.950 0.941
Independent testing AAE* 6.1 5.3 7.3 24.3 25.2 28.6
MAE* 32.3 45.3 33.7 120.6 140.7 160.7
RMSE* 8.3 7.7 9.9 33.2 34.6 40.9
R2 0.947 0.957 0.953 0.925 0.916 0.919
B747-8 Training algorithm CGFf CGBd CGPg CGPg OSSd CGBe
ANN architecture 17-20-1 17-10-1 17-45-1 19-5-5-1 19-5-5-1 19-5-5-1
Training AAE* 4.1 6.0 8.6 20.7 20.1 20.7
MAE* 23.4 70.7 57.5 127.9 119.8 143.3
RMSE* 5.3 8.3 11.8 28.5 27.5 29.2
R2 0.978 0.950 0.912 0.940 0.943 0.943
Independent testing AAE* 6.4 6.7 9.7 22.6 22.7 23.9
MAE* 29.7 36.1 56.0 142.0 105.1 136.2
RMSE* 9.1 8.9 12.6 31.7 30.7 33.3
R2 0.912 0.907 0.872 0.929 0.930 0.926
A380-800 Training algorithm OSSd CGBe CGBe CGBe LMh OSSd
ANN architecture 17-5-5-1 17-15-15-1 17-15-1 19-5-5-1 19-5-5-1 19-5-5-1
Training AAE* 8.8 6.0 8.4 20.6 20.8 22.3
MAE* 66.8 36.5 51.5 130.3 112.0 140.3
RMSE* 12.2 7.8 11.5 27.7 28.0 30.1
R2 0.937 0.977 0.956 0.958 0.956 0.953
independent testing AAE* 8.6 8.7 10.0 21.9 22.6 24.7
MAE* 53.1 63.0 48.6 126.1 174.2 126.8
RMSE* 12.2 12.1 13.3 29.5 31.7 33.4
R2 0.929 0.930 0.930 0.959 0.949 0.947
A340-500 opt Training algorithm CGPg CGFf CGBe CGFf LMh CGFf
ANN architecture 17-10-10-1 17-10-1 17-5-5-1 19-5-5-1 19-5-5-1 19-5-5-1
Training AAE* 4.2 5.1 5.7 19.2 19.2 19.6
MAE* 38.5 55.1 52.7 131.7 115.3 145.9
RMSE* 5.6 7.0 8.3 26.5 26.0 27.8
R2 0.970 0.954 0.941 0.950 0.950 0.949
Independent testing AAE* 5.2 6.4 6.5 21.3 21.3 23.0
MAE* 57.1 41.5 34.2 94.6 103.8 125.8
RMSE* 8.5 8.9 9.6 29.2 28.7 32.2
R2 0.925 0.905 0.910 0.943 0.948 0.935
a
σXX-Max-Top-Tens.
b
σYY-Max-Top-Tens.
c
σ1-Max-Top-Tens.
d
One-step Secant.
e
Powell-Beale Restarts.
f
Fletcher-Reeves Update, (cgp).
g
Polak-Ribiére Update.
h
Levenberg-Marquardt.
*The unit is (psi).

4.9 to 6 psi for training and 5.3to 8.7 psi for independent test- to top-down cracking in concrete pavements using ANN
ing. For predicting σYY-Max-Top-Tens. under combined loading models.
conditions, the R2 value ranged from 0.943 to 0.956 for train- Table 2 shows that, for mechanical loading only, the ANN
ing and 0.916 to 0.948 for independent testing. The results for model accuracy in terms of R2 varies from 0.912–0.956 for
the combined loading condition also demonstrate that the training and 0.872–0.953 for independent testing, and for sim-
ANN model can on average predict σYY-Max-Top-Tens. with ultaneous temperature and mechanical loading the accuracy
only a 28.7–34.6 psi difference from the NIKE3D FE solution ranged from 0.941–0.953 for training and 0.919–0.947 for inde-
as the stress varies from 0 to 600 psi. pendent testing. These results demonstrate that the ANN
models for B777-300 ER, B747-8, A380-800, and A340-500
opt can predict σ1-Max-Top-Tens. with only a 6.5–10.0 psi error
Maximum principal tensile stress (σ1-Max-Top-Tens.) on average when pavement is subjected to mechanical loading
Table 2 also shows that the σ1-Max-Top-Tens.-prediction ANN only and with only a 23–28.6 psi error when pavement is sub-
models developed for different types of aircraft can accurately jected to simultaneous temperature and mechanical loading.
replicate the NIKE3D-FAA FE solutions. This can help to ANN models were found to successfully replicate FEA-
obtain one of the most important responses corresponding FAA/NIKE3D-FAA pavement response solutions for all four
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 11

types of airplanes. It is also important to note that test sets Inputsnor includes the normalised inputs value obtained by
produced high accuracy similar to that of the training set Equation (8):
for all pavement response types, proving ANN models’ gener-
(Xi − Xmin
i
)
alisation (i.e. they did not memorise the relationship), and Normalised Xi = 2 × −1 (8)
showing them to be robust and valid. Comparing the results (Xmax − Xmin )
i i

for all three responses shows that ANN models predict top i
where Xi is the actual value of the input i, Xmin is the minimum
tensile stress responses accurately, so ANN models should i
of the input i in the dataset, and Xmax is the maximum of the
be able to greatly help in analysing top-down cracks by pre- input i in the dataset. The matrix size of IW (S1 × R) is deter-
dicting the most critical associated stress responses. The mined by number of elements R in the Inputsnor and the num-
most accurate ANN models presented in this section are ber of neurons in the first hidden layer (S1), and the size of the
used in the new rapid rigid airfield pavement analysis tool weight matrices are determined by the number of neurons in
described in the next section. each layer (S1 and S2). For a one-hidden-layer ANN architec-
Besides, the authors studied the effects of different input par- ture the LW1 size is 1 × S1, while for a two-hidden-layer archi-
ameters on the responses (Rezaei-Tarahomi et al. 2019). The tecture the size of LW1 is S1 × S2 and LW2 is 1 × S2. The bias
objective of the study was validating the developed ANN matrix size for B1 is S × 1 and for B2 is 1 × 1 for one hidden
models using sensitivity analysis, by both NIKE3D-FAA sol- layer. For an architecture with two hidden layers the sizes of
utions and those predicted by ANN models, of the critical ten- B1 and B2 are S1 × 1 and S2 × 1, respectively, and B3 is 1 × 1
sile stresses on top of rigid airfield pavements. Also, the study in this study.
helped to improve the models and using the most accurate Table 3 shows the execution time for ANNFAA to predict
ones in developing the ANN-based analysing software called one critical response. As shown in the table, for 1 million
ANNFAA which is explained in the next section. cases, it takes 18.8 s to obtain the inputs, weights, and bias
matrices from text files, and 6.3s to compute the critical
responses (σXX-Max-Top-Tens., σYY- Max-Top-Tens., or σ1- Max-Top-
Tens.) for all cases of a one-hidden-layer ANN model, while
Artificial neural networks based federal aviation the same process for a two-hidden-layer ANN model takes
administration rigid pavement analysis tool: 18.4 and 15.9 s, respectively, to import data and compute the
ANNFAA critical responses.
A new software tool (ANNFAA) for replicating maximum In the ANNFAA program the matrix multiplication is per-
stress responses on top of slabs in rigid airfield pavements formed in parallel so the full processor capacity is utilised. Such
has been developed. ANNFAA was developed using the C# parallel computation significantly increases program execution
programming language. This tool is based on the ANN models speed and decreases run time to less than one-third that of a
trained for critical responses for all types of airplanes provided non-parallel program execution.
in the FEAFAA 2.0 software. Figure 8 shows the user interface
of the program. Working with the program is similar to using
FEAFAA, although in the new program creating a mesh and Conclusions
running NIKE3D_FAA are not required. The same inputs as
The main objective of this study was to create a process for
in FEAFAA 2.0 are defined for the ANNFAA (Figure 8), and
developing and using critical stress response prediction models
a set of inputs can be imported into ANNFAA. Results are pre-
for multiple slab-rigid airfield pavements and to generate soft-
dicted for all input sets at once. ANNFAA uses weights and
ware for rapid analysis of the rigid pavements. A batch run
biases matrices exported from the neural networks trained in
automation tool for implementing 3D-finite element (3D-FE)
MATLAB. Equations 6 shows the calculation used within the
simulation and post-processing the outputs has been created.
program for predicting the critical stresses. Weight matrices
An extensive database of input-output records from
(IW and LW) and biases (B1 and B2) are obtained from the
NIKE3D-FAA FE analysis have been obtained for each aircraft
optimal ANN models for each aircraft. Equations 6 is used
and for each loading condition. Moreover, an artificial neural
by ANNFAA if the optimum ANN model for the aircraft and
network (ANN) model-development tool was created in
response of interest has one hidden layer, while if the model
MATLAB to facilitate the process of ANN model development
has two hidden layers Equations 7 is used.
and enhance the speed of importing data, training models,
Y = B2 + LW1 × tansig(B1 + IW × Inputsnor) (6) and exporting and determining the optimum models. Accu-
racies of the ANN critical response prediction models for (a)
Y = B3 + LW 2 σXX-Max-Top-Tens., (b) σYY-Max-Top-Tens., and (c) σ1-Max-Top-Tens.
× tansig((LW 1 × tansig(B1 + IW × Inputsnor)) + B2 ) responses for loading applied by four types of airplanes:
B777-300 ER, B747-8, A380-800, and A340-500opt, have
(7)
been presented in this study. Furthermore, a new tool, ANN
where B1, B2, and B3 = Biases matrices; LW1 = Weight matrix based FAA rigid pavement analysis tool (ANNFAA), has
(layer 1); LW2 = Weight matrix (layer 2); IW = Input weight been developed using C# programming to replicate maximum
matrix; tansig = Hyperbolic tangent sigmoid transfer function stress responses at slab tops of rigid airfield pavements for all
(see Figure 9); Inputsnor = Matrix of normalised inputs types of aircraft provided by Finite Element Analysis FAA
between (−1, 1); Y = Network output of interest. (FEAFAA) version 2.0 software that makes use of
12 A. REZAEI TARAHOMI ET AL.

Figure 8. ANNFAA user interface (a) input and (b) prediction tabs.

NIKE3D_FAA as a stand-alone tool for 3D-FE analysis of mul- each case, automatically. It also facilitated and handled cre-
tiple-slab rigid airport pavements and overlays. The major con- ation of a synthetic database consisting of FEAFAA input par-
clusions of this study are: ameters and associated critical pavement responses.
. The program provided a very easy-to-use platform for post-
. An automation program was developed to perform FEAFAA, processing the data and gathering all the data required for
save input files, run NIKE3D, and save the output files for further ANN development.
 INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 13

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.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Funding
This work was supported by Federal Aviation Administration [grant num-
ber15-G-01].
Figure 9. Hyperbolic tangent sigmoid transfer function.

ORCID
Table 3. ANNFAA computation run times for different data set size. Adel Rezaei Tarahomi http://orcid.org/0000-0001-5405-896X
Run time (s)
One hidden Two hidden
References
Importing Importing
Data set inputs, weights inputs, weights Beale, M. H., Hagan, M. T., and Demuth, H. B., 2017. Neural network tool-
size and biases Prediction and biases Prediction box user’s guide. Natick, MA: The MathWorks, Inc..
(Cases) matrices computation matrices computation Bennett, J., 2017. AutoIt. https://www.autoitscript.com/site/autoit-script-
5000 0.08 0.10 0.38 0.15 editor.
500,000 9.43 3.43 9.36 7.71 Brill, D. R., 1998. Development of advanced computational models for air-
1,000,000 18.84 6.33 18.37 15.89 port pavement design. Springfield, VA: National Technical Information
Service. Rep. No. DOT/FAA/AR-97.
. Use of the automation program significantly increased the Brill, D. R., 2000. Field verification of a 3D finite element rigid airport pave-
ment model. Springfield, VA: National Technical Information Service.
efficiency of the batch run execution by decreasing the num- Rep. No. DOT/FAA/AR-00/33.
ber of mistakes and time required for performing the finite- Ceylan, H., 2002. Analysis and design of concrete pavement systems using
element analysis for 156 different types of airplanes. artificial neural networks. Urbana, IL: University of Illinois at
. The MATLAB ANN model development tool supported Urbana-Champaign.
development of a total of 6,000 ANN models for mechanical Ceylan, H., et al., 2013. Sensitivity quantification of jointed plain concrete
pavement mechanistic-empirical performance predictions.
and simultaneous mechanical and temperature loading con- Construction and Building Materials, 43, 545–556.
ditions for each type of aircraft. Ceylan, H., Bayrak, M. B., and Gopalakrishnan, K., 2014. Neural networks
. The rapid rigid airfield pavement analysis tool ANNFAA applications in pavement engineering: a recent survey. International
can predict critical tensile stresses for 1 million cases in 6– Journal of Pavement Research and Technology, 7 (6), 434–444.
16 s, while executing the same cases by NIKE3D-FAA Ceylan, H., and Gopalakrishnan, K., 2007. Neural networks based models
for mechanistic-empirical design of rubblized concrete pavements.
would take several weeks. Geotechnical Special Publication No. 169, Soil and Material Inputs for
. Transforming all the ANN model data from MATLAB Mechanistic-Empirical Pavement Design, ASCE, 1–10. https://doi.org/
workspaces into text files and making them ready to for 10.1061/40913(232)5.
use in an executive program like ANNFAA is a very useful, Ceylan, H., Tutumluer, E., and Barenberg, E., 1999. Artificial neural net-
fast, and easy method for utilising the developed ANN works for analyzing concrete airfield pavements serving the Boeing B-
777 aircraft. Transportation Research Record: Journal of the
models for all aircraft. Transportation Research Board, 1684 (99), 110–117. https://doi.org/
. The ANNFAA tool can be used as a tool for testing ANN 10.3141/1684-13.
models during further investigation and model development. Chai, T., and Draxler, R. R., 2014. Root mean square error (RMSE) or
. Critical response surrogate models are very useful alterna- mean absolute error (MAE)? – Arguments against avoiding RMSE in
tives for use in design iterative processes, making routine the literature. Geoscientific Model Development, 7 (3), 1247–1250.
https://doi.org/10.5194/gmd-7-1247-2014.
design more efficient and practical while taking into account Elbagalati, O., et al., 2017. Development of an artificial neural network
more failure modes during design. model to predict subgrade resilient modulus from continuous deflection
. Since the models developed in this study are for rigid pave- testing. Canadian Journal of Civil Engineering, 44 (9), 700–706. https://
ments consisting of nine multiple slabs, the critical responses doi.org/10.1139/cjce-2017-0132.
of the pavement systems with some other number of slabs Ghanizadeh, A. R. and Ahadi, M. R., 2015. Application of artificial neural
networks for analysis of flexible pavements under static loading of stan-
might not be correctly predicted by the models presented dard axle. International Journal of Transportation Engineering, 3 (1),
in this study. 31–43. https://doi.org/10.22119/ijte.2015.13361.
Ghasemi, P., et al., 2019a. Predicting dynamic modulus of asphalt mixture
using data obtained from indirect tension mode of testing. Chemical
and Biological Engineering Publications, 375, .https://lib.dr.iastate.edu/
Acknowledgements cbe_pubs/375.
Ghasemi, P., et al., 2019b. Principal component neural networks for
The authors gratefully acknowledge the Federal Aviation Administration modeling, prediction, and optimization of hot mix asphalt
(FAA) for supporting this study under grant number15-G-01. The con- dynamics modulus. Infrastructures, 4 (3), 53. https://doi.org/10.3390/
tents of this paper of this paper reflect the views of the authors, who are infrastructures4030053.
14 A. REZAEI TARAHOMI ET AL.

Gopalakrishnan, K, 2010. Effect of training algorithms on neural networks International Journal of Environmental Science and Technology, 16
aided pavement diagnosis. International Journal of Engineering, Science (10), 5835–5856.
and Technology, 2 (2), 83–92. https://doi.org/10.4314/ijest.v2i2.59147. Rezaei-Tarahomi, A., et al., 2017a. Neural networks based prediction of
Gopalakrishnan, K., Ceylan, H., and Attoh-Okine, N. O., 2009. Intelligent critical responses related to top-down and bottom-up cracking in airfi-
and soft computing in infrastructure systems engineering: recent eld concrete pavements. Tenth International Conference on the Bearing
advances. New York City, NY: Springer. Capacity of Roads, Railways and Airfields.
Gopalakrishnan, K., Thompson, M. R., and Manik, A., 2006. Rapid finite- Rezaei-Tarahomi, A., et al., 2017b. Sensitivity quantification of airport
element based airport pavement moduli solutions using neural net- concrete pavement stress responses associated with top-down and
works. Computational Intelligence, 3 (1), 63–71. bottom-up cracking. International Journal of Pavement Research
Grace-Martin, K., 2013. Assessing the fit of the model. 153–225. http://doi. and Technology, 10 (5), 410–420. https://doi.org/10.1016/j.ijprt.2017.
wiley.com/10.1002/9781118548387.ch5. 07.001.
Hagan, M. T., et al., 2014. Neural network design. Boston, MA: PWS Pub. Rezaei-Tarahomi, A., et al., 2018. Neural network algorithms for rigid airfi-
Hamm, L., Brorsen, B. W., and Hagan, M. T., 2007. Comparison of sto- eld pavement responses. Doha: International Conference on Advances in
chastic global optimization methods to estimate neural network Materials and Pavement Performance Prediction.
weights. Neural Processing Letters, 26 (3), 145–158. Rezaei-Tarahomi, A., et al., 2019. Artificial neural network models for
Kaya, O., et al., 2018. Neural network–based multiple-slab response models airport rigid pavement top-down critical stress predictions: sensi-
for top-down cracking mode in airfield pavement design. Journal of tivity evaluation. airfield and highway pavements 2019: Innovation
Transportation Engineering, Part B: Pavements, 144 (2), 4018009. and sustainability in highway and airfield pavement technology –
Ling, M., et al., 2017. Numerical modeling and artificial neural network for selected papers from the international airfield and highway pave-
predicting j-integral of top-down cracking in asphalt pavement. ments conference 2019, July, 302–312. https://doi.org/10.1061/
Transportation Research Record, 2631, 83–95. https://doi.org/10.3141/ 9780784482476.030.
2631-10. Salehi, H., and Burgueño, R., 2018. Emerging artificial intelligence
Rastegaripour, F., et al., 2019. Simultaneous management of water and methods in structural engineering. Engineering Structures, 171, 170–
wastewater using ant and artificial neural network (ANN) algorithms. 189. https://doi.org/10.1016/j.engstruct.2018.05.084.