Journal of Microscopy, Vol. 254, Issue 1 2014, pp. 31–41

doi: 10.1111/jmi.12113

Received 25 September 2013; accepted 7 January 2014

Neural network modelling of asphalt adhesion determined by AFM R.A. TAREFDER & S. AHSAN Department of Civil Engineering, University of New Mexico, Albuquerque, New Mexico, U.S.A.

Key words. Adhesion, atomic force microscopy, neural network, polymer and lime.

Summary This study constructs a neural network (NN) model to quantify adhesion from atomic force microscopy (AFM) data. AFM data contain five-point force–distance values. A total of 760 observations are used to build NN model. To train the network, AFM tip-sample distance data, percentage of lime, type and percentage of polymer and asphalt chemical functional groups are given as inputs and AFM force as an output. To select the NN architecture, one and two hidden layers with varying neurons are tried with 10 input nodes in the input layer and 5 output nodes in the output layer. Two hidden layers with 9 and 17 nodes in the first and second layer, respectively, show the best performance. A 10-9-17-5 NN is selected as the final structure of the NN model. Test results for the trained model show good prediction ability. The model is further applied to evaluate the effect of five different percentages of lime on the adhesion of asphalt. Results show that increase in the percentage of lime is very effective at reducing moisture damage in a styrene butadiene polymer modified asphalt sample. However, increase in lime percentage above 1.5% does not help reduce moisture damage in the styrene butadiene styrene polymer modified sample.

Introduction Adhesion of asphalt is an important parameter in quantifying moisture damage of asphalt. Moisture damage can be quantified comparing adhesion force of wet and dry asphalt samples. However, it is not possible to determine adhesion force from bond chemistry, as asphalt chemical structure is not fully explored till now. Therefore, an atomic force microscopy (AFM) laboratory test was performed to determine adhesion of asphalt (Arifuzzaman, 2010). Development of a predictive model that uses AFM laboratory experiment data is useful in determining the adhesion force for the values of test factors that were Correspondence to: Rafiqul A. Tarefder, Department of Civil Engineering, Associate Professor and Regents’ Lecturer, University of New Mexico, MSC01-1070, Albuquerque, NM 87131-0001, U.S.A. Tel: 505-277-6083; fax: 505-277-1988; e-mail: [email protected]

 C 2014 The Authors C 2014 Royal Microscopical Society Journal of Microscopy 

not incorporated in the AFM laboratory tests. The test factors and outputs of the AFM laboratory tests have nonlinear and complex relationships. A neural network (NN) has the ability to recognize and trace the complex relationship trend existing between inputs and outputs (Kim et al., 2004; Xiao & Amirkhanian, 2009a). Also, traditional modelling techniques require some assumptions, but NN does not require any assumption for model development. Therefore, in this study, an NN is chosen to develop the predictive model using AFM laboratory test data. In the past 15 years, there has been an increased interest in NNs, a computational intelligence system, in pavement system applications. Past studies revealed that there have been several successful studies that incorporated NNs to predict the pavement structural parameters such as pavement moduli, pavement layer thickness, etc. using falling weight deflectometer (FWD) deflection data (Meier & Rix, 1995; Williams & Gucunski, 1995; Meier et al., 1997; Gucunski et al., 1998; Kim & Kim 1998; Lee et al., 1998; Saltan & Terzi, 2004; Ceylan et al., 2008). The researchers concluded NNs to be more efficient and a better technique compared to conventional and traditional tools in this regard. In addition, an NN proved to a better prediction tool in predicting low temperature performance of modified asphalt mixtures in comparison to a general linear model (Tasdemir, 2009) due to its versatile and complex pattern tracing and computational capabilities. Different state departments of transportation (DOT) developed models on the basis of NNs to compute pavement-related parameters. For example, Illinois DOT successfully used NNs with low average error in comparison to ILLI-PAVE analysis in estimating pavement moduli, maximum stress and strain from FWD data (Ceylan et al., 2004). Texas DOT also developed a methodology to calculate remaining life of flexible pavement with the help of NN (Ferregut et al., 1999). NNs are not only used in predicting pavement structural parameters, they are also used in determining pavement performance parameters by researchers. Owusu-Ababio (1998) developed an NN-based model to predict cracking performance of pavement. Pavement cracking depends on many parameters. Though cracking maintained complex and critical relationships with the parameters, NN was able to pick the

32

R.A. TAREFDER AND S. AHSAN

trend of the relationship and predict cracking on the basis of those parameters. Yang et al. (2003) applied an NN in predicting the pavement crack index and pavement condition rating, which was very difficult to forecast using traditional techniques and tools. Alsugair & Al-Qudrah (1998) and Huang & Moore (1997) utilized an NN technique to develop a module that suggested suitable maintenance strategies on the basis of distress data and corresponding maintenance offered for pavement. From the laboratory test data for bituminous materials, NN was used to predict dynamic modulus of hot mix asphalt (Ceylan et al., 2008) and fatigue life (Huang et al., 2007). Xiao & Amirkhanian (2009b) involved NN to predict stiffness behaviour of rubberized asphalt concrete mixtures with reclaimed asphalt pavement. Tarefder et al. (2005a) used NN to determine asphalt rutting. They also determined permeability from asphalt mix properties using NN. All NN models showed satisfactory performance, better efficiency compared to traditional regression models and so, were being accepted by researchers. All of the past studies imply that an NN is capable of tracing the two complicated relationships existing between the input and output parameters of a model. However, an NN has not been used extensively in determining pavement material characterization. The AFM laboratory data used in this study have an intricate relationship between the test parameters and adhesion value of asphalt. Therefore, an NN model is chosen to develop a model that will be able to predict adhesion of asphalt on the basis of test factors provided to the model. Objective

Fig. 1. A generalized NN structure.

receive information and process to obtain output. Input layer has nodes but they are passive. It means, the nodes in input layer only receive the input information and pass information without modification to nodes in intermediate layer through connection weights. Each node in intermediate layer contains a summation function. A summation function accumulates the multiplication of the input value (xi , where i = 1,2, . . . n) and weight (wi , where i = 1,2, . . . n) of the connection through which it is fed to the node. The node passes the summed value of inputs through an activation function such as sigmoid, pure linear, etc., to compute the output:  Summation function, = x1 w1 + x2 w2 + x3 w3 + · · · · · · · · · + xn wn (1)

The objectives of this study is to

r r

Develop an NN model that maps the relationship between the adhesion force of asphalt and testing factors or variables involved in the nano-scale evaluation of asphalt adhesion by the AFM laboratory test. Use the developed model to predict asphalt adhesion for five percentages of lime keeping all other input factors as constant. AFM laboratory test was done for 0.5%, 1.0% and 1.5% of lime. Two extrapolated percentages (2.0 and 2.5) of lime are added to the model to investigate the effect of varying percentage of lime on the adhesion force of asphalt.

Basics of NN An NN is a nonlinear computational model, which contains nodes that perform mathematical functions. An NN contains inputs, outputs and intermediate (also known as hidden) layers with a set of nodes and connection between the nodes. Nodes can be considered as computational units. A function of the computation unit in NN is illustrated in Figure 1. Nodes

Activation function, ϕ



=

1  . 1 + e −( )

(2)

The summation function of a node (see Eq. 1) in a layer depends on the connection weights and inputs of the previous layer. In addition, bias term can be added to the summation function. Bias is used to create hyperplanes to separate the input dimensions. Without bias, all hyperplanes are constrained to go through the origin of the hyperspace defined by the inputs. Output of a node depends on the activation function applied on the summation function. If more than one set of the intermediate or hidden layer with neurons is used, then the activation output of the first hidden layer is used as inputs for the second hidden layer’s nodes. The same calculation is done in each node of the single hidden layer until it reaches the output layer:    ∼ ϕ ϕ . (3) = ϕ2

 C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

NEURAL NETWORK MODELLING

33

Fig. 2. Sample force–distance curve for AFM laboratory test.

The sigmoid function is nonlinear. Therefore, addition of each hidden layer gives a higher order of nonlinearity to the activation output as nonlinear activation output of previous layer is transmitted through nonlinear activation function of following layer (see Eq. 3). Thus, this computational model is very suitable to establish a prediction model by addressing a nonlinear higher order polynomial relationship between input and output parameters. AFM data Basics of AFM testing AFM is a scanning probe technique that measures force as a function of distance between two molecules or atoms. It uses a laser beam deflection system along with a probe and sample. A probe is a sharp tip placed on a cantilever. Laser is reflected from the reflective surface of the cantilever and onto a position sensitive detector. Using this system, beam position can be determined. When a tip with the cantilever comes in close distance to a sample surface, an attraction or repulsion force occurs in between the surface and probe depending on the distance between them. This force is not directly measured by AFM, rather the deflection of the cantilever due to the attractive or repulsive force is recorded. Knowing the stiffness of the cantilever, force is determined from Hooke’s law as follows: F = − kz,

(4)

where F is the force, k is the stiffness of the lever and z is the distance the lever is bent. Figure 2 shows schematic five-point force–distance curve that can be attained from the AFM test. It can be seen that, initially, when the tip is far away from the sample’s surface, there is no attraction or repulsion between the tip and sur-

 C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

face. Then, the tip approaches to a distance when at first the tip jumps onto the sample’s surface due to some attraction (position B). Pushing the tip further toward the sample’s surface then produces a strong repulsive force (position C). Now, retracting the tip from the sample’s surface decreases the repulsion and increases the attraction force. On the retraction path, there is a distance beyond which the attraction tends to decrease. A ‘pull off’ force is needed at position D to completely remove the tip from the influence of the strong attraction force. The force value at position D is commonly regarded as the adhesion or cohesion force of that particular sample. Finally, the tip is fully retracted from the force influence zone at position E.

Tip functionalization Asphalt is a mixture of highly condensed polycyclic aromatic hydrocarbons. Till present, asphalt chemical structure is not fully explored. However, presence of some compounds in asphalt has been detected such as carboxyl (–COOH), methyl (–CH3 ), ammine (–NH3 ) and hydroxyl (–OH) groups (Testa, 1995). AFM tips were functionalized with these functional groups to measure cohesion within the asphalt. Another tip, made of silicon nitride (Si3 N4 ), is used to probe the asphalt surface to measure adhesion between asphalt and aggregate. Adhesion is a general term, force between two molecules. If the two molecules are of similar type, which is the case of when an asphalt functionalized tip molecule is probing an asphalt molecule, one could term adhesion as cohesion. Though a certain functionalized tip is used to measure adhesion of the asphalt, it is not known which molecule of asphalt sample the tip is interacting with. Therefore, the term adhesion is used all through the paper instead of using adhesion and cohesion terms separately.

34

R.A. TAREFDER AND S. AHSAN

Sample preparation Polymer modification on asphalt is commonly done to improve the resistance of rutting, fatigue and thermal cracking of asphalt pavements. Two elastomer types of polymers are used to modify asphalt sample used in AFM test. They are styrene butadiene styrene (SBS) and styrene butadiene (SB). SB and SBS type polymers have emerged as effective polymers to be used in asphalt with optimum performance, reliability, ease of use and economy. Elastomer modification of binder reduces permanent deformation. In addition, modified binder becomes less susceptible to temperature and more resistive to thermal cracking and rutting. Also, polymer-modified asphalt shows better performance under moisture effect in past literatures. Polymers effect viscosity of asphalt in a way that improves the adhesion ability of asphalt binder to aggregate surface. Therefore, asphalt sample modified with 3% and 5% of SB and SBS modified samples are used to determine bond force of asphalt in AFM test. These polymers are suggested to be used by 3–5% in asphalt. Therefore, for this study, 3% and 5% polymer modification are considered. Dry and wet conditioned asphalt samples are used to determine asphalt adhesion by AFM testing. At first AFM test samples are prepared for polymer modified and lime added asphalt in regular procedure. This sample is considered as dry sample. Then, those dry samples are vacuum saturated for half an hour and soaked under 3-inch depth de-ionized water for 72 h. Before AFM testing, those samples are removed from water and dried overnight inside a draft oven at 40°C to reduce the surface wettability and probable capillary action may cause by surface water. Then this sample is ready to be tested and considered at wet sample. To maintain the same surface environment, dry samples are also put in oven overnight at the mentioned temperature.

Laboratory AFM data processing An AFM test is conducted on an asphalt sample using five different types of AFM tips to probe the sample. A force–distance curve attained from the AFM test is shown in Figure 3. The figure is for an asphalt sample probed with a tip modified with the –OH functional group. Figure 3 has the same characteristic as the schematic force–distance curve shown in Figure 2. Indeed, force and distance values are tabulated corresponding to the points A, B, C, D, E in Figures 2 and 3. From the force–distance curve, the force value corresponding to point D is considered as the adhesion or cohesion force. On the basis of what type of tip is used to probe the asphalt surface, the force can be named as adhesion or cohesion.

NN data Laboratory data are collected from the AFM test on asphalt to develop an NN model. Laboratory test data contain five-

Fig. 3. Force–distance curve obtained for asphalt sample probed with –OH tip by AFM laboratory test.

point force–distance measurements for different types of asphalt samples. To observe the effect of moisture in the adhesion and cohesion force of asphalt, asphalt samples were categorized as wet and dry samples. For dry and wet condition case, base asphalt with no polymer and asphalt modified with two types of polymer named SB and SBS were prepared. The percentage of polymer varied from 3% to 5% for both the polymer types. In addition, a lime anti-stripping agent was used in all the asphalt samples with percentage varied from 0.5 to 1.5 at an increment of 0.5%. For each combination of asphalt sample, the AFM test was conducted using five different tips mentioned previously in the tip functionalization section. Table 1 shows the variables of AFM testing on the asphalt samples. The variables are separated into two groups: input and output. Variables that fall under the input category are selected as inputs and five forces corresponding to five distances are selected as outputs to construct a supervised NN model. In Table 1, subscripts of X and F denote the positions of the probe referring to Figure 3. Selection of NN algorithm A multilayer feed-forward NN was employed with error back propagation algorithm. Multilayer feed-forward network has been applied successfully to map nonlinear relationships between input–output variables in a supervised manner using backpropagation algorithm (BPA) (Saltan & Terzi, 2004; Tasdemir, 2009). Association of BPA with the feed-forward network helps the network to learn from the error and apply error minimization technique to determine output from input information. The error is calculated by comparing the network output and actual output. There are many different BPAs available for error minimization. In this study, error minimization is done using Levenberg–Marquartz algorithm in MATLAB. This technique is fast and requires less memory use in  C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

NEURAL NETWORK MODELLING

35

Table 1. Input and output variables for NN model. Inputs Moisture condition Wet Dry

Polymer type

Polymer %

Antistripping agent %

Tip type

Five-point distances

Outputs Force corresponding to distances

None SB SBS

0 3 4 5

0.5 1.0 1.5

Silicon Ammine Methy1 Carboxy1 Hydroxy1

X1 X2 X3 X4 X5

F1 F2 F3 F4 F5

analysis. Also, Gopalakrishnan (2010) showed this algorithm performs better in developing prediction model in comparison with other error minimization techniques.

(6)

where n = total number of data sets, o = network output, p = number of outputs, t = target output and o mean =average of network output.

Selection of NN architecture In this study, a multilayer feed-forward backpropagation NN is developed to predict lime-treated asphalt adhesion and cohesion force from AFM laboratory test data. The inputs and outputs of the model are shown in Table 1. A total of 760 observations are used to construct the model. The data set is divided into three sets: training, validation and testing set. The training set is used to train the network to trace the nonlinear and complex relationship between input and output. The validation set is used to decide the stopping criteria and estimate network performance. The stopping criterion helps prevent the network from overfitting. The testing set is used to check prediction ability of the network. A fully connected feed-forward NN is chosen for this study. There were 10 test parameters that were incorporated in the AFM laboratory test on the determination of adhesion force of the asphalt. These are provided as inputs for the NN model. The output of the model is set as five-point adhesion forces for five distances. Thus, an NN with 10 input neurons and 5 output neurons is developed. Like the neurons in our brain, input neurons receive input data provided from outside the model environment and transmit the information to the hidden layers. The hidden layer neurons perform calculations necessary to process the data and extract features to mapping output from the input space. Outputs produced from the NN model are stored in the output layer neurons. The basic concern is to choose a number of neurons in the hidden layer and also how many hidden layers are to be incorporated in the model structure. There is no general rule to selecting the number of neurons in the hidden layer. A number of models are run for the single hidden layer with 4– 20 neurons and two hidden layers each with 4–20 neurons. Then the performance of the network is determined by MSE and R2 expressed as follows: p  n  2  o i j − ti j

Mean Squared Error, MSE =

 (o−t)2 , Goodness of fit, R =  (o−o mean )2 2

j =1 i =1

n. p

 C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

,

(5)

Training and validation of NN The total data set is split into 90% and 10% for training and testing, respectively. The training data set is divided into two sets: training and validation. Initially, the network is trained using the training data, and calculations for MSE for the validation data simultaneously to check performance of the trained network. Training is done for 10 000 epochs to determine the number of hidden layers and number of nodes in each hidden layer. An epoch is a measure of the number of the passes all the training vectors are used once to update the weights. While training, overfitting occurs after a certain point during training when the network is trained to get as low error as possible. Due to the overfitting, network establishes intermediate points between two observations which disrupt NN in generalization. There are various techniques available to avoid overfitting: early stopping is one of the widely used among the techniques. In this study, two sets of data are used during training: training and validation set. Training set is used to update weights and validation set is used to calculate the error using the updated network. If the error calculated by validation set decreases, training continues on. If, error of validation set starts increasing, then the training is stopped. Also, for some further epochs, training can be continued to observe whether the validation set error starts decreasing or continue on increasing. For those specific epochs, if the validation error still continues on increasing, then the training is stopped, otherwise, training is continued on. Training of NN model with training data is carried out in two phases. In the first phase, number of hidden layer and number of nodes in each layer is determined. NN models with one and two hidden layers with 4–20 nodes in each layer are tried. Initially, random NN structure is tried with randomly generated connection weights. With each epoch, connection weights are updated to make prediction result closer to

36

R.A. TAREFDER AND S. AHSAN

network architecture of NN model is depicted in Figure 5. The inputs and outputs of the model are also shown on the same figure. After selection of the NN architecture, the 10-9-17-5 network is again trained but now for 10 000 randomly generated connection weights. From the MSE and R2 performance of validation and testing set, respectively, the 100 best performed network connection weights are stored. These weights are further fed to the network and used in prediction. The maximum frequent value of the outputs of the prediction models for 100 connection weights is taken as predicted output. Thus, NN model is developed and ready to be used in predicted as the NN structure as well as the connection weights are known. Testing the NN The test data are introduced to the established network to determine the adhesion value of asphalt at five different distances. The network is simulated for 100 test runs for 100 connection weight sets. The network provides an average MSE value of 1.534E-15. The model shows good prediction ability with a regression coefficient of 0.971.

Fig. 4. Performance plot at training of NN. Table 2. Sample results for selection of NN structure. No. of neurons in the first hidden layer 9 9 9 9 9

No. of neurons in the second hidden layer

Validation MSE data

Test data R2

15 16 17 18 19

1.21E-15 1.27E-15 1.18E-15 1.22E-15 1.26E-15

0.980119 0.978471 0.979596 0.980211 0.979014

actual result. From those analyses, an initial idea of connection weight range is gathered. That range is further used to generate 100 sets of random weights for NN structure selection. In addition, training data are also randomized 75 times for better training. Each trial NN model is run for 100 weights and for each weight, data are randomized 75 times. Thus, each trial NN is run for 7500 times and maximum likelihood value of the performance parameters for all 7500 runs are taken as final performance of the trial NN. In this way, all trail NN model combinations for one and two hidden layers having varied NN are run and performance parameters are stored for each case. On the basis of performance parameters, final NN architecture is selected. Performances of the training, testing and validation data are shown in Figure 4. After the network is trained, the network is tested with the test data. Goodness of fit of the model predicted data and actual data are determined. On the basis of R2 and MSE, the best performed NN architecture, respectively, is selected with 2 hidden layers and 9 and 17 nodes in the first and second hidden layer. Performance of the some of the networks is tabulated in Table 2. From Table 2, it can be seen that 2 hidden layers with 9 and 17 nodes in the first and second layers, respectively, show the best performance. The selected

Input contributions to adhesion The relative contribution of the input parameters on the prediction of the output parameter can be calculated using Garson’s scheme (Garson, 1991; Tarefder et al., 2005b). This scheme provides a technique so that the connection weights of the hidden layers and output layers can be broken into components and partitioned in an association with each input node. The connections weights give weightage to the input data that are transferred from layer to layer to predict output. Therefore, partitioning connection weights with respect to each input parameter can infer relative contribution of the input parameters on the predicted output. The model built in this paper is a 10-9-17-5 network, i.e. the input layer (i) has 10 nodes, two hidden layers (h1 and h2 ) have 9 and 17 nodes, respectively, and the output layer (o) has 5 nodes. To apply the Garson’s scheme, the relative influence of the inputs on a single output (output node = 1) is considered at a time. Also, the bias weights are not involved in the calculation. At first, the connection weights between the h1 -h2 and h2 -o are taken for the calculation (shown in Table 3). Dimensions of the weight matrices in between h1 -h2 and h2 -o are 17 × 9 and 17 × 1, respectively, considering one output only. Now, absolute product of the wh1-h2 and wh2-o are taken and stored as β factor of each h2 node using the following equation: β pq = |wq p .w po |,

(7)

where p = node identification in the second hidden layer, range from 1 to 17,  C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

NEURAL NETWORK MODELLING

37

Fig. 5. Structure of fully connected NN selected for model development. Connections for first and last node are shown for convenience.

q = node identification in the first hidden layer, range from 1 to 9, wqp = weight between node q in h1 and node p in h2 , wpo = weight between node p in h2 and output node o (=1). For each value of p, i.e. for each hidden 2 node, β pq are calculated changing the q from 1 to 9. Then, θ is calculated from β using the following equation: β pq θ pq = 9 . q =1 β pq

(8)

Now, θ values are summed up to calculate another variable X, which has a dimension of h1 ×1. The equation to calculate X is as follows: Xq 1 =

17 

θ pq .

(9)

p=1

Here, q is given input from 1 to 9 at a time. Results for θ pq and X q 1 are tabulated in Table 4.  C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

This Xp1 is used as the output connection weights and same calculations are to be repeated. That is, for this case, instead of weights between hidden 1 and hidden 2, weight between input and hidden 1 will be used and instead of weights between hidden 2 and output, Xp1 will be used. After calculating Xinput for each input node, the percentage contribution of each input parameter can be calculated from the Xinput values. The contributions of each input in terms of percentage are summarized in Table 5. From Table 5, it can be concluded that tip type is the most significant input factor in determining the five-point forces by NN model. Tip type, which means asphalt’s chemical composition (-COOH, -OH, NH3 , etc.), shows the highest per cent contribution calculated using the connection weights.

Application of NN model The developed NN model is applied to observe the effect of lime on adhesion of the asphalt sample probed with the –OH

38

R.A. TAREFDER AND S. AHSAN

Table 3. Weight matrix in between hidden layer h1 - h2 and hidden layer h2 – output.

Wqp

p, q = node no. for h1 and h2 , respectively

q=1

q=2

q=3

q=4

q=5

q=6

q=7

q=8

q=9

Wpo o=1

p=1 p=2 p=3 p=4 p=5 p=6 p=7 p=8 p=9 p = 10 p = 11 p = 12 p = 13 p = 14 p = 15 p = 16 p = 17

−0.5642 −0.8331 −0.1407 0.1078 0.3763 −0.1605 0.5651 0.2014 0.3445 1.7815 −0.1286 0.5369 0.1671 −0.2173 1.0206 −0.1332 1.3474

0.9894 1.1865 0.3994 2.1470 −0.7972 1.9256 0.0181 0.2577 −0.1790 0.3999 1.3722 −0.2227 −0.4992 0.2528 0.5540 0.0085 −0.4193

0.5532 −0.9121 1.0503 1.1310 0.0697 −0.5951 0.6520 0.7116 0.7654 0.8974 0.5688 0.5892 0.0014 1.1002 0.5069 0.3480 1.5051

0.1377 0.1831 1.3467 −0.6610 0.7432 1.1997 0.2386 1.0780 0.1195 0.3664 0.5659 0.6690 −0.3496 1.6409 0.2062 −0.0294 0.2014

0.7926 0.1506 −0.2387 2.1617 0.7016 0.3957 0.9507 0.4942 0.9234 0.8416 0.8820 0.8529 1.2543 −0.6686 0.7117 2.6521 1.3865

−0.4533 0.3526 0.3342 0.4993 −1.9496 0.6340 −0.7858 −0.3299 −0.2831 1.3948 −0.3719 −0.4085 1.4623 0.7790 0.7356 2.1356 1.0188

0.4497 −1.5260 0.2030 −0.2880 1.0570 −0.1410 −0.2418 −0.4311 0.1929 2.1984 1.0743 1.2039 0.1464 1.9162 1.0013 0.8522 0.8096

1.0612 0.9021 −0.1403 −0.3304 −0.7026 −0.2227 −0.3526 0.7772 −0.5874 0.6136 0.8555 0.8320 −0.1261 −0.3292 0.7710 0.0675 0.2112

0.6833 0.5847 −0.8341 0.8050 0.5930 0.6351 1.4906 −0.3067 0.7496 0.6020 0.6285 −1.4377 0.7624 −1.6200 1.2584 0.4607 2.0621

0.3803 0.0224 −0.0531 0.0043 −0.2896 0.1373 0.3028 −0.1722 −0.1569 −0.0871 −0.1616 −0.3357 −0.2938 −0.0733 0.1289 0.0680 −0.0730

Wqp

Table 4. Calculated values of θ and X using (6)–(8). p, q = node no. for h1 and h2 , respectively θ pq

Xpq

p=1 p=2 p=3 p=4 p=5 p=6 p=7 p=8 p=9 p = 10 p = 11 p = 12 p = 13 p = 14 p = 15 p = 16 p = 17

θ pq q=1 0.0993 0.1256 0.0300 0.0133 0.0538 0.0272 0.1067 0.0439 0.0831 0.1959 0.0199 0.0795 0.0350 0.0255 0.1508 0.0199 0.1504 1.2599

q=2

q=3

q=4

q=5

q=6

q=7

q=8

q=9

0.1740 0.1789 0.0852 0.2640 0.1140 0.3259 0.0034 0.0562 0.0432 0.0440 0.2128 0.0330 0.1047 0.0297 0.0819 0.0013 0.0468 1.7990

0.0973 0.1375 0.2241 0.1391 0.0100 0.1007 0.1231 0.1551 0.1847 0.0987 0.0882 0.0873 0.0003 0.1291 0.0749 0.0520 0.1680 1.8700

0.0242 0.0276 0.2873 0.0813 0.1063 0.2030 0.0451 0.2350 0.0288 0.0403 0.0878 0.0991 0.0733 0.1925 0.0305 0.0044 0.0225 1.5889

0.1394 0.0227 0.0509 0.2659 0.1004 0.0670 0.1795 0.1077 0.2228 0.0925 0.1368 0.1263 0.2630 0.0784 0.1052 0.3966 0.1547 2.5099

0.0797 0.0532 0.0713 0.0614 0.2789 0.1073 0.1484 0.0719 0.0683 0.1534 0.0577 0.0605 0.3066 0.0914 0.1087 0.3194 0.1137 2.1517

0.0791 0.2301 0.0433 0.0354 0.1512 0.0239 0.0457 0.0940 0.0465 0.2417 0.1666 0.1783 0.0307 0.2248 0.1480 0.1274 0.0903 1.9571

0.1867 0.1360 0.0299 0.0406 0.1005 0.0377 0.0666 0.1694 0.1417 0.0675 0.1327 0.1232 0.0265 0.0386 0.1140 0.0101 0.0236 1.4452

0.1202 0.0882 0.1780 0.0990 0.0848 0.1075 0.2815 0.0668 0.1809 0.0662 0.0975 0.2129 0.1599 0.1900 0.1860 0.0689 0.2301 2.4183

functionalized tip only. From the test data of the NN model, one asphalt sample data are chosen that is probed with –OH tip. Using these data, five synthetic data are produced. In the synthetic data, all the input factors are kept constant except the lime percentage is varied from 0.5% to 2.5% at an increment of 0.5%. This procedure is adopted for a total of eight different combinations of asphalt samples shown in the test matrix

below: 2(Wet, dry) × 2(SB, SBS) × 2(3% polymer, 5% polymer) ×5(lime percent) = 40. Now, five synthetic data along with the actual laboratory test data are fed into the NN model and the adhesion forces are predicted.

 C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

NEURAL NETWORK MODELLING

39

Table 5. Per cent contribution of input factors. Input variables Per cent contribution

Wet/dry

SB/SBS/none

% of polymer

% of antistripping agent

Tip type

X1

X2

X3

X4

X5

9.6

10.1

11.4

5.7

22.2

12.6

3.2

11.9

9.0

4.2

Fig. 6. Force versus distance plots for wet and dry polymer modified asphalt samples.

Figure 6 shows force–distance plots (A), (B), (C) and (D) for asphalt samples modified with SBS. Each plot contains six curves. Five of the six curves are synthetic data produced force– distance curves for which the forces are predicted by the NN model. The remaining curve is the force–distance curve of the actual laboratory test data. The NN predicted force–distance curve showed success in mapping almost similar patterns as the laboratory force–distance curve. In addition to that, effects in the adhesion determination due to change in lime percentage are depicted in Figure 7. Except for the dry sample modified  C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

with 3% SBS polymer (Fig. 6B), the predicted curve pattern for all other samples indicates that samples with a higher percentage of lime give higher values of adhesion force. Figures 7(A) and (B) are plotted for lime-treated asphalt samples modified with 3% and 5% SBS and SB polymers. Lime varies from 0.5% to 2.5% at an increment of 0.5% and adhesion forces of the asphalt are predicted by the NN model for asphalt samples probed with the –OH functionalized tip. From both Figures 7(A) and (B), it is noted that adhesion forces of wet samples are higher than that of corresponding

40

R.A. TAREFDER AND S. AHSAN

Fig. 7. Adhesion force versus antistripping agent for wet and dry polymer modified samples.

dry samples for all lime percentages. In AFM testing, weaker samples give higher values of adhesion forces than the stronger samples (Arifuzzaman, 2010). That is, the higher the adhesion force, the weaker the sample. Therefore, all the samples have gone through moisture damage. For the SBS sample shown in Figure 7(A), all samples except the dry 3% sample show an increasing trend in adhesion value with the increase in lime percentages. The dry 3% SBS sample seems to have a reverse effect of lime increment in determining adhesion force. Wet and dry samples modified with 5% SBS have almost the same rate of increment in the adhesion value with the increase in lime percentage. On the other hand, wet asphalt samples modified with SB polymer, especially 3% SB samples, depict decreasing trend with the increase in lime percentages. Increase in the lime percentage shows an insignificant effect on adhesion forces of the dry SB samples. This may be due to the fact that in SB polymer modified asphalt, lime cannot be activated without the presence of water. Thus, Adhesion force in wet 3% SB asphalt samples decreases at a high rate with the increase in lime percentage compared to corresponding dry samples. The decreasing trend of adhesion indicates that the sample becomes stronger with the increase in lime percentage. Thus, the increase in lime percentage helps to reduce the adverse moisture effect in 3% SB modified samples. In addition, 3% SB modified sample provides better platform for lime to resist moisture effect compared to 5% SB modified sample Comparing Figures 7(A) and (B), it can be concluded that increase in percentage of lime is effective in reducing moisture damage on 3% SB modified samples. Increasing lime on wet 3% SB samples shows significant reduction in moisture damage. Five per cent SB and SBS samples maintain almost same difference in wet and dry adhesion with the increase in percentage of lime. Therefore, these samples show negligible effect of addition of lime in reducing moisture damage. However, it should be noted that these results and trends are for the –OH

functionalized tip. The trends can vary if using a functionalized tip other than the –OH. Conclusions

r

r

r

The study constructs an NN that can be fitted to AFM data to predict adhesion. A four-layer feed-forward network is developed and the established network shows good prediction ability for the test input variables that were not introduced to the network before. The relative contribution of each input variable is calculated using the connection weights in between the NN layers. The results indicate the tip type (–OH, –COOH, etc.) contributed highest percentage (22.2%) on the determination of five-point forces. That is, asphalt chemistry has the highest significant effect on adhesion force. The developed network is applied to observe the effect of the change in percentage of lime on the determination of adhesion force for eight different types of asphalt samples. The results indicate that in the SBS sample, adhesion values increase with an increase in lime percentage. That means lime does not help reduce moisture damage potential in the SBS sample. The trend is opposite in the SB sample. Therefore, lime is effective in SB but not in SBS polymer modified samples.

Acknowledgements This project was funded by National Science Foundation (NSF), the prestigious NSF CAREER award, Award Number: 0644047. References Alsugair, A.M. & Al-Qudrah, A.A. (1998) Artificial neural network approach for pavement maintenance. J. Comput. Civil Eng. 12, 249– 255.

 C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

NEURAL NETWORK MODELLING

Arifuzzaman, M. (2010) Nano-scale evaluation of moisture damage in asphalt. PhD dissertation, University of New Mexico, Albuquerque. Ceylan, H., Guclu, A., Tutumluer, E., Gomez-Ramirez, F. & Thompson, M.R. (2004) Neural network-based structural models for rapid analysis of flexible pavements with unbound aggregate layers. In 6th International Symposium on Pavements Unbound, Nottingham Centre for Pavement Eng., School of Civil Eng., University of Nottingham, England. Ceylan, H., Gopalakrishnan, K. & Bayrakc, M.B. (2008) Neural networks based concrete airfield pavement layer moduli backcalculation. Civil Eng. Environ. Sys. 25, 185–199. Ferregut, C., Abdallah, I., Melchor, O. & Nazarian, S. (1999) Artificial neural network-based methodologies for rational assessment of remaining life of existing pavements. Report, Center for Highway Materials Research, The University of Texas at El Paso, El Paso, Texas. Garson, G. (1991) Interpreting neural network connection weights. Artif. Intell. Expert 6(4), 47–51. Gopalakrishnan K. (2010) Effect of training algorithms on neural networks aided pavement diagnosis. Int. J. Eng. Sci. Tech. 2(2), 83–92. Gucunski, N., Krstic, V. & Maher, M.H. (1998) Backcalculation of pavement profiles from the SASW test by neural networks. Artificial neural networks for civil engineers: advanced features and applications (ed. by I. Flood & N. Kartam), Amer. Soc. Civil Eng. Reston. pp. 191– 222. Huang, Y. & Moore, R.K. (1997) Roughness level probability prediction using artificial neural networks. Trans. Res. Rec. 1592, 89–97. Huang, C., Najjar, Y.M. & Romanoschi, S.A. (2007) Predicting asphalt concrete fatigue life using artificial neural network approach. Paper No. 07-1607, In 86th Transportation Research Board Annual Meeting (CD-ROM), Trans. Res. Board (TRB), National Research Council, Washington, DC. Kim, Y. & and Kim, Y.R. (1998) Prediction of layer moduli from falling weight deflectometer and surface wave measurements using artificial neural network. Trans. Res. Rec. 1639(1), 53–61. Kim, J.I., Kim, D.K., Feng, M.Q. & Yazdani, F. (2004) Application of neural networks for estimation of concrete strength. J. Mater. Civil Eng. 16, 257–264.

 C 2014 The Authors C 2014 Royal Microscopical Society, 254, 31–41 Journal of Microscopy 

41

Lee, Y.C., Kim, Y.R. & Ranjithan, S.R. (1998) Dynamic analysis-based approach to determine flexible pavement layer moduli using deflection basin parameters. Trans. Res. Rec. 1639(1), 36–42. Meier, R.W. & Rix, G.J. (1995) Backcalculation of flexible pavement moduli from dynamic deflection basins using artificial neural networks. Trans. Res. Rec. 1473, 72–81. Meier, R., Alexander, D. & Freeman, R. (1997) Using artificial neural networks as a forward approach to backcalculation. Trans. Res. Rec. 1570, 126–133. Owusu-Ababio, S. (1998) Effect of neural network topology on flexible pavement cracking prediction. Comput.-Aided Civil Infrastruct. Eng. 13, 349–355. Saltan, M. & Terzi, S. (2004) Backcalculation of pavement layer parameters using artificial neural networks. Indian J. Eng. Mat. Sci. 11(1), 38–42. Tarefder, R.A., White, L. & Zaman, M.M. (2005a) Development and application of a rut prediction model for flexible pavement. Trans. Res. Rec. 1936, 201–209. Tarefder, R.A., White, L. & Zaman, M. (2005b) Neural network model for asphalt concrete permeability. J. Mat. Civil Eng. 17(1), 19–27. Tasdemir, Y. (2009) Artificial neural network for predicting low temperature performances of modified asphalt mixtures. Indian J. Eng. Mater. Sci. 16, 237–244. Testa, S.M. (1995) Chemical aspects of cold-mix asphalt incorporating contaminated soil. J. Soil Contamination 4(2), 191–207. Williams, T.P. & Gucunski, N. (1995) Neural networks for backcalculation of moduli from SASW test. J. Com. Civil Eng. 9(1), 1–8. Xiao, F. & Amirkhanian, S.N. (2009a) Effects of binders on resilient modulus of rubberized mixtures containing RAP using artificial neural network approach. J. Test. Eval.: ASTM. 37(2), 129–138. Xiao, F. & Amirkhanian, S.N. (2009b) Artificial neural network approach to estimating stiffness behavior of rubberized asphalt concrete containing reclaimed asphalt pavement. J. Trans. 135(8), 580–589. Yang, J., Lu, J.J. & Gunaratne, M. (2003) Application of neural network models for forecasting pavement crack index and pavement condition rating. Final report. Department of Civil and Environmental Engineering, University of South Florida, Tampa, Florida.

Neural network modelling of asphalt adhesion determined by AFM.

This study constructs a neural network (NN) model to quantify adhesion from atomic force microscopy (AFM) data. AFM data contain five-point force-dist...
901KB Sizes 0 Downloads 0 Views