Академический Документы
Профессиональный Документы
Культура Документы
ISSN 1866-7511
Arab J Geosci
DOI 10.1007/s12517-014-1304-3
1 23
Your article is protected by copyright and
all rights are held exclusively by Saudi
Society for Geosciences. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.
1 23
Author's personal copy
Arab J Geosci
DOI 10.1007/s12517-014-1304-3
ORIGINAL PAPER
permeability, 320 data sets for modeling MDD and OMC, and
105 cases for modeling effective friction angle of shearing.
Different types of soils were used in these tests; therefore, the
results of this research could be valid for all types of soils.
As presented in Fig. 1, the database is obtained from
different sources, mainly from the geotechnical engineering
laboratory of the Federal University of Bahia (UFBA), Brazil.
Also, other cases from Wang and Huang (1984) and Mousavi
et al (2011) were added to the former source.
For each data set, the values of permeability, OMC, MDD,
compaction degree, friction angle, and soil index properties
(grain size curve, Atterbeg limits, and specific density) were
available. Table 1 gives the descriptive statistics of the vari-
ables used for the compaction characteristics and permeability
model developments. The variation ranges of the parameters
Fig. 1 Data sources used to compile the database used for effective friction angle model are summarized in
Table 2. It is noteworthy that the gravel content (Gc) was a
57 synthetic soils, developed several regression equations to coarse aggregate having a particle size coarser than 4.75 mm,
correlate the compaction characteristics and permeability with and the grain size of sand content (Sc) ranged from 4.75 to
the soil’s index properties. Koltermann and Gorelick (1995), 0.075 mm. As well, particles smaller than 0.075 mm were
Boadu (2000), Chapuis (2004), Sinha and Wang (2008), and, named as fine content (Fc). In addition, a quantity called
more recently, Cote et al. (2011) developed different numeri- fineness modulus (FM) is also computed. The value of fine-
cal and analytical models to estimate the hydraulic conductiv- ness modulus multiplied by 100 is equal to the sum of the
ity of soils based on their index parameters. percentages of particles coarser than 3/4 in., 3/8 in., no. 4, and
In recent years, some efforts have been made to develop a no. 100 mm. Note that the friction angle values in the database
correlation between the effective friction angle and the basic were determined by consolidated–drained triaxial tests.
properties of soil. Kayadelen et al. (2009) used artificial neural
network (ANN), genetic programming (GP), and adaptive Modeling method
neuro-fuzzy methods to predict the φ′ value of soils.
Mousavi et al. (2011) used GP and orthogonal least squares Artificial neural networks were employed to develop predic-
algorithm to present a correlation between the internal friction tion models. To evaluate the importance of each factor on the
angle and the physical properties of soils such as the fine and prediction models, a set of sensitivity analyses have been
coarse content, density, and liquid limit. Sezer (2013) performed.
employed nonlinear multiple regression, neuro-fuzzy, and Artificial neural networks are information-processing sys-
ANN methods to predict this soil parameter. tems whose architectures essentially mimic the biological
system of the brain (Goh 1994). ANNs have been successfully
applied to link independent variables to a series of dependent
Materials and methods ones, mainly where it is diverse to establish numerical equa-
tions. The use of ANNs has increased during the last decades
Database in various fields of geotechnical engineering, such as lique-
faction (Ural and Saka 1998; Najjar and Ali 1998), foundation
A database including a total number of 580 data sets was settlements (Sivakugan et al. 1998), reinforced soil (Ghiassian
compiled, in which 155 data sets were used for modeling the et al. 2006), and compaction characteristics of soils (Sinha and
Table 1 Descriptive statistics of the variables used for MDD, OMC, and permeability model
Parameter Gc (%) Sc (%) Fc (%) FM LL (%) PL (%) Gs Cd (%) MDD (kN/m3) OMC (%) Ka (cm/s)
Fig. 5 Comparison of the ANN results on MDD model by excluding each parameter
Author's personal copy
Arab J Geosci
Fig. 6 Comparison of the ANN results on OMC model by excluding each parameter
Xn
Evaluation method ðM i −Pi Þ
COD ¼ 1 − Xi¼1
n
ð1Þ
Different statistical approaches were used to evaluate the perfor- i¼1
M 1 −M
mance of the prediction model. These parameters were coeffi-
cient of determination (COD), root mean squared error (RMSE), sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn
and coefficient of residual mass (CRM). The following equa- i¼1
ðPi −M i Þ2
RMSE ¼ 100 ð2Þ
tions are the mathematical expressions of these parameters. n
Outputs Statistics Total Ex. w1a (Gc) Ex. w2 (Sc) Ex. w3 (Fc) Ex. w4 (Gs) Ex. w5 (LL) Ex. w6 (PL)
Ex. excluding
a
Scalar weights
Author's personal copy
Arab J Geosci
(PL). The only output was MDD. Hence, the input layer has
six neurons and the output layer has one neuron. Among 320
measured data sets, 290 sets (90 %) have been used for training
and 30 sets (10 %) have been used for testing the model.
Figure 3 shows the predicted values of MDD versus the
experimental data. As is clear from the graphs, the MLP model
gives very reliable estimates of the maximum dry density.
A similar ANN structure with the same input parameters
(Gc, Sc, Fc, Gs, LL, PL) has been used to estimate OMC.
Figure 4 shows a comparison between the results of the
developed ANN model and the experimental data. The ob-
tained values of COD and RMSE demonstrate the accuracy of
the developed model.
The results of the sensitivity analysis performed on the
MDD prediction model is presented in Fig. 5. In each graph,
one input variable is excluded (by setting its scalar weight to
zero) and the ANN model is trained for the five remaining
input parameters. As is obvious from the graphs, excluding
each parameter causes some extra scatter in the prediction
Fig. 7 Comparison between the predicted values of permeability coeffi- model for MDD.
cient and the actual data
The results of the sensitivity analysis performed on the OMC
prediction model are also presented in Fig. 6. The graphs
Xn
ðPi Þ illustrate that LL and PL are the most important variables in
CRM ¼ 1 − Xni¼1 ð3Þ the prediction of OMC. Moreover, it can be seen easily that
i¼1
ðM i Þ removing Fc leads to perturbation in the results.
Table 3 also presents a summary of ANN performance.
where Mi and Pi are the measured and predicted values, Based on the results, in the case of MDD, excluding Fc from
respectively, M is the mean of the measured values, and n ANN increases the RMSE to 390.9, indicating the significant
is the number of samples. The RMSE is the variance of the role of this parameter on the ANN predictions. Also, a CRM
residual error and should be minimized when the outputs fit a of 0.21 shows an underestimation of the MDD. LL and PL are
set of data. In the case of a perfect fitting, the RMSE is zero. the other important parameters in the prediction of MDD.
The lower the RMSE, the higher is the accuracy of the model Moreover, it is clear from the results that excluding Gs and
predictions. The coefficient of residual mass, CRM, is an PL causes an overestimation of the ANN prediction.
analysis of the difference between the measured and predicted Considering the OMC results, Fc, LL, and PL are the most
values. The optimum value of CRM is zero. Positive values of important variables in the ANN prediction.
CRM indicate underestimation and vice versa.
ANN prediction model for permeability
Results and discussions Six input variables were used for the ANN model for perme-
ability coefficient, including FM, LL, gravel content (Gc), sand
ANN prediction model for MDD and OMC content (Sc), fine content (Fc), and compaction degree (Cd). The
only output was LogK. Thus, the input layer has six neurons
Six input variables were used for the ANN model for MDD, and the output layer has one neuron. Among 155 measured data
including gravel content (Gc), sand content (Sc), fine content sets, 140 sets (90 %) have been used for training and 15 sets
(Fc), specific density (Gs), liquid limit (LL), and plastic limit (10 %) have been used for testing the model. Figure 6 shows the
Output Statistics Total Ex. w1 (Gc) Ex. w2 (Sc) Ex. w3 (Fc) Ex. w4 (FM) Ex. w5 (LL) Ex. w6 (Cd)
Fig. 8 Comparison of the ANN results on permeability model by excluding each parameter
predicted values of LogK versus the experimental data. Figure 7 conducted hydraulic conductivity tests on compacted clays.
shows an almost perfect prediction of hydraulic conductivity They plotted the variation of permeability at different densities
based on the employed index parameters; however, in the (Fig. 9). As could be seen, for the same soil, permeability
testing cases, the predictions exhibit a higher scatter. could be changed almost in one order of magnitude. Ahangar-
Table 4 and Fig. 8 illustrate the results of the sensitivity Asr et al. (2011) have presented the same results for the effects
analysis performed on the permeability prediction model. In of compaction degree on the permeability of soils. The results
accordance with Table 4, removing Cd increases the RMSE to of their parametric analysis show a decreasing trend for per-
431.4, which indicates that the compaction degree is of great meability by increasing the degree of compaction. In addition,
importance in predicting the permeability of soils. In this case, Mesri and Olson (1971), by comparing the permeability of
the obtained value for CRM shows an overestimation of different type of clays, stated that the size and arrangement of
permeability. As is clear from the results, Gc, Sc, and LL are particles have a significant effect on the permeability coeffi-
the other important variables in the ANN prediction model. cient. According to their studies, decreasing the void ratio,
Another aspect that is obvious from the results is that exclud- which is a result of increasing the compaction degree, will
ing Fc and FM causes an underestimation of the ANN model. cause the permeability of soils to decline.
As mentioned above, the degree of compaction has the
highest effect on the prediction of permeability coefficient. It ANN prediction model for effective friction angle
is well known that the volume of voids decreases due to an
increase in the degree of compaction, which causes the per- Five input variables were used for the ANN model for effec-
meability of a soil to decrease. Boynton and Daniel (1985) tive friction angle of shearing, including coarse content (Cc),
Author's personal copy
Arab J Geosci
fine content (Fc), liquid limit (LL), soil bulk density (γ), and
shearing rate (Sr). The only output was φ′. Thus, the input
layer has six neurons and the output layer has one neuron.
Among 105 measured data sets, 95 sets (90 %) have been used
for training and 10 sets (10 %) have been used for testing the
model. Figure 10 shows the predicted values of φ′ versus the
experimental data.
Table 5 and Fig. 11 also present the results of the sensitivity
analysis performed on the ANN model of effective friction
angle. It can be concluded from the results that the soil bulk
density has the highest influence on the prediction of the
effective friction angle. Another aspect that is coherent from
Fig. 11 Comparison of the ANN results on friction angle model by excluding each parameter
the result is that excluding the soil bulk density increases University of Bahia (UFBA), Brazil was utilized; however
the value of CRM significantly and leads to an underes- some other experimental data cases from the literature were
timation of effective friction angle. Also, the sensitivity added to this data bank.
analysis conducted by Mousavi et al. (2011) indicates the The results of the prediction models have been compared
importance of the bulk density on the friction angle of with the experimental data. Comparison of the results demon-
soils. strates that the developed ANN models provide highly accu-
rate predictions and that the existing models can be improved
with increasing the database. A major strength of the ANN
Summary and conclusion prediction models is their ability to improve as more data
become available without repeating the development proce-
This paper presents a set of ANN models for predicting the dures from the beginning.
compaction characteristics, permeability, and effective fric- Furthermore, a set of sensitivity analyses have been
tion angle of shearing based on the soil index properties. performed to illustrate the influence of each parameter
Since laboratory tests to determine these key parameters are on the ANN’s performance. To evaluate the performance
laborious and time-consuming, it is desirable to develop of the prediction models, three statistical approaches have
prediction models to estimate these parameters based on been utilized: coefficient of determination (COD), root
index parameters which are easy to measure. To do this, mean squared error (RMSE), and coefficient of residual
the databank of geotechnical laboratory of the Federal mass (CRM).
Author's personal copy
Arab J Geosci