Machine learning techniques applied to

prediction of residual strength of clay

 Prediction of residual clay strength is done based on index properties using

two machine learning techniques described in the paper.

 Different Artificial Neural Network (ANN) models and Support Vector

Machine (SVM) techniques have been used.

 Stability of natural slopes or landslides depends upon the shear

strength parameters of clay, which varies significantly between
the peak and residual states

 Residual strength Is the load or force that a damaged object or

material can still carry without failing.
Material toughness, fracture size and geometry as well as its orientation all
contribute to residual strength

Peak strength is the maximum value of the shear stress or the

maximum value of the ratio of shear stress to effective mean or
normal stress.
 At a residual state, the clay has undergone a large strain
and hence remoulded, with previous bonds broken and
prior fabric destroyed. As a result, in a residual state the
clay exhibits no cohesion. Further, the friction angle (φr)
is substantially less than the friction angle at a peak state
(φp), resulting in lower shear strength.

Methodology
In the present study artificial intelligence techniques ANN
And SVM have been used to examine the effects of the
Index properties on the residual friction angle. The salient
Features of the ANN and SVM

Artificial Neural Network

modelling
is an artificial intelligence system which is an alterative
statistical tool
Inspired by the behaviour of the human brain and nervous
system. It has been applied successfully to a
wide range of geotechnical problems. In the present
study the ANN models are developed using two different
training algorithms, Bayesian Regularization Neural
Network(BRNN)and Differential Evolution Neural
Network(DENN)

Support Vector Machine (SVM)

 originated from the concept of statistical learning theory
pioneered by Boser
et al. Our study uses the SVM as a regression
technique by introducing an error (ε)ε-insensitive loss function.
In this section , a brief introduction of how to
construct an SVM for regression problems is presented.
Further details can be found in many publications

2.1.2. Differential Evolution Neural Network

(DENN)
The training of the feed- for ward neural network using
differential evolution optimization is known as a Differential
Evolution Neural Network (DENN)[14]. The DE optimization
is a population-based heuristic global optimization
method. Unlike o the revolutionary optimization techniques
, in DE the vector sin current population s are randomly sampled
and combined to create vectors for the next
Generation . The real valued crossover factor and mutation
Factor govern the convergence of the search process . The
Details of DENN are available in the literature[14].The
Results of an ANN model trained with DENN and BRNN
Are compared with that from the commonly used LMNN
To discuss the quality of the prediction of the different
networks.
LMNN is the most widely used ANN method and has been
Extensively used in geo techn

Stability of natural slopes or landslides depends upon the shear

strength parameters of clay, which varies significantly between
the peak and residual states. At residual strength, due to
remoulding, clay exhibits negligible cohesion and a decreased
value of friction angle as compared to the peak state. Right from
the early studies, residual strength has been associated with
both reactivated landslides and first-time slope failures in terms
of residual friction angle (ϕr).

Skempton was the first to study the effect of drained residual

shear strength of soil for stability analysis of reactivated
landslides and suggested that the decrease of shear strength is
partly due to changes in orientation of clay particles upon
unidirectional shearing. Kenney reported the effect of
mineralogical composition of soils on their residual strength.
Based on an analysis of 99 cases of landslide failure in 36 types
of soft clays, stiff clays and clay shales, Mesri and Shahien
observed that residual strength also develops in first-time slope
failures.
Several attempts have been made in the past to correlate the
residual friction angle of soils and their index properties such as
Atterberg limits and clay fraction (CF). Skempton related the
residual friction angle (ϕr) value with the clay fraction. Many
researchers postulated relationships between ϕr and plasticity
index (PI). Relationships between ϕr and liquid limit (LL) were also
proposed. For sedimentary soil observed that the type of minerals
and percent of clay governs the value of ϕr. Using LL as an
indicator of clay mineral, they proposed correlations of ϕr with LL
for various ranges of CF. Wesley observed that, most of the soil
above the A-line have the ϕr < 10°, while those below the A-line
have higher values of ϕr. A good relationship was found between ϕr
and deviation from the A-line (ΔPI) for soils with LL > 50. The ΔPI
is denoted as(1)ΔPI=PI−0.73×(LL−20)
Based on direct shear test results on simulated soil–rock mixtures
that were developed by mixing kaolinite clay with sand, Vallejo and
Zhou indicated that the shear strength of the whole mixture was
governed by the concentration of sand in the mixture. For sand
content <50%, the shear strength was influenced by clay. For sand
concentration between 50 and 80%, the shear strength was
provided partly by the shear strength of kaolin and partly by the
frictional resistance between sand grains.
Based on an experiment on 80 specimens, Tiwari and Marui (2005)
presented a triangular correlation chart to calculate ϕr based on
mineralogical composition of soils. The chart provided correlation
of ϕr with the liquid limit, the plasticity index, the clay fraction, the
specific surface area and the proportion of the clay mineral
smectite. This model gave good results for the specimen tested by
them, but failed to correctly predict the values for 53 other samples
tested by other researchers.
Wen et al. (2007) examined the residual strength of soils from the
slip zones of about 170 landslides in the Three Gorges Project
(TGP) area, China, and concluded that clay content and Atterberg
limits could be used to estimate the residual strength of soils finer
than 2 mm, but they are not appropriate for the evaluation of
residual strength of soils containing a considerable amount of
gravel-sized particles.
Studies done by Kaya and Kwong (2007) on soil properties of some
active landslides in Hawaii showed a poor correlation between soil
index properties and ϕr for colluvial soils, which are rich in an
amorphous phase. Another study by Yanrong (2009) on slip zones
of large landslides in the Three Gorges Project, China observed
Atterberg limits, particle size distribution, normal stress, particle
shape and shearing rate as the most influential factors affecting
residual shear strength of composite soil. Thus the previous studies
suggested that ϕr is affected by a number of index properties. But
most of the relationships cited earlier are in the form of graphs and
are not easy to use by geotechnical engineers in practice.
Nowadays Artificial Intelligence (AI) techniques like artificial
neural network (ANN), support vector machine (SVM), and genetic
programming (GP) are being used as alternate techniques to
statistical methods in different fields of science and technology.
Yaghouby and Ayatollahi (2009) used SVM for multi-classification of
cardiac arrhythmias into five classes with an accuracy of 99.78%.
Yaghouby et al. (2009) used heart rate variability (HRV) signal to
classify the cardiac arrhythmias into four classes using ANN
analysis and found that it was very efficient with 100% accuracy.
The use of ANN has been also found to be efficient as a PID
controller (Dong et al., 2014). However, GP was found to be more
efficient compared to radial basis function (RBF) neural network in
the automatic detection of atrial fibrillation based on HRV signals
(Yaghouby et al., 2010). AI techniques have been found to be better
prediction tools for geoscience problems than conventional
techniques served that for prediction of peak ground acceleration,
ANN based model with strong motion has better prediction
performance compared to conventional nonlinear regression
models. Similarly, reported that for liquefaction susceptibility
analysis of ground using in-situ data, ANN based model is more
efficient compared to available empirical models. Das presented a
comprehensive review of the successful application of ANN in
different geotechnical engineering problems.
Das and Basudhar (2008) used artificial neural network (ANN)
modelling to predict the ϕr of clay, but their study was limited to
tropical soil of a specific region only. Das et al. (2011) provided an
equation for the calculation of ϕr of soil based on their analysis of
data using ANN and SVM.
However, ANN has poor generalization, attributed to attainment of
local minima during training and needs iterative learning steps to
obtain better learning performances. SVM has better generalization
compared to ANN, but the parameters (C) and insensitive loss
function (ɛ) need to be fine-tuned by the user. Moreover, these
techniques will not produce a comprehensive model equation and
are also called as ‘black box’ system
Recently, a new prediction method, functional network (FN), which
is based upon the structure of the physical world has been used in
many fields of science and engineering including), signal
processing, pattern recognition, function approximations, real-time
flood forecasting, science mining, and structural engineering FN
was introduced by Castillo), and Castillo.as a powerful alternative
to ANN.
FN as a new modelling scheme has been used in solving both
prediction and classification problems. Hence, in the present study
an attempt has been made to predict the residual friction of soil
using FN based on a set of index properties including LL, PI, CF
and ΔPI. The data set used for the study is the same as used by
Functional Networks have not been applied to geotechnical
engineering issues to the best of the knowledge of the authors. The
following section briefly describes the concepts of FN. The results
from FN have been compared with the results from ANN and SVM
as obtained

4. Results and discussions

Two different models have been analyzed in the present study
with the use of FN. The first model (Model 1) consists of four
inputs (LL, PI, CF, ΔPI) with one output (ϕr). The second (Model
2) model consists of two inputs (ΔPI and CF) with one output ϕr.
The data were normalized in the range [0, 1] for analysis. The FN
was implemented

4.1. Model 1
Out of the 131 available data, 98 (75%) randomly selected data were used for
training and the rest was used for testing. Using prescribed neural function and
degree of the selected function, the model in functional network is developed. By
an increase in the degree the obtained results would be more accurate, but at the
same time the complexity of the problem also increases. shows a plot between the
degree of neural function and R value in testing obtained for polynomial,
exponential, sin and cos functions. It can be seen from the plot that there is a sharp
rise in the R value initially. After a degree 10 of the neural function, the R value
remains approximately the same. If a degree above 10 is taken, it increases the
complexity of the model without much increase in the accuracy. Hence a tradeoff
was made in the present study and a polynomial neural function with degree 10
was selected to model the FN.

4.2. Model 2
Similar to Model 1, 98 (75%) randomly selected data were used for
training and the rest were used for testing. shows a plot between
the degree of neural function and R value in testing obtained for
polynomial, exponential, sin and cos functions. For Model 2 also, a
sharp rise in the R value is observed initially. The R values
remained approximately same from degree 10 to 15 and then there
was a decrease in the R value for polynomial, exponential and cos
functions. However, a rise in the R value is observed after degree
15. In accordance with these observations, a FN model was
selected with sin as the neural function with degree 18. The
resulting associative FN is shown in The prediction equation for
Model 2 can be obtained in the same manner as Model 1. The
results obtained for Model 2 are also presented in. A plot of
predicted vs. measured values for training and testing data are
shown in a , respectively. Based on the 80% prediction limit lines, it
can be seen that most of the data points lie within the range of 80%
prediction