Professional Documents
Culture Documents
www.elsevier.com/locate/wear
PII: S0043-1648(16)30370-2
DOI: http://dx.doi.org/10.1016/j.wear.2017.02.028
Reference: WEA102082
To appear in: Wear
Received date: 4 October 2016
Accepted date: 10 February 2017
Cite this article as: D.A. Pandya, B.H. Dennis and R.D. Russell, A
Computational Fluid Dynamics Based Artificial Neural Network Model to
Predict Solid Particle Erosion, Wear,
http://dx.doi.org/10.1016/j.wear.2017.02.028
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
A Computational Fluid Dynamics Based
a
Dept. of Mechanical and Aerospace Engineering, The University of Texas at Arlington;
b
Baker Hughes
Abstract
Solid particle erosion plays a critical role in the design and reliability of equipment employed in the oil
and gas industry. Significant erosion occurs due to solid particle loading, especially in applications
involving sand production. Low particle loading in drilling fluids (<10%) is also a source of erosion
inside downhole tools and in rig equipment at the surface. Accurate prediction of erosion rates can save
money and lives by predicting failure accurately and helping to maintain the safety of the equipment.
Empirical and mechanistic models to predict erosion were primarily developed based on observations of
extensive experiments and field studies. Computational fluid dynamics (CFD) has emerged as an
alternative tool to predict erosion in recent years. The ability to simulate multiphase flows in complex
geometries using CFD makes it a valuable and less-expensive method to predict erosion flow loop
experimentations and field trials. Various empirical relations have been established to predict erosion
1
Now with Shell Global Solutions US Inc.
using CFD. These methods often predict erosion regions accurately, but typically are highly inaccurate in
This study employs machine learning approach along with CFD-based methodology to develop robust
erosion models. A generalized model is developed based on experiments conducted on 90-degree elbows
of 1-inch diameter and made from Inconel 718, Nickel Alloy 825, 25% Cr, Nickel Alloy 925, and 13% Cr
L-80 materials. The Baker Hughes erosion model developed in 2008 is studied as a baseline. Statistical
analysis was performed on CFD output parameters to identify those that most affect erosion rates. A
correlation analysis and non-parametric statistical analysis is performed resulting in the development of
two new regression models based on turbulent kinetic energy, and surface shear stress was developed. A
25-percent improvement is observed in the predictions of cumulative erosion rate error compared to
baseline. An artificial neural network with multilayer feed-forward model with the back-propagation
algorithm and Levenberg-Marquardt training was developed. This model, along with Bayesian
regularization, reduced cumulative error to less than 10%, compared to more than 40% in the baseline
Introduction
Solid particle erosion can be defined as material damage caused by solid particles that impinge a surface.
Erosion damage is seen in almost any industrial application where a solid-fluid multiphase flow occurs.
Particles carried across the streamline due to momentum, strike the walls and cause material damage.
Erosion can be the limiting factor in equipment design and might cause failure. Consequently, the erosion
process has been studied extensively for decades, although CFD-based erosion modeling is fairly recent
when compared to experimental investigations. Erosion is a complex phenomenon and is of great interest
in the oil and gas community. Accurate prediction of erosion rates can save money and lives by predicting
failure accurately and helping to maintain the safety of the equipment. Drill bits and other downhole tools
are affected by erosion from sand production at various high-rate wells. Low sand concentration flows
Artificial neural networks (ANN) are systems inspired from human brain functions. ANN are based on
simulated nerve cells or neurons that are connected in various ways to form a network. Much like the
brain, ANN can learn, memorize and create relationships among what may seem like random data.
Advances in computing power have enabled the application of ANN to real-world problems. McCulloch
and Pitts (1943) developed an ANN model based on their knowledge of neurology. Frank Rosenblatt can
be credited for the development of the first simple learning algorithm called Perceptron (1958). In the
1970s and 1980s, significant innovations occurred in the field of ANNs. The adaptive resonance theory
(ART) network was developed, based on Grossberg and Carpenter’s school of thought, which explored
resonating algorithms (1985). Klopf (1972, 1975), developed a basis for learning in artificial neurons
inspired from the biological principle of neuronal learning called heterostasis. However, one of the most
important contributions to learning was made by Paul Werbos (1974, 1988), the so-called back-
propagation learning method. It is one of the most extensively used methods today in ANN. Rezaul Begg
(2006) calls back propagation “a Perceptron with multiple layers, a different threshold function in
neurons, and a more robust and capable learning rule.” Today, ANN are used for classification,
forecasting, modeling and pattern recognition in fields like business, engineering, science, and medicine.
Artificial neural networks have been successfully developed for reservoir modeling and estimation in the
oil and gas industry (Mohaghegh et al., 1995, Shahab et al., 1997). They have also been implemented in
log data analysis (H Eskandari et al., 2004) and for continuous oil field optimization Saputelli, (2002).
Velten et al. (2000) and Zhang et al. (2003) were among the first to implement ANN in analyzing the
wear of polymer composites. More recently, work by Suresh et al. (2009) reports successful
implementation of ANN in predicting solid particle erosion in composites. Shamshirband S. et al. (2015)
have developed adaptive neuro-fuzzy inference system (ANFIS) to predict total and maximum erosion
rates in a 900 elbow. The results are promising when compared to CFD. This work majorly differs in the
methodology of using a hybrid approach compared to stand alone soft computing technique developed in
algorithm has been used widely in wear prediction (Zhang et al. 2003), and the Bayesian regularization
training algorithm was found to be more accurate for erosion modeling by Zhang et al. (2003), Danaher et
Erosion is a complex phenomenon that varies primarily as a result of material strength. In brittle
materials, erosion occurs by crack formation. Solid particles are hitting the surface form cracks, which
then propagate by subsequent impacts. Depending on the material hardness, when the impact load
exceeds a defined level, plastic defragmentation occurs along the fracture (Levy, 1995). Evans, Gulden,
and Rosenblatt (1978) explained how planar cracks form along the interface when the area is unloaded
after particles rebound from the surface and pieces are then removed by subsequent particle impacts.
Finnie (1960) presented one of the earliest analytical models to predict erosion in ductile materials. The
model assumes that erosion in ductile materials is a result of a micro-cutting process. In this process, the
surface material is removed in the form of cuttings. Particles that strike the surface at low impact angles
form a crater, and material is continuously removed by further impacts. This theory suffers from a
drawback in that it neglects the effect of particles hitting at a right angle. Finnie (1978) later corrected this
by modifying the model and considering the effect of surface material piling when particles impact the
surface at larger angles. A more detailed two-step model was presented by Tilly (1973), which stated that
first the impacting particles produce an indentation and may remove a chip. Subsequent impacts break up
the hole and project fragments radially from the primary site.
Particle impact velocity can be considered the single-most important factor causing erosion. Many
researchers conclude that the erosion rate (ER) is proportional to particle velocity (Vp) raised to the power
n.
Various values of n have been proposed. Finnie (1978) proposed two as the value of n. Smeltzer et al.
(1970) and Grewal et al. (2013) also found negative values for n by curve-fitting the experimental data in
some cases. More recently, Oka et al. (2005a, 2005b) proposed that the value of n was not a constant, but
a function of hardness. The impingement angle is also an important factor in determining erosion. At low
impact angles in ductile materials, cutting action and platelet formation is more prevalent, whereas
Erosion occurs in various applications. The particles in each of these applications vary in hardness, size,
shape, etc. This variation in particle properties also affects erosion rate and nature. Particle properties like
size, shape, and hardness have been observed to affect erosion rate. Oka et al. (2005a, 2005b)
incorporated constants in the model to account for the effect of particle shape and size. The University of
Tulsa’s Mclaury (1997, 1998, and 1999) and Ahlert (1994) developed models that incorporated shape
factors ranging from 0.2 for round particles to 1.0 for sharp particles. Levy (1983) concluded that angular
particles could cause up to four times more erosion than round particles. Tilly (1973) proposed a very
well-fitted empirical model that considers erosion rate as a function of particle size. Particle size, in the
form of a power law, was also included in the Oka et al. (2005a, 2005b) model. The effect of particle
concentration was explained regarding a shielding mechanism by Andrews and Horsfield (1983). The
sand volume concentration ranged from 0.38% to 8.61% in tests by Turenne et al. (1990). They
discovered that the erosion ratio decrease followed a power law of sand volume concentration.
Fluid properties, such as the effect of turbulence were studied by Pourahmadi and Humphry (1983, 1990).
Impingements occurred in regions of high turbulence due to higher momentum transfer to the particles.
As fluid density and viscosity increase, the drag force acting on the particle also increases. This
phenomenon can be directly related to the conclusion of research conducted by Smeltzer et al. (1970) that
indicated the erosion rate decreases at higher test temperatures. Clark and Burmeister (1992) proposed a
squeeze film model to account for the cushioning effect of a fluid boundary layer. Clark (1992) observed
a phenomenon of boundary layer filtration in which particles were deflected due to high normal gradients
in a flow.
The novelty of this work lies in the hybrid approach of combining CFD and machine learning models like
regression and ANN to build a predictive modeling pipeline for erosion rates. General purpose CFD
code, ANSYS Fluent is used to simulate fluid flow and particle transport for corresponding experiments
on a 900 elbow. Exploratory data analysis and correlation analysis was performed on eight CFD output
parameters to identify their significance in erosion rate predictions. Two new models based on shear
stress and turbulent kinetic energy is developed by curve fitting. Experimental data of erosion rates is
used as a target variable, and CFD output is used as input to develop ANN model. A detail of the
Computational fluid dynamics (CFD)-based erosion modeling has been employed and explored for many
years. CFD-based erosion models offer a low-cost, fast solution to the erosion modeling problem.
However, erosion modeling using CFD is a complex task because it involves various phenomena that
must be considered. Flow modeling, particle interaction with fluid, walls and other particles, fluid
properties, and erosion modeling are some of these phenomena. Also, geometric modeling and meshing
(discretization of the domain) are critical in the overall CFD-based erosion modeling procedure. The
biggest advantage of CFD-based erosion modeling over other simplistic models is that complex
geometries and flow fields can be easily modeled using CFD. It is a powerful tool that can predict erosion
regions accurately. Consequently, it presents a low-cost, qualitatively accurate and timely method to
predict erosion in various tools and geometries. The Erosion Corrosion Research Center (ECRC) at The
University of Tulsa has been one of the pioneers in CFD-based erosion prediction research. McLaury
(1993, 1996) proposed a widely followed CFD-based erosion prediction procedure. More recently, Zhang
et al. (2009) presented a procedure and improved guidelines for using commercial CFD code for erosion
prediction. The authors have also recently published work (Pandya, Dennis, and Russell, 2014) presenting
the effect of CFD modeling parameters on erosion prediction and a new improved CFD-based erosion
model. The work in this paper can be considered an extension of this previous work.
CFD-based erosion modeling consists of fluid flow modeling, discrete particle modeling, and erosion
ANSYS Fluent is a finite-volume CFD code that solves the Navier-Stokes equation to model fluid flow.
ANSYS automated meshing was used to discretize the domain. An unstructured mesh for flow region
with structured refinement on the walls is used. The wall y + values of around 30 are observed, and
scalable wall functions are used to simulate flow in near-wall regions. The realizable k-epsilon model was
used to model turbulence parameters. It has proved to be a better at capturing physics of localized
circulation in the elbow when compared with Standard k - epsilon. (Pandya, Dennis, and Russell, 2014).
Simulation is initiated with a first order discretization scheme and then switched to second order scheme
for pressure, Momentum, TKE, and dissipation rate. A SIMPLE scheme is used for pressure – velocity
coupling. The discrete phase (sand) was solved by tracking a large number of particles Lagrangian
approach after fluid field simulations were converged. In the Lagrangian approach, the fluid (water in this
case) is treated as the continuum, but the discrete phase (sand) is treated as single particles, where particle
trajectories, representing a stream of particles, are calculated as a result of forces acting on them. Water
is the primary phase, and the Sand particle is secondary phase in this multiphase simulations. The
secondary phase (sand) in our case had a very low concentration, so this approach was justified. The
effect of sand particles on fluid was considered negligible for low concentrations. This model gave
various output parameters such as particle velocity, particle impingement angle, wall shear stress, etc.
Depending upon which parameters were required in the empirical erosion model, erosion rates were
calculated as a post-processing step and displayed using custom field functions or user-defined functions
within ANSYS Fluent. A detailed modeling approach presented by authors (Pandya, Dennis, and Russell,
2014) was followed, the details of which are out of scope for this paper.
All analysis performed in this research was based on experimental results provided by Baker Hughes.
Tests were conducted at Baker Hughes for four different flow conditions considering two different water
flow velocities (i.e., 50 ft/sec and 85.8 ft/sec) and for two different particle sizes (i.e., 256 microns and 25
microns) (Russell,2004). Flow velocities of 50 ft/sec and 85.8 ft/sec correspond to Reynolds number of
3.83 x 105 and 6.5 x 105 respectively. The test was performed on a 90-degree elbow with a 1- inch inner
diameter. Inconel 718, Nickel Alloy 825, 25% Cr, Nickel Alloy 925, and 13% Cr L-80 material elbows
were placed in series. The sand concentration was approximately 1% by mass and 0.38% by volume
(Russell, 2004). The erosion rate was measured at 230 locations in each elbow. The experimental setup is
Mark McCasland et.al. (2004) and Ronnie Russell et. al. (2004)
Figure 3 Experimental setup for erosion experiments (Russell, 2004)
The ERC-2003 erosion model was developed by Baker Hughes in 2003. Ronnie Russell et al. (2004)
presented this model that was based on extensive experiments on a 90-degree elbow. A series of erosion
tests were performed with several materials and under various flow conditions and sand concentrations.
More than 95% of the impact angles were between 0-30 –degree with few around 90-degree in the
stagnation zone. A predictive model was then created based on the obtained results. The model is
The Baker Hughes ERC-2008 model offers some accuracy improvements over the Baker Hughes ERC-
2003 model for flow velocities less than 30 ft/sec (Russell and Marsis, 2013). Baker Hughes ERC -2008
Different data analysis techniques such as regression, exploratory data analysis, and artificial neural
networks were used in this work. MATLAB was used for all data analysis. A correlation analysis was
performed on outputs of CFD erosion models and experimental results. The parameters with high
correlation were identified for model generation. Regression for curve fitting and artificial neural network
models were developed by considering the erosion rate as a function of these identified parameters.
Two variables are said to be associated if the behavior of one affects the behavior of the other.
Correlation coefficients are measures of association. They assign a numerical value to the degree of
association or strength of the relationship between two variables (Gibbons, 1993). Several correlation
coefficients are proposed to measure a degree of correlation. Pearson correlation coefficient (Pearson,
1901), Spearman’s rank correlation coefficient and Kendall’s rank correlation coefficient (Kendall, 1948)
were obtained to establish the association between various independent parameters to erosion, the
dependent parameter in each case. A general interpretation based on the values of these coefficients
Pearson’s correlation is a measure of the linear correlation between two variables. It is also called
Pearson-product-moment correlations and is represented by ‘r’. Linear correlation may exist, even if
variables have a nonlinear relation to one another. Correlation varies from -1 to 1, where 0 means no
correlation, 1 means complete positive linear correlation, and -1 means complete negative linear
correlation.
encountered in this work do not have a strong linear or parametric relation to erosion, a nonparametric
approach is important to consider while performing correlation analysis. Spearman correlation can give a
perfect value (1 or -1) when the two variables are related by any monotonic function, in contrast to only a
Like Spearman’s rho, Kendall’s tau is also a nonparametric measure of correlation. It was first introduced
by Maurice Kendall (1948). Although in many cases, as in ours, the interpretation of Spearman’s rho and
Kendall’s tau is very similar, some people have argued that Kendall’s tau has an advantage of very direct
observation and interpretation, considering agreeable (concordant) and non-agreeable (discordant) pairs
(Bolboaca and Jantschi, 2006). Details about the correlation coefficients and their formula can be found in
Appendix A
The data are resampled by bootstrapping to reaffirm the inference made based on parameter values.
Bootstrap resampling has a great advantage of being a simple and straightforward way to derive
properties of estimators. It measures the properties from some resamples created by random data
sampling with replacement from the data set. The number of elements in each bootstrap sample is equal to
that in the original data set. It helps to determine how certain are the conclusions made on a parameter
value.
A variety of fits, algorithms, and methods are explored to find the best fit for the identified CFD
parameters. MATLAB curve fitting toolbox is used for this purpose (2013). For all the methods
implemented, input data is first centered and scaled to normalize the predictor (input) data. The input
variables have wide differences in scale in our data set. Consequently, normalizing the data improved the
fit. A weighted least-squares method is helpful when the assumption that the response data is not of equal
quality. The method improves fit by adding weight (scale factor) in the fitting process. The weight
determines how much the response value influences the parameter estimates (i.e., a high-quality data
point influences the fit more than a low-quality data point). The disadvantage of least squares as being
highly sensitive to outliers is rectified in the robust scheme. Two methods are available in MATLAB
Least-absolute residuals (LAR): To reduce the effect caused by squaring the absolute residuals of outliers,
LAR calculates the curve that minimizes the absolute difference. This reduces the influence of outliers on
the fit.
Bisquare weights: Weights are assigned to each data point based on its distance from the fitted line. Points
that are much farther from the line get a zero weight. This minimizes the weighted sum of squares. The
Bisquare method is preferred most of the time because it finds the curve that fits the majority of the data
A robust iteratively re-weighted least-square algorithm is proposed by Holland and Welsh (1977).
A detailed discussion about development and implementation of each of the methods and algorithms is
out of the scope of this work, MATLAB implementation of Robust fitting with Bisquare weights (2013)
is presented in Appendix B, and it can be obtained from respective reference. Its application in MATLAB
The basic building blocks of an ANN are network architecture, transfer functions, and training methods.
A multilayer network is a network with multiple layers with different weight and biases for each layer.
Each layer may have different network architecture. The layer that produces the final output of the
network is called output layer. The remaining layers are usually called hidden layers. A simple three-layer
network with an input layer, one hidden layer, and one output layer is depicted in Figure 3.
Figure 4 Schematic of a typical multilayer neural network model (Hagan et al., 1996)
Linear and hyperbolic tangent sigmoid transfer functions were used in this study. As the name suggests, it
algorithms (discussed later in detail) have been discussed by Vogl et al. (1988) and Harrington (1993).
The method of setting weights in an ANN enables the process of learning. Training is generally referred
to as the process of modifying the weights in the connection between network layers to achieve expected
output (Shivanandam at el. 2006). The process that takes place while the network is trained is called
learning. There are different types of training including supervised, unsupervised and reinforcement. The
back-propagation algorithm used in this study is a type of supervised training, where the network is
provided with sample inputs and the outputs are compared to expected responses. There are target output
vectors for comparing the sequence of training input vector. The weights are adjusted according to
different algorithms.
A feed-forward network with the back-propagation algorithm (Rumelhart and McClelland, 1986) is one
of the most commonly used networks. In this network, neurons are organized in layers and send their
signals "forward." The errors are then propagated backward. The back-propagation algorithm falls under
the category of supervised learning. The central idea for a back-propagation algorithm is that of
minimizing the error until the network is trained. As in many other networks, random weights are
assigned at the beginning of training, and they are adjusted until the minimization criteria for error is
reached. Hagan and Menhaj (1994) first demonstrated the development of the Levenberg-Marquardt
algorithm for neural networks. It has been reported to train the neural network 10 to 100 times faster than
the popular gradient descent back-propagation method. A Bayesian regularization algorithm to avoid
A detailed derivation of any of the methods is available in the literature mentioned and will not be
combination of squared errors and weights. This implementation is one of the approaches to stop over-
fitting a problem. It also reduces the need to test a different number of hidden neurons for a problem.
The typical performance function for a feed-forward network is a mean sum of squares of the errors.
∑
A term consisting of a mean sum of squares of the network weights is added to performance to improve
generalization.
Where α and β are parameters optimized in the Bayesian framework (MacKay, 1992). This results in
smaller weight and forces a smooth response. In this framework, the weights are assumed to be random
variables with a specified distribution. The regularization parameters are associated with the variance of
More details, proof, and derivation of the algorithms can be obtained in MacKay (1992) and Foresee and
Hagan (1997).
Before we discuss the results obtained by implementing various erosion-prediction models, it is important
to discuss the parameters used throughout results to measure efficiency and effectiveness of the models.
Following are quantities measured and their implications are related to results in this work.
The most frequently used quantity to measure the quantitative performance of the model throughout this
plotted. An ideal situation would be when this value is 1, which would mean that predicted value and
The sum of squares due to error, denoted by SSE, measures the total deviation of the predicted value form
Values closer to 0 mean the model has a smaller random error and is a better fit.
R-squared is the most widely used measure to determine a quality of fit for statistical models like
regression or neural network. It measures how well the model explains the variation in data. It is called
R-square is defined as the ratio of the sum of squares of regression (SSR) and the total sum of squares
Hence,
The root mean squared error, is commonly called the standard error of fit or standard error of regression.
It is a measure of estimation of the standard deviation of the random component in the data:
√ ∑
Correlation analysis was performed to identify parameters that most affect erosion rate. This analysis was
performed on 920 data points obtained from CFD analysis of all four test cases. The aim was to develop a
more robust erosion rate prediction model. Three different correlation coefficients were calculated.
Pearson coefficient ‘R' is most commonly used correlation coefficient. However, it evaluates the linear
correlation between parameters (Note: a linear correlation between parameters with the nonlinear
relationship is possible). Consequently, Kendal's ‘tau' and spearman's ‘rho,' both of which rank
correlations, were also calculated. Values closer to one depict strong correlations. Figure 4 shows
Correlation Coefficients
0.0229
Impact Angle 0.0189
0.025
Turbulent Kinetic 0.9
0.79
Energy 0.76
0.78
Surface Shear stress 0.73
0.72
0.49
Concentration 0.33
0.29
0.69
Velocity 0.63
0.62
A bootstrap resampling method is applied to calculate the correlation coefficient for 100 resamples. The
results further support our statistical inference of the correlation analysis mentioned above. Pearson’s
coefficient for each of these resampled data is calculated, and distribution is presented in Figure 6.
Figure 6 Bootstrapping: Resampling distributions of the Pearson’s correlation coefficients
The results agree well with the correlation analysis performed. It can be observed that turbulent kinetic
energy is identified as most influential parameter followed by surface shear stress, velocity, and
concentration. The impact angle does not have any conclusive correlation with erosion rate prediction. It
is also verified when overall performance of ANN was degraded by including impact angle as an input
Concentration has some positive correlation with the erosion rate. An analysis is performed to check if the
Discrete particle (DPM) –Sand concentration can be used as a conditioning parameter in velocity-based
erosion models. A co-plot for velocity, concentration and erosion rate presented in in Figure 7 shows that
for every range of concentration, a similar trend line is observed. Consequently, use of DPM
Surface shear stress has a strong positive correlation with the erosion rate, so curve fitting through
regression was performed to develop an erosion model as a function of surface shear stress. Two different
models were developed, a polynomial model and exponential model. Results are presented in Table 1.
Table 1 Shear stress model - goodness-of-fit parameters
R -square error
An exponential fit has a lesser percentage error of 42-percent mean error. A regression plot for an
erosion rate. Thus, as expected, the TKE-based model proved to be very efficient. Table 2 presents
R -square error
Three different models were developed including linear fit, fifth-degree polynomial fit, and an
exponential fit. Only exponential fit, which showed the best results, is presented in Figure 9.
of-fit parameter values. The R-squared value is 0.9972, which is one of the best seen so far. Also, the
mean percentage error is 29% which is excellent. It should also be observed that TKE model provides a
An ANN model was developed as a black box to predict the erosion rate. The four most influential
parameters, velocity, concentration, TKE and surface shear stress, were taken as an input parameter to the
ANN and the erosion rate was considered target variable. A multilayer feed-forward/back-propagation
training algorithm gave the best results. More than 45 different networks were modeled and analyzed. The
description and goodness-of-fit parameters for 11 different ANN models are presented in Table 3.
Table 3 ANN model description and goodness-of-fit parameters
Figure 10 shows the percentage error for each of the 11 networks described.
Percentage error
90
80
Percentage Error 70
60
50
40
30
20
10
0
0 2 4 6 8 10 12
ANN Model number (Refer Table 90 )
Input parameters for 4-[10-10-10]-1 are velocity, DPM concentration, surface shear stress and turbulent
kinetic energy and target are experimentally observed erosion rates. A network representation can be seen
A regression plot for ANN with three hidden layers and structure 4-[15-10-15]-1 with and without
Bayesian regularization is presented next. Bayesian regularization avoids over-fitting the model and
visible better fit is obtained by Bayesian regularization. The mean percentage error is reduced from 27%
to as low as 7% after regularization is forced on the Lavernberg-Marquardt training algorithm. The 5-[10-
10-10]-1 network represents the inclusion of impact angle as an input parameter, but it has a negative
Figure 12and Figure 13 show a graph, plotting measured and predicted values for all points without and
Experimental value
Predicted Value
Experimental value
To verify the reproducibility of the results by ANN , a bootstrap sampling on input parameter was
performed. A 100-sample input was created from the current data set. The mean percentage error is
50
40
30
20
10
0
1
6
56
11
16
21
26
31
36
41
46
51
61
66
71
76
81
86
91
96
Precentage Error
The best of each model was selected, and their prediction capability was compared regarding percentage
error (Figure 15) and error ratio (Figure 16). The Baker Hughes model was the base case for comparison.
Overall Percentage Error comparision
ANN 7.1
TKE 29
Shear stress 42
Baker Hughes 57
0 10 20 30 40 50 60
Mean percentage error
New
10
TKE Model
Upstream Downstream
0.01
0 2 4 6 8 10 12
Non dimensional point location (0-5 upstream, 5-12 downstream)
concentration and velocities were tested, but it was still the same elbow case. Consequently, there was no
change in geometry. Verification was necessary with a different geometric case. A hold out data
validation is vital for any machine learning approach. Experimental results for a test on a completion tool
were compared with the models developed. Due to proprietary data restrictions, the geometry of the tool
is not presented here. To give a comparative idea, flow conditions will be discussed. The erosion rate of
the tool was measured for 2% sand loading with 53-micron sand size and 66.47 ft/sec inlet velocity.
Erosion rate at 243 points was measured and compared with simulated results from the new Baker
Hughes model, the shear stress model, the turbulent kinetic energy model, and the ANN model.
Regression results are presented below in Figures 16. Mean percentage errors are represented in a graph
below.
Figure 17 Predicted vs. Experimental Erosion Rate (mm/hr.) for Model verification
It can be observed that in the verification case that each model performs better in comparison to elbow
erosion rate prediction. This can be attributed to the fact that the verification model geometry does not
have major flow path variations. The model has an essentially undisturbed streamline flow and is more
comparable to upstream points in case of the elbow. Even in the elbow, the mean error for upstream
points is much lower compared to downstream where complex flow structures like vortices are generated.
Conclusion
CFD best practice was followed to simulate fluid flow and particle transport in a 90-degree elbow.
Experimental results were considered as the target parameter, and three new models based on statistical
analysis of CFD output parameters were developed. Correlation analysis was performed on eight different
CFD output parameters to identify ones affecting erosion rate prediction the most. Velocity, surface shear
stress, DPM concentration and turbulent kinetic energy were found to have a significant correlation.
A new shear stress-based erosion rate prediction model was developed by implementing robust least-
square curve fitting. The new shear stress-based model had an average percentage error of 42%. A
turbulent kinetic energy-based erosion model was also developed, and it exhibited and error of 29%,
which is an improvement of about 20% from the baseline model of Baker Hughes ERC -2008.
Artificial neural network models were explored to act as a black box where CFD output was taken as
input parameter and erosion was the output. More than 45 different networks with a different number of
neurons in hidden layers as well as a different number of hidden layers were modeled. A multilayer feed-
forward network with a back-propagation algorithm was implemented for each model. Training
algorithms like Scaled Conjugate gradient and Levenberg-Marquardt were implemented. A Bayesian
approach gave the best results. The results from the ANN model were excellent, with the mean percentage
error of 7%, which is a further improvement of more than 20% points and 300% from the shear stress or
the TKE model. Overall, it brings down error from 57 percent to 7 percent compared to the baseline.
ANN’s have potential to be developed further as a tool to predict erosion rate. The ANN model was also
verified on a hold out data set consisting of CFD analysis on a completion tool. Developed model had not
seen the data from completion tool at all and hence served as a good candidate for model verification.
ANN model showed improvement in predicting erosion rates on completion tool when compared to the
elbow case. This is counter-intuitive, but the verification model has very simple flow path with no
circulation regions. It can be compared to upstream data points of the elbow where erosion rate prediction
capabilities were found to be more accurate when compared to downstream points. Consequently, the
All the models were developed based on experimentation on elbows. This is just one case, so care should
be taken not to generalize it. Experimentation should be performed on more complex tools that are used in
the oil and gas industry and which face daily erosion problems. This approach will help in the
For any statistical model, more data points lead to better models, and this can be achieved by more
experimentation.
Material properties of the tool are believed to have the great effect on erosion rate. Dependence of erosion
Lastly, although our studies show no significant effect of impact angle, other experimental studies have
found it to be a significant parameter affecting erosion (Finnie et al. 1992, Sheldon 1970). More effort
For paired x and y values in a sample, Pearson’s correlation coefficient r is given by:
r=
r=
r
( x X )( y Y )
( x X ) ( y Y )
2 2
Nonlinearity and outliners are a major factor that affects the value of r.
After the data are collected, it is ranked by giving highest rank to the highest value, and so on. Then the
difference in rank of each data set is calculated, and the values are inserted in the formula above.
Kendall’s Tau
The general formula for calculating the Kendall’s Tau is given below:
For a given sample of variables x and y, both with sample size of ‘n’, there are nC2 combinations possible
for selecting distinct pairs (xi., yi) and (xj., yj). These pairs are defined as concordant if (xi. > xj and yi >
yj ) or (xi. < xj and yi < yj ); discordant if (xi. > xj and yi < yj ) or (xi. < xj and yi > yj ) and neither if
xi = xj or xi = yj .
Then the number of concordant pairs and discordant pairs are inserted into the above formula to derive a
coefficient that lies between -1 to 1.
Appendix B – MATLAB implementation of Robust fitting with Bisqare
weights
Following steps are implemented for Robust fitting with Bisquare weights (2013).
ri is the usual least-square residuals and hi are the leverages that adjust the weight reducing weight of data
points that had significant effects on the least-square fit.
Where K is the tuning constant with a value 4.685 and s is the robust variance given by:
| |
| |
5. Stop if fit converges, else go back to step one for the next iteration.
A nonlinear least-squares method is employed to fit a model that has an equation with nonlinear
coefficients or is a combination of linear and nonlinear coefficients. Exponential and power function fits
were obtained using nonlinear methods.
y = f(X,β) +ε
where,
X is a n-by-m (matrix with n rows and m columns) design matrix for the model
An iterative approach is required to solve for the coefficients in a nonlinear model. The following steps
are performed by a MATLAB function to solve for the coefficients in a nonlinear least-square regression
for curve fitting (2013):
3) Adjust the coefficients following a fitting algorithm, and see if the fit improves. Two different
fitting algorithms are available in MATLAB:
a. Trust-Region – This algorithm was described by More and Sorensen (1983) and Cartis et
al. (2009). It is used if the constraints on coefficients are known.
b. Levenberg-Marquardt – This is one of the most popular algorithms used for decades.
Proposed by Levenberg (1944) and improved upon by Marquardt (1963), it has proven to
give a good fit for a wide range of nonlinear models.
4) Stop if the convergence criteria are reached, else return to step 2 for next iteration.
The steps of back-propagation for a connection between hidden layer neuron A and output neuron B are
explained in below (MacLeod, 2013):
1) Initialize random weights, apply inputs and calculate the outputs.
2) Calculate the error for neuron B. Remember that The transfer function is sigmoid function, so the
error is:
4) Calculate the error for the hidden layer neuron by back-propagating them from the output layer.
This is done by taking errors from output neurons and running them back through the weight to
get hidden layer errors. If a neuron is connected to two output neurons, B and C, then:
The training process for the Lavenberg-Marquardt algorithm is described in the following steps (Yu and
Wilamowski, 2012):
1) With the initial weights (randomly generated), evaluate the total error (SSE).
2) Perform the update as directed by the Levenberg-Marquardt update rule to adjust weights with the
new weights and evaluate the total error.
3) If the current total error increases as a result of the update, then retract the step (such as reset the
weight vector to the previous value) and increase the combination coefficient μ by a factor of 10
or by some other factor. Go to step ii and try an update again.
4) If the current total error decreases as a result of the update, then accept the step (such as keep the
new weight vector as the current one) and decrease the combination coefficient μ by a factor of
10 or by the same factor as step 4.
5) Go to step 2 with the new weights until the current total error is smaller than the required value.
5) Solve to find δ.
c. ⁄
Reference
2013 MATLAB Curve Fitting Toolbox Documentation. Version 3.4. Natick, Massachusetts: The
MathWorks Inc.
2013 MATLAB R2013b Documentation. Version 3.4. Natick, Massachusetts: The MathWorks Inc.
Ahlert, K.R. 1994. Effects of particle impingement angle and surface wetting on solid particle erosion of
Andrews, D. and Horsfield, N. 1983. Particle collisions in the vicinity of an eroding surface. Journal of
Andrews, D. and Horsfield, N. 1983. Particle collisions in the vicinity of an eroding surface. Journal of
Begg, R. and Palaniswami, M. 2006. Computational intelligence for movement sciences: neutral networks
Bolboaca, S. and Jäntschi, L. 2006. Pearson versus Spearman, Kendall's tau correlation analysis on
structure-activity relationships of biologic active compounds. Leonardo Journal of Sciences 5 (9): 179-
200.
Carpenter, G.A. and Grossberg, S. 1985. Category learning and adaptive pattern recognition: A neural
network model. Presented at the Proceedings, Third Army Conference on Applied Mathematics and
Curtis, C., Gould, N.I. and Toint, P.L. 2009. Trust-region and other regularizations of linear least-squares
Danaher, S., Datta, S., Waddle, I. and Hackney, P. 2004. Erosion modeling using Bayesian regulated
Demuth, H. and Beale, M. 1993. Neural network toolbox for use with MATLAB. Version 4. Natick,
Drukker, D.M. 2003. Testing for serial correlation in linear panel-data models. Stata Journal 3 (2): 168-
77.
Eskandari, H., Rezaee, M., and Mohammadnia, M. 2004. Application of multiple regression and artificial
neural network techniques to predict shear wave velocity from wireline log data for a carbonate reservoir
Evans, A., Gulden, M. and Rosenblatt, M. 1978. Impact damage in brittle materials in the elastic-plastic
response regime. Proceedings of the Royal Society of London.A.Mathematical and Physical Sciences 361
(1706): 343-65.
Finnie, I. and McFadden, D. 1978. On the velocity dependence of the erosion of ductile metals by solid
Finnie, I. 1960. Erosion of surfaces by solid particles. Wear 3 (2): 87-103. doi: 10.1016/0043-
1648(60)90055-7.
Finnie, I., Stevick, G. and Ridgely, J. 1992. The influence of impingement angle on the erosion of ductile
https://dx.doi.org/10.1109/ICNN.1997.614194.
Gibbons, J.D. 1993. Nonparametric measures of association: Sage University Paper series on
Grewal, H.S., Arora, H.S., Agrawal, A. et al. 2013. Slurry Erosion of Thermal Spray Coatings: Effect of
doi: http://dx.doi.org/10.1016/j.proeng.2013.12.210.
Hagan, M.T., and Menhaj, M.B. 1994. Training feedforward networks with the Marquardt
Harrington, P.d.B. 1993. Sigmoid transfer functions in backpropagation neural networks. Analytical
Holland, P.W., and Welsch, R.E. 1977. Robust regression using iteratively reweighted least-squares.
Humphrey, J. 1990. Fundamentals of fluid motion in erosion by solid particle impact. International
Kendall, M.G. 1948. Rank correlation methods, first edition. London: Charles Griffin & Co. Ltd.
Klopf, A.H. 1975. A comparison of natural and artificial intelligence. ACM SIGART Bulletin (52): 11-3.
Klopf, A.H. 1972 Brain function and adaptive systems: a Heterostatic Theory. Special Report No. 133
Lee Rodgers, J. and Nicewander, W.A. 1988. Thirteen ways to look at the correlation coefficient. The
Mathematics 2: 164-8.
Levy, A.V. 1995. Solid particle erosion and erosion-corrosion of materials, second edition, Materials
Levy, A.V. and Chik, P. 1983. The effects of erodent composition and shape on the erosion of steel. Wear
89 (2): 151-62.
MacLeod, C. 2013. The Back Propagation Algorithm In An Introduction to Practical Neural Networks
McCasland, M., Barrilleaux, M., Gai, H. et al. 2004. Predicting and mitigating erosion of downhole flow-
control equipment in water-injector completions. Presented at the SPE Annual Technical Conference and
Exhibition.
Marquardt, D.W. 1963. An algorithm for least-squares estimation of nonlinear parameters. Journal of the
McCulloch, W.S. and Pitts, W. 1943. A logical calculus of the ideas immanent in nervous activity. The
McLaury, B.S. 1996. Predicting solid particle erosion resulting from turbulent fluctuations in oilfield
McLaury, B.S. 1993 A model to predict solid particle erosion in oilfield geometries. M.S.. Thesis,
Production. Presented at the SPE Annual Technical Conference and Exhibition, Houston, Texas, 3-6
McLaury, B. and Shirazi, S. 1998. Predicting sand erosion in chokes for high pressure wells. Presented at
the SPE Annual Technical Conference and Exhibition. New Orleans, Louisiana, 27-30 September. SPE-
49308-MS http://dx.doi.org/10.2118/49308-MS
McLaury, B., Wang, J., Shirazi, S., Shadley, J. and Rybicki, E. 1997. Solid particle erosion in long radius
elbows and straight pipes. Presented at the SPE Annual Technical Conference and Exhibition. San
Mohaghegh, S. and Ameri, S. 1995. Artificial neural network as a valuable tool for petroleum engineers.
Moré, J.J. 1978. The Levenberg-Marquardt algorithm: implementation and theory. In Numerical analysis,
Moré, J.J. and Sorensen, D.C. 1983. Computing a trust region step. SIAM Journal on Scientific and
Oka, Y.I., Okamura, K. and Yoshida, T. 2005. Practical estimation of erosion damage caused by solid
particle impact: Part 1: Effects of impact parameters on a predictive equation. Wear 259 (1): 95-101.
Oka, Y. and Yoshida, T. 2005. Practical estimation of erosion damage caused by solid particle impact:
Part 2: Mechanical properties of materials directly associated with erosion damage. Wear 259 (1): 102-9.
Pandya, D., Dennis, B. and Russell, R. 2014. An Improved Computational Fluid Dynamics (CFD) Model
for Erosion Prediction in Oil and Gas Industry Applications. Presented at the ASME 2014 33rd
Edinburgh, and Dublin Philosophical Magazine and Journal of Science 2 (11): 559-72.
Pourahmadi, F. and Humphrey, J. 1983. Modeling solid-fluid turbulent flows with application to
Rosenblatt, F. 1958. The perceptron: a probabilistic model for information storage and organization in the
Rumelhart, D.E. and McClelland, J.L. 1986. Parallel distributed processing: explorations in The
Russell, R., Shirazi, D. and Macrae, J. 2004. A new computational fluid dynamics model to predict flow
profiles and erosion rates in downhole completion equipment. Presented at the SPE Annual Technical
http://dx.doi.org/10.2118/90734-MS
Russell, R., Barrilleaux, M., Gai, H. et al. 2004. Design, analysis, and full-scale erosion testing of a
downhole flow control device for high rate water injection wells. Presented at the SPE Annual Technical
Russell, R. and Marsis, E. 2013. A State-of-the-Art Computational Fluid Dynamics Simulation for
Erosion Rates Prediction in a Bottom Hole Electrical Submersible Pump. Presented at the SPE Heavy Oil
MS
Saputelli, L., Malki, H., Canelon, J. and Nikolaou, M. 2002. A critical overview of artificial neural
network applications in the context of continuous oil field optimization. Presented at the SPE Annual
Technical Conference and Exhibition. San Antonio, Texas, 29 September-2 October. SPE-77703-MS
http://dx.doi.org/10.2118/77703-MS
Shahab, M., Bogdan, B. and Samuel, A. 1997. Permeability determination from well log data. SPE
Shamshirband, S. 2015. "Performance investigation of micro- and nano-sized particle erosion in a 90°
Sheldon, G. 1970. Similarities and differences in the erosion behavior of materials. J BASIC ENG TRANS
Sivanandam, S., Sumathi, S. and Deepa, S. 2006. Introduction to neural networks using Matlab 6.0: Tata
McGraw-Hill Education.
Smeltzer, C., Gulden, M. and Compton, W. 1970. Mechanisms of Metal Removal by Impacting Dust
Souza, C. 2009. Neural Network Learning by the Levenberg-Marquardt Algorithm with Bayesian
Suresh, A., Harsha, A. and Ghosh, M. 2009. Solid particle erosion studies on polyphenylene sulfide
composites and prediction on erosion data using artificial neural networks. Wear 266 (1): 184-93.
Tilly, G.P. 1973. A two-stage mechanism of ductile erosion. Wear 23 (1): 87-96. doi: 10.1016/0043-
1648(73)90044-6.
Turenne, S., Chatigny, Y., Simard, D. et al. 1990. The effect of abrasive particle size on the slurry erosion
Velten, K., Reinicke, R. and Friedrich, K. 2000. Wear volume prediction with artificial neural networks.
Vogl, T.P., Mangis, J., Rigler, A., Zink, W. and Alkon, D. 1988. Accelerating the convergence of the
Werbos, P.J. 1988. Backpropagation: Past and future. Presented at the IEEE International Conference on
Yu, H. and Wilamowski, B. 2012. Neural Network Training with Second Order Algorithms. In Human–
Zhang, Y., McLaury, B.S. and Shirazi, S.A. 2009. Improvements of particle near-wall velocity and
erosion predictions using a commercial CFD code. Journal of fluids engineering 131 (3).
Zhang, Z., Barkoula, N., Karger-Kocsis, J. and Friedrich, K. 2003. Artificial neural network predictions
Highlights
Turbulent kinetic energy and shear stress based erosion models are established
CFD – Artificial neural network hybrid model to predict erosion rate is developed
Significant performance boost is obtained compared to baseline Baker Hughes model
Novel CFD – machine learning method for erosion prediction is presented