ISSN: 0974-8571 Vol. 13 No.

2, December, 2021

International Journal of Computational Intelligence in Control

Prediction of Sand-Casting Defects by Artificial

Neural Network
Anjul Rai and S.K.Ganguly
anjul.rai@bitdurg.ac.in , sapan.bit@gmail.com

Date of Submission: 24th November 2021 Revised:25th December 2021 Accepted: 30Th December 2021

Abstract - Sand casting defects appear unexpectedly and first layer. The output of this input layer will become the
prediction of defects can help decide the ranges of their input to the next hidden layer which consists of the number
responsible process parameters to reduce these defects. of neurons with some activation functions, the output of this
Artificial Neural Networks (ANN) is one of the famous hidden layer will become input to the output layer frame and
tools of Artificial Intelligence that can be utilised in a wide give the desired output (1).
extent of problems and are getting a wider application in
the foundry industries also. In this paper, a discussion The yearly casting production around the world from
about the prediction of three grey iron sand casting 2015 to 2019 has recognized remarkable growth (2), and
defects: Blow, Penetration and Erosion using ANN is India is the second-largest producer of castings just after
presented. This is a method of assessing the suitability of china (3). Metal casting is a versatile primary manufacturing
sand casting process parameters by finding correlations process that can produce products from a few grams to
between individual parameters and defects by means of several tons from molten metal (4). The sand casting
artificial neural network systems. The implementation of process generally means pouring molten metal into a
ANN was performed with the application of three refractory mould that contains a cavity of the required shape
programs first a statistical software SPSS, second another and allows it to solidify this final solidified object is called
software MATLAB and third using python programming casting. The sand-casting process is explained stepwise in
language. The data narrating to sand casting processing
Figure 1.
parameters and relevant defects in castings is raised from
a sand casting foundry. The ANN model gave better
results in terms of predicted accuracy. The systems are
fairly tested on real-life data acquired from the foundry,
attending to a significant decrease in the rejection rate of
the casting due to defects. The ANN prediction system can
help foundry engineers to take decisions regarding
preventing possible defects in advance.
The control of defects is an essential issue for the foundry
industry and utilized several Artificial Intelligence
Index Terms - Sand casting, Blow holes, Penetration, algorithms and proven their fruitfulness for making
Erosion, Artificial Neural Networks. decisions, by training AI models using past available data.
The defects in foundry can be significantly decreased by
predicting its occurrence and regulating the appropriate
process parameters through the utilization of comprehensive
INTRODUCTION
domain knowledge. In work for the prediction of three grey
Artificial Neural Networks is one of the most popular iron sand casting defects: Blow, Penetration and Erosion the
algorithms in Artificial Intelligence and is the functional ANN is implemented using statistical softwires SPSS,
unit of deep Learning. ANN is a computing system that is MATLAB and using python programming language. The
designed to simulate the way the human brain analyses and data required for training and testing of ANN model is
processes information. ANN is a network of artificial collected from a foundry for grey cast iron component of
neurons (or nodes) arranged in different layers and (FG 250, Composition: C 3.3-3.7, Si 2.2-2.6, Mg 0.65-0.67,
connections between the nodes. The first layer is the input S 0.00-0.04, P 0.03-0.05, Induction furnace) overall size of
layer next multiple hidden layers the last layer is known as 1000 × 550 × 113 mm and weighs 200 kg; its wood pattern
the output layer. The input is provided to the neurons of the

Copyrights @Muk Publications Vol. 13 No. 2, December, 2021

International Journal of Computational Intelligence in Control

199
 Prediction of Sand-Casting Defects by Artificial Neural Network

weighs 27kg. The yield is between 70 and 75 %. The based models and witnessed that the neuro-fuzzy model was
detailed drawing of the component is shown in figure 2. more precise than the neural network model. (11) Indicated
that the Radial Basis Function Artificial Neural Network
(RBFANN) model could adequately estimate the parameter
values of phosphate graphite sand. To determine the
relationship between input-output parameters of cement-
bonded moulding system (12) used propagation algorithm
and a genetic algorithm, based neural networks and proved
it can be used for both forward and reverse mapping of input
and output parameters. (13) apply back-propagation neural
networks and genetic neural networks for both forward and
reverse modelling of input and output parameters of green
sand moulding after that (14) adopted back-propagation as
well as genetic neural network for determining the
Figure 2. Sand casting part considered in the study
relationship between input-output parameters of sodium
silicate bonded carbon dioxide hardened sand moulding.
The foundry typically carried out 4–5 heats (cycles of (15) Presented a method of selection of the proper kind of a
melting and pouring) on each working day, each heat being neural network for prediction a sand moistness on the bases
used to pour 4-5 castings. For each heat, 7 data items were of certain moulding sand properties such as: permeability
identified for collection included Moisture percentage, and friability. To analyse the influence of input variables on
Permeability, Binder Percentage, Pouring Temperature, compressive strength of Furan No-bake mould system (16)
Grain Fineness Number, Mould Hardness, Green Strength formed ANN modal. Artificial Neural Network has also
and assurance of the three defects. The data from 493 heats been applied in the foundry industry to control melting
(approximately 1590 castings) were collected over six processes in furnaces. (17) presented Artificial Neural
months. The sample dataset of 494 data entry which Network-based model to determine the active bentonite
content in green moulding sand based on sand properties
contains the total seven parameters for evaluation propose
such as permeability, compactibility and the compressive
from that we can predict the output variables like blow,
strength, and letter (18) applied Neural Networks for
Penetration and Erosion. Each record has a class value that controlling the quality of bentonite moulding sand. ANN
indicate whether the casting suffered any defect or not has been used in the sand-casting industries for defect
within six-month measurements. identification, quality assessment of the castings, prediction
of possible defects, selection and control of sand mould
The created sand-casting data has certain characteristics properties and for optimization of casting input parameters.
that makes the analysis very challenging and attractive as
well. The next Literature review about ANN in the realm To summaries Artificial Neural Networks have been used
of metal sand casting defect reduction is discussed. to carry out input-output modelling of the metal sand
casting process. That Artificial Neural Networks are
excellent in the sand-casting quality and mould quality
LITERATURE REVIEW predictions and for optimization of casting input parameters.
Next in this work discussion about data collection from sand
To support sand foundry production researchers have casting foundry is presented.
adopted different types of Neural Networks differently.
Now a review about the relevance of ANN in foundries for ARTIFICIAL NEURAL NETWORK AND SAND-
casting defect reduction is presented. For cause detection of CASTING DEFECT PREDICTION
gas porosity defects in sand casting (5) trained ANN using
process parameters, materials used and workers involved. To see whether the input parameters can predict the
(6) Applied a backpropagation neural network to predict defects blow, penetration and erosion using the available
some significant defects in sand castings. For shrinkage data, an artificial neural network (ANN) is built using SPSS
location recognition(7) trained back propagation neural program (a statistical software). A Multilayer perceptron
network with the results of Finite Element Methods.(8) neural network is built in SPSS, with the seven input
Used artificial neural network-based methodology to the variables. The information about the network is given in the
identification of flaws in the aluminium alloy castings by table 2.
radio graphical image analysis. (9) Offered an Artificial
Neural Network based model to foresee collapsibility of
CO2 sand moulds. To assess the green compressive strength
Table 2 Network Information
of clay bonded moulding sand (10) offered neuro-fuzzy
Input Covariates 1 x1
Copyrights @Muk Publications Vol. 13 No. 2, December, 2021
International Journal of Computational Intelligence in Control

200
 Anjul Rai, S. K. Ganguly and R.L. Shrivastava

Layer 2 x2 Average Overall Relative Error .104

3 x3 Blow .102
4 x4 Relative Error for Scale
Penetration .102
5 x5 Dependents
6 x6 Erosion .109
7 x7 1 consecutive step(s)
Number of Unitsa 7 Stopping Rule Used with no decrease in
Rescaling Method for Covariates Standardized errora
Number of Hidden Layers 1 Training Time 0:00:00.20
Hidden
Number of Units in Hidden Layer 1a 8 Sum of Squares Error 1.851
Layer(s)
Activation Function Sigmoid Average Overall Relative Error .141
1 Blow
Testing Blow .133
Dependent Variables 2 Penetration Relative Error for Scale
3 Erosion Penetration .132
Output Dependents
Number of Units 3 Erosion .152
Layer
Rescaling Method for Scale Dependents Normalized
Activation Function Sigmoid The accuracy of the prepared ANN model is checked by
Error Function Sum of Squares
R_Squared values because it is a regression model. R2 is a
statistical measure of how close the data are to the filled
The network is having seven input variables in the input
regression line. The R2 values range from 0 to 1. From these
layer one hidden layer with eight nodes listed in the table.
values, one can estimate how well the model will perform
The network architecture is shown in figure 3. Which is
while doing the predictions. The R2 value of the ANN
having the topology of 7-8-3.
model is shown in table 4. The R2 value of regression for
the Blow prediction is 0.89, indicating 89% variance in the
data was analysed and was explained by the ANN model,
similarly, the R2 value of regression for the penetration
prediction is .893 and R2 value of regression for the Erosion
prediction is .882. So, the model can be used for the
prediction of blow, penetration and erosion defects for the
given input parameter values.

for blow (R2 = 0.890)

for penetration (R2 = 0.893)
for erosion (R2 = 0.882)

Figure 3 Network Diagram with topology 7-8-3

Error computations based on the training sample of the Fig. 4 R2 values for Blow, Penetration and Erosion
overall modal is represented in table 3. The sum of square
error of the training data is 3.880 and for testing data is
1.851 which shows the Built neural network is performing Nest residual of the model is estimated. Residual is the
well. difference between the observed value and the predicted
value. The values for a good predictive model should have
less residual or error. The residue charts are shown in the
Figure 5. For Blow, penetration and erosion. It can be seen
there is random scatter in all the data and the residue is near
TABLE 3. MODEL SUMMARY about zero for the predicted values of blow, penetration and
Training Sum of Squares Error 3.880
Copyrights @Muk Publications Vol. 13 No. 2, December, 2021
International Journal of Computational Intelligence in Control

3
201
 Prediction of Sand-Casting Defects by Artificial Neural Network

erosion.
Nest residual of the model is estimated. Residual is the
difference between the observed value and the predicted
value. The values for a good predictive model should have
less residual or error. The residue charts are shown in the
Figure 5. For Blow, penetration and erosion. It can be seen
there is random scatter in all the data and the residue is near
about zero for the predicted values of blow, penetration and Figure 6. Network topology
erosion.
For training of the ANN Levenberg Marquard Training
algorithm is used. The validation performance of the neural
network is displayed in figure 7 the error was changing for
different iterations starting from zeroth iteration, the best
validation performance is 0.0061801 at the tenth epoch.

Figure 5. Regression for prediction of Blow, Penetration

and Erosion Figure 7 Validation performance of ANN Model

The validation check is shown in figure 8. Here validation

SAND-CASTING DEFECT PREDICTION USING check is at 16 epochs. It menaces at 16th epoch it has a
ARTIFICIAL NEURAL NETWORK better performance.
Here MATLAB is used to create an ANN model and to
see whether the input parameters can predict the defects
blow, penetration and erosion using the available data. The
procedure of developing the ANN model is discussed here.
First, the dataset is loaded into MATLAB. The total size
of the data is 493*10, as it is having 493 observations and
seven independent variables (moisture%, permeability,
Binder%, Pouring temperature, GFN, Mould Hardness and
Green strength) and three independent variables (Blow,
Penetration and Erosion). Then data is separated into
training and testing datasets. 70% of data is divided for
testing and 15% of data is used as the validation data set.
The remaining 15% of the data will be used for testing the
developed model. The network architecture of the build
ANN model is shown in figure 6. It has 7 neurons as there Figure 8. Validation check
are 7 input variables in the dataset, and 3 neurons in the
output layer. And there are 10 neurons in the hidden layer, The amount of error is shown in the output error histogram
figure 9.
so the network topology is 7-10-3.

Copyrights @Muk Publications Vol. 13 No. 2, December, 2021

International Journal of Computational Intelligence in Control

202
 Anjul Rai, S. K. Ganguly and R.L. Shrivastava

discussion about implementation of ANN sand casting

defect prediction using the same data is presented.

SAND-CASTING DEFECT PREDICTION

Using the same data one more ANN model is
implemented using python and its in-built packages. The
implementation process is discussed next stepwise. First all
the required libraries like pandas, numpy, sklearn along with
keras is imported. Keras is basically a frame work in python
that allows to implement ANN models. Next the data is
visualised to see the target and impacting variables, shown
in figure 11.

Figure 9. Error histogram

To check the performance of the ANN model the

regression plots of the model is shown in figure 10. For
training data set R=0.95429, for the validation dataset,
the performance of the model is 0.95077 and for the
testing dataset the performance of the model is 0.95747 (b)
(a)
and for all datasets, the performance of the model is
0.95417.

(d)
(c)

(f)
(e)

Figure 10. R Values of Training, Validation, and

Testing and for All data set.

As the performance the model is very good and can be

used for prediction of sand-casting defects for the given
input parameter values. Hance both the models
prepared by SPSS and MATLAB can be used to predict
the sand casting defects: Bow, penetration and erosion.
ANN can also be implemented using python, next

Copyrights @Muk Publications Vol. 13 No. 2, December, 2021

International Journal of Computational Intelligence in Control

5
203
 Prediction of Sand-Casting Defects by Artificial Neural Network

(g) occur at given level of input parameters. And actually, the

Figure 11. Relative frequency distribution of variables, (a) moisture, (b)number of times the penetration occurred were 387.
permeability, (c) binder, (d) pouring temperature, (e) grain finesse number, (f)
mould hardness, (g) green strength
TABLE 5. PERFORMANCE OF THE ANN MODEL FOR PENETRATION
Heat No Prediction for Actual Performance
After visualising the data, it is prepared for modelling the penetration penetration of the model
whole data is spited into the ratio of 80:20 for training and
DM001 Penetration Penetration 0
testing the model. After that the model is trained and build 1.00
DM070 Penetration No penetration
the input, hidden and output layer activation function,
DM111 Penetration Penetration 1.00
weights and bias are defined in block of cods. Keras is used
to build the model sequentially layer by layer. Next step is DM231 Penetration Penetration 1.00
to compile the mode means to keep the model ready to DM284 Penetration Penetration 1.00
consume the data. After compiling the model run it is up to DM401 No penetration No penetration 1.00
100 epochs in the batch size of 20. In every epoch the DM434 Penetration Penetration 1.00
weights and bises get adjusted and model accuracy - - - -
improves. The model accuracy started from 0.624 and after No penetration No Penetration 0.00
DM493
the weights adjustment and going back and forth the model
Total 413 387 453.00
gave 95% accuracy. Here we defined keras model and then
called the keras and got 95% accuracy in test data. Hance
ANN easily implemented using python also in a very simple To see the performance of the ANN mode for the
way and the performance of the model is very good. prediction of erosion all the 493 predictions and actual data
are tabulated and compared to determine the accuracy of the
RESULT AND DISCUSSION model on the bases of its correct predictions shown in Table
6. The model predicted 335 times erosion can occur at given
To see the performance of the ANN model for the level of input parameters. And actually, the number of times
prediction of Blow all the 493 predictions and actual data the erosion occurred were 331.
are tabulated and compared to determine the accuracy of the
TABLE 6. PERFORMANCE OF THE ANN MODEL FOR EROSION
model on the bases of its correct predictions shown in Table
4. The model predicted 264 times blow can occur at given Heat Prediction Actual Performance
level of input parameters. And actually, the number of times No for erosion erosion of the model
the blow occurred were 276. DM053 No erosion No erosion 1
DM195 No erosion No erosion 1
TABLE 4. PERFORMANCE OF THE ANN MODEL FOR BLOW DM200 Erosion No erosion 0
Heat No Prediction for Actual Performance DM253 No erosion No erosion 1
Blow Blow of the model DM297 Erosion Erosion 1
DM006 No Blow No Blow 1.00 DM340 Erosion Erosion 1
DM067 Blow Blow 1.00 DM369 Erosion Erosion 1
DM106 No Blow No Blow 1.00 DM412 erosion Erosion 1
DM121 No Blow No Blow 1.00 - - - -
DM374 No Blow No Blow 1.00 DM493 No erosion No erosion 1
DM409 No Blow No Blow 1.00 Total 335 331 488.00
DM410 No Blow No Blow 1.00
DM432 No Blow Blow 0.00 The computed accuracy of the model for blow,
DM462 Blow Blow 1.00 penetration and erosion is summarized in table 6. As the
- - - - performance of the ANN model for all three defects is quite
DM493 Blow Blow 1.00 satisfactory. Model prediction for erosion is best with the
Total 264 276 416.00 accuracy of 98.99% followed by prediction of prediction
and then blow is 83.16%.
To see the performance of the ANN mode for the
TABLE 6 PERFORMANCE OF BAYESIAN-BASED SAND-CASTING PREDICTION
prediction of penetration all the 493 predictions and actual SYSTEM
data are tabulated and compared to determine the accuracy Blow Penetration Erosion
of the model on the bases of its correct predictions shown in Actual no of 276 387 331
Table 5. The model predicted 413 times prediction can
Copyrights @Muk Publications Vol. 13 No. 2, December, 2021
International Journal of Computational Intelligence in Control

204
 Anjul Rai, S. K. Ganguly and R.L. Shrivastava

defects in Data [9] Ohdar RK, Pushp PT. Prediction of collapsibility of moulds and cores
of CO 2 sands using a neural network. J Eng Manuf. 2003;217:475–
Predictions by the 265 413 335 87.
model [10] Babu NN, Ohdar RK, Pushp PT. Evaluation of green compressive
Accuracy 84.38% 91.89% 98.99% strength of clay bonded moulding sand mix: neural network and
neuro-fuzzy based approaches., Int J Cast Met Res,2006,19(2),
PP.110-114
Hance ANN easily implemented using python also in a very [11] Cai A, Chen H, Tan J, Chen M, Li X, Zhou Y, et al. Optimization of
simple way and the performance of the model is very good Compositional and Technological Parameters for Phosphate
Graphite Sand. J Mater Eng Perform. 2008;17(4):465–71.
so the model can be used for prediction of blow, penetration [12] Parappagoudar MB, Pratihar DK, Datta GL. Modelling of input–
and erosion defects for the given input parameter values. output relationships in cement bonded moulding sand system using
neural networks. Int J Cast Met Res. 2007;20(5):265–74.
CONCLUSION [13] Parappagoudar MB, Pratihar DK, Datta GL. Forward and reverse
mappings in green sand mould system using neural networks. Appl
ANN models have been used to carry out input-output Soft Comput J. 2008;8(1):239–60.
[14] Parappagoudar MB, Pratihar DK, Datta GL. Neural network-based
modelling of the metal sand casting process. Nowadays approaches for forward and reverse mappings of sodium silicate-
different ways are available to build the ANN models and bonded, carbon dioxide gas hardened moulding sand system. Mater
use them in the wared range of problem domains. It can be Manuf Process. 2009;24(1):59–67.
[15] J. Jakubski and Dobosz SM., Selected parameters of moulding sands
used for defect prediction and quality assurance for the grey for designing quality control systems, Arch Foundry Eng.
iron sand casting using different statistical software like 2010;10(3):11–16.
SPSS and MATLAB. And it can be implemented using [16] Acharya SG, Vadher JA. Parametric Analyses on Compressive
Strength of Furan No Bake Mould System Using ANN. Arch Foundry
python also in very simple steps. All the ANN models Eng. 2016;16(4):5–10.
presented in this work gave better results in terms of the [17] Jakubski J, Dobosz SM, Major-Gabryś K. Active Binder Content as a
predicted accuracy of the defects. This type of prediction factor of Control System of the Moulding Sand Quality. Arch Foundry
Eng. 2011;12(4):49–52.
system can help foundry engineers to take decisions [18] Jakubski J, Malinowski P, Dobosz STM, Major-Gabryś K. ANN
regarding preventing possible defects in advance. Modelling for the Analysis of the Green Moulding Sands Properties.
conclusion section is usually required. Although a Arch Metall Mater. 2013;58(3):961–3.
conclusion may review the main points of the paper, do not
replicate the abstract as the conclusion. A conclusion might
elaborate on the importance of the work or suggest AUTHOR INFORMATION
applications and extensions.

ACKNOWLEDGMENT Anjul Rai is an assistant professor in mechanical

engineering at Bhilai Institute Technology, Durg,
The authors thanks the foundry "Abha Power & Steel Pvt. Chhattisgarh, India. She holds a bachelor degree in
Ltd." located at Hardikala Silpahari Industrial Area, mechanical engineering from Guru Ramdas Khalsa Institute
Bilaspur, Chhattisgarh, India consortium members for their of Science & Technology, Jabalpur and master degree in
support and data for this project. computer integrated manufacturing from Shri Govindram
Seksaria Institute of Technology and Science, Indore. Her
REFERENCES research interests are in sand casting, CNC, Advanced
[1] S. R. Peter and Norvig, “Artificial Intelligence: A Morden Approach, manufacturing systems and Artificial Intelligence.
6th ed, Prentice Hall: 2010, ch. 3, pp. 121-123
[2] Modern Casting Staff Report, “Census of World Casting Production - S. K. Ganguly is a professor in mechanical engineering at
Total Casting Tons Dip in 2019”. American Foundry Society.
January 2021; pp. 28–31.
Bhilai Institute Technology, Durg, Chhattisgarh, India. He
[3] Modern Casting Staff Report, “50th Annual Census of World Casting holds a bachelor degree in mechanical engineering from
Production”, American Foundry Society. January 2016, pp. 25–28. Government Engineering College, Raipur (now National
[4] Ravi B. “Metal casting - computer aided design and analysis”, 6th Institute of Technology, Raipur), master degree in
ed. New Delhi: P.H.I. Learning Private Limited; 2008.
[5] Perzyk M, Kochański A. Detection of causes of casting defects
production engineering from Visvesvaraya Regional
assisted by artificial neural networks. Proc Inst Mech Eng Part B J College of Engineering, Nagpur (now Visvesvaraya
Eng Manuf. 2003;217(9):1279–84. National Institute of Technology) and PhD from National
[6] Karunakar DB, Datta GL. Prevention of defects in castings using Institute of Technology, Raipur in mechanical engineering.
back propagation neural networks. Int J Adv Manuf Technol.
2008;39(11–12):1111–24. His research interest is in advanced manufacturing systems,
[7] Zhang P, Xu Z, Du F. Optimizing casting parameters of ingot based quality engineering, measurement system analysis. A
on neural network and genetic algorithm. Proc - 4th Int Conf Nat number of PhD candidates had graduated under his
Comput ICNC 2008,4:545–8. supervision.
[8] Dobrzański LA, Krupinski M, Sokolowski JH. Computer aided
classification of flaws occurred during casting of aluminum. J Mater
Process Technol. 2005;167(2–3):456–62.

Copyrights @Muk Publications Vol. 13 No. 2, December, 2021

International Journal of Computational Intelligence in Control

7
205