Professional Documents
Culture Documents
www.springerlink.com/content/1738-494x
DOI 10.1007/s12206-013-0111-1
(Manuscript Received March 23, 2012; Revised August 22, 2012; Accepted October 30, 2012)
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Abstract
A model was proposed for optimization of stereolithography (SLA) process parameters to achieve the minimum shrinkage of H-
shaped parts. A neural network was designed to correlate the input parameters to dimensional error of the parts manufactured by SLA.
For this purpose, the data of a previous study from the literature was used that investigated the effect of three important parameters (layer
thickness, hatch overcure and hatch spacing) of the SLA process by measuring the H-shaped parts manufactured by SLA 250. Then, the
neural network model was imported into two optimization algorithms (genetic algorithm and simulated annealing) and the optimal values
were determined. Results showed that the combination of neural network and optimization algorithms could determine the optimal input
parameters for the minimum shrinkage with good accuracy.
Keywords: SLA; Neural network; Shrinkage; Optimization; Genetic algorithm; Simulated annealing
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
accuracy, we may have reasonable results. Fig. 3. MSE vs. number of neurons for two different transfer functions.
Table 2. MSE and epoch number for different transfer functions and
hidden layers with training method of Levenberg-Marquardt.
Table 3. Optimization results by GA with different GA parameters. Table 4. Parameters of GA optimization which led to the best result.
Stall generation
Creation function Constraint dependent
Crossover frac-
Migration frac-
Population size
HO ( mm)
HS (mm)
LT (mm)
Scaling function Rank
tion
tion Objective
function Selection Function Stochastic uniform
value Elite count 2
Reproduction
20 0.8 0.2 50 0.01064 0.109 -0.1 0.181 Crossover fraction 0.8
20 0.8 0.2 100 0.01064 0.109 -0.1 0.181 Mutation fraction Constraint dependent
20 0.9 0.2 100 0.01052 0.132 -0.044 0.25 Crossover function Scattered
20 0.9 0.3 100 0.01533 0.116 0.026 0.25 Direction Forward
30 0.8 0.2 100 0.01040 0.1 -0.038 0.25 Migration Fraction 0.2
30 0.9 0.2 100 0.01040 0.1 -0.038 0.25 Interval 20
30 0.8 0.3 100 0.01040 0.1 -0.038 0.25 Constraint Initial penalty 10
40 0.8 0.2 100 0.01064 0.109 -0.1 0.181 parameters Penalty factor 100
50 0.8 0.2 100 0.01064 0.109 -0.1 0.181 Generations 150
60 0.9 0.3 100 0.01040 0.1 -0.038 0.25 Stopping criteria Stall generations 100
60 0.8 0.2 100 0.01064 0.109 -0.1 0.181 Function tolerance 1.00E-06
60 0.8 0.2 100 0.01052 0.132 -0.043 0.25
30 0.8 0.2 100 0.01040 0.1 -0.038 0.25
30 0.8 0.2 50 0.01041 0.1 -0.039 0.25 lations for each generation and minimum stall generations. It
30 0.8 0.2 100 0.01040 0.1 -0.038 0.25 was reasonable to adopt 50 for stall generation number, but in
some cases, there was little improvement after the 50th gen-
eration and it was safer to adopt 100.
using the outputs of ANN as follows: There are some other operators which influence the GA per-
formance and are assumed constant in our optimization. Table
F= (dH − top)2 + (dB − top)2 + (dWaist)2 + (dAnkle)2 + (dLateral)2 4 shows a complete set of important GA options which led to
(1) the best results in this research. Considering the minimum
value for the objective function (= 0.01040), the optimal val-
where the values of dH-top, dB-top, dWaist, dAnkle and dLat- ues for SLA setup parameters are as follows:
eral are obtained from the output of ANN. Also, the ranges of
process parameters used to obtain the optimized values are the Optimal LT = 0.1 mm
same as those previously mentioned in Table 1. Optimal HO = -0.038 mm
Optimal HS = 0.25 mm
4.2.1 Optimization results using GA
In this phase, we tried to obtain the optimal results from GA Here, in the optimal case we can see that LT is at its mini-
by examining many options. We changed population size, mum acceptable value, HO is near its minimum and hatch
crossover fraction, migration fraction and stall generation. spacing is at its maximum acceptable value. These results are
Table 3 shows the results obtained in each case. It is obvious in accordance with the other researchers’ outcomes which
that in some cases GA is trapped in local minima and the final mention that LT and HO have negative effect and HS has
objective function value is far from the minimum value and positive effect on the total dimensional inaccuracy of SLA
even increasing the population size does not help us overcome parts [1-5, 9].
this drawback. However, in most of the trials, GA was reliable It is also interesting that a very near optimal result is ob-
and the convergence speed was so high that before reaching tained in some cases. The value of objective in this near opti-
50 generations, the best result was obtained. mal case is 0.0106 (just 0.0002 higher than the best result) but
It is almost impossible to be certain about the best options the SLA setup parameters are quite different (LT = 0.109; HO
for GA optimization, considering plenty of operators that in- = -0.1; HS = 0.181 mm). In this case, the value of LT is near
fluence its accuracy and convergence speed, but from the re- minimum, HO value is absolute minimum and hatch spacing
sults gathered in Table 3, we can conclude that the most pre- is near maximum. This result is also compatible with previous
cise and reliable mode may be as follows: population size = 30, studies and shows there may be near optimal results that can
crossover fraction = 0.8, migration fraction = 0.2 and stall help in the situations with special limitation. For example,
generation = 100. These values for GA operators were manufacturing time may be reduced in this near optimal case
adopted and corresponded to the case with the minimum value by increasing the LT from 0.1 to 0.109 mm.
for objective function with high repeatability, minimum calcu-
862 E. R. Khorasani and H. Baseri / Journal of Mechanical Science and Technology 27 (3) (2013) 857~863
( mm)
Initial
temp.
(mm)
(mm)
SLA process. In this paper, based on previous works done in
HO
Objective
HS
LT
Start point
function this area, we investigated the effect of three important SLA
value setup parameters (layer thickness, hatch overcure and hatch
[0.175,-0.025,0.15] 100 0.01062 0.1 -0.033 0.25
spacing) on the shrinkage and distortion of H-shaped parts
[0.125,-0.075,0.1] 100 0.01053 0.1 -0.034 0.249
manufactured by SLA250, and to find the optimal values that
[0.15,-0.05,0.2] 100 0.01052 0.102 -0.039 0.25
would lead to the minimum dimensional error in SLA parts.
[0.15,-0.1,0.2] 100 0.01053 0.103 -0.039 0.25
Since an analytical approach to solve this problem is very
[0.15,0.025,0.2] 100 0.01053 0.1 -0.042 0.249
difficult due to the complexity of the process, a multilayer
[0.15,-0.1,0.1] 100 0.01052 0.102 -0.039 0.25 perceptron was used to precisely and reliably relate three input
[0.1,-0.1,0.1] 100 0.01040 0.1 -0.038 0.25 parameters to five outputs (H-shaped parts measured dimen-
[0.1,0,0.1] 100 0.01040 0.1 -0.038 0.25 sions). After establishing a well-trained neural network (using
[0.25,0.05,0.05] 100 0.1052 0.1 -0.039 0.249 140 sets of inputs and outputs), it was possible to accomplish
[0.2,0,0.15] 100 0.1053 0.102 -0.039 0.25 optimization phase using GA and SA methods. Based upon
[0.2,-0.025,0.2] 100 0.01040 0.1 -0.038 0.25 these optimization results, the following conclusion can be
[0.175,-0.025,0.15] 150 0.01058 0.1 -0.032 0.25 realized:
[0.15,-0.05,0.2] 150 0.01058 0.105 -0.032 0.25 The proposed neural network (with one tansig hidden layer
[0.125,-0.075,0.1] 200 0.01046 0.1 -0.035 0.25 and 23 neurons) is found to reliably and precisely model the
[0.175,-0.025,0.15] 100 0.01062 0.1 -0.033 0.25 dimensional error (shrinkage + distortion) in SLA process.
The optimal values for SLA setup parameters using both
GA and SA methods were equal (layer thickness = 0.1; hatch
Table 6. Constant values used for SA optimization. overcure = -0.038; hatch spacing = 0.25) which can be con-
firmed by the results of previous works done by different re-
Annealing function Fast annealing
searchers. They mention that layer thickness and hatch over-
Reannealing interval 100
cure have negative effect and hatch spacing has positive effect
Temperature update function Exponential
on the total dimensional inaccuracy of SLA parts.
Max Function evaluations 9000 In the present study, both of the optimization algorithms
Function tolerance 1.00E-06 used (GA and SA), were suitable and found the optimal result.
Stall iterations 1500 GA worked fast and exactly, but in some cases was trapped in
local minima. It is essential to increase the number of trials
with GA to be certain of achieving the best fit point. SA
showed that it is a slow but precise and reliable method and
4.2.2 Optimization results using SA not easily trapped in local minima.
In the optimization with SA method, we tried different start Finally, it is considerable the presented model is only valid
points with various initial temperatures. Since the results were for SLA of H-shaped parts. To model the other shapes, first
very close and precise, it was not necessary to vary other SA the critical characteristics should be defined and measured,
operators. Table 5 shows the optimization results with differ- and then the optimization methods can be applied.
ent start points. Also, Table 6 shows the SA constant values
used in all trials.
Here, the best result is equal to the previous optimization
References
using GA. The interesting point is that, referring to the results [1] J. G. Zhou, D. Herscovici and C. Chen, Parametric process
in Table 5, SA comes to the final answer in much higher itera- optimization to improve the accuracy of rapid prototyped
tion numbers, but it is not trapped in local minima and all of stereolithography parts, International Journal of Machine
the results are in the same range with very little difference. In Tools and Manufacture, 40 (1999) 1-17.
fact, the final value for the objective function is just 0.0002 [2] S. H. Lee, W. S. Park, H. S. Cho, W. Zhang and M. C. Leu,
higher in the worst case. Besides, it is obvious that SA is not A neural network approach to the modelling and analysis of
sensitive to changing the start point or even the initial tem- stereolithography processes, Proceedings of Institution of
perature. Also, the near optimal result found by GA is not seen Mechanical Engineers Part B, 215 (2001) 1719-1733.
here that shows how powerful and reliable the SA method is. [3] S. Jayanthi, M. Keefe and E. P. Gargiulo, Studies in stereo-
These results confirm the optimal value for SLA setup pa- lithography: Influence of process parameters on curl distor-
rameters obtained in the previous optimization by GA. We can tion in photopolymer models, Proceedings of the Solid Free-
say that SA is good for researchers who are seeking a slow, form Fabrication Symposium, USA (1994) 250-258.
repeatable and precise method. [4] P. Minetola, The importance of a correct alignment in con-
E. R. Khorasani and H. Baseri / Journal of Mechanical Science and Technology 27 (3) (2013) 857~863 863
tactless inspection of additive manufactured parts. Interna- [9] T. H. Pang, M. D. Guertin and H. D. Nguyen, Accuracy of
tional Journal of Precision Engineering and Manufacturing, stereolithography parts: mechanism and modes of distortion
13 (2) (2012) 211-218. for a ‘letter-H’ diagnostic part. In Proceedings of the Solid
[5] B. S. Raju, U. Chandrashekar, D. N. Drakshayani and K. Freeform Fabrication Symposium, USA (1995) 170-180.
Chockalingam, Determining the influence of layer thickness [10] S. H. Lee, Optimal selection of process parameters and in
for rapid prototyping with stereolithography (SLA) process, process measurement of cross-sectional shapes of products
International Journal of Engineering Science and Technol- in stereolithography process. MS thesis, Korea Advanced In-
ogy, 2 (7) (2010) 3199-3205. stitute of Science and Technology, Korea (1998).
[6] S. L. Campanelli, G. Cardano, R. Giannoccaro, A. D. Ludo- [11] S. S. Rao, Engineering optimization: theory and practice,
vico and E. L. J. Bohez, Statistical analysis of the stereo- John Wiley & Sons, New Jersey (2009).
lithographic process to improve the accuracy. Computer-
Aided Design, 39 (1) (2007) 80-86.
[7] G. C. Vosniakos, T. Maroulis and D. Pantelis, A method for Hamid Baseri is an assistant professor
optimizing process parameters in layer-based rapid prototyp- of mechanical engineering in Babol
ing. Proceedings of Institution of Mechanical Engineers university of technology in Iran. He has
Part B, 221 (8) (2007) 1329-1340. ten years experience in manufacturing
[8] A. Gebhardt, E-manufacturing based on additive manufac- research. The subjects of his interest
turing technology (rapid prototyping, rapid tooling, rapid include machining process and intelli-
manufacturing), Aachen, Germany (2005). gent technologies.