You are on page 1of 11

Process Optimization of Injection Molding

Using an Adaptive Surrogate Model With


Gaussian Process Approach

Jian Zhou, Lih-Sheng Turng


Polymer Engineering Center, Department of Mechanical Engineering,
University of Wisconsin-Madison, Madison, Wisconsin 53706

This article presents an integrated, simulation-based erence handbooks, and later adjusted and fine-tuned by
optimization procedure that can determine the optimal trial-and-error on the shop floor. This approach is highly
process conditions for injection molding without user dependent on the experience of molding operators and
intervention. The idea is to use a nonlinear statistical
regression technique and design of computer experi- can be costly and time consuming, especially with a new
ments to establish an adaptive surrogate model with resin and/or new applications.
short turn-around time and adequate accuracy for sub- With advances in numerical modeling and computer
stituting time-consuming computer simulations during simulation techniques, there have been tremendous efforts
system-level optimization. A special surrogate model made to develop computer simulation tools to facilitate
based on the Gaussian process (GP) approach, which
has not been employed previously for injection molding injection molding design and process setups [2]. These
optimization, is introduced. GP is capable of giving commercial and research tools are generally capable of
both a prediction and an estimate of the confidence predicting the flow phenomena in a complex geometry,
(variance) for the prediction simultaneously, thus pro- thus helping engineers gain process insights, pinpoint
viding direction as to where additional training samples potential design problems, and make rational decisions
could be added to improve the surrogate model. While
the surrogate model is being established, a hybrid [3–5]. However, going through the vast amount of com-
genetic algorithm is employed to evaluate the model to puter-generated data manually to qualify the design and
search for the global optimal solutions in a concurrent process setups could become a daunting task, which
fashion. The examples presented in this article show requires numerous iterations of CPU-intensive evaluation
that the proposed adaptive optimization procedure of simulations and an understanding of the complex, non-
helps engineers determine the optimal process condi-
tions more efficiently and effectively. POLYM. ENG. SCI., linear relationships among the design and process parame-
47:684–694, 2007. ª 2007 Society of Plastics Engineers ters as well as objectives. All of these problems make it
difficult to effectively identify the optimal design and pro-
cess conditions for injection molding.
INTRODUCTION The purpose of this study is to develop an integrated
Injection molding, which is suitable for the mass pro- CAE optimization system that can intelligently, automati-
duction of complex plastic parts with precise dimensions, cally, and adaptively determine the optimal process condi-
represents a major polymer processing method and tions for injection molding in a timely fashion without
accounts for one-third of all plastics processed [1]. During user intervention. To facilitate the optimization process
the injection molding process, many process conditions, for injection molding while avoiding an excessive number
such as melt temperature, mold temperature, ram speed, of numerical iterations using computationally intensive
pack/hold pressure and duration, cooling time, etc., have simulations, an alternative and more effective approach is
to be properly set to ensure the quality of the molded adopted in this study. The idea is to use Gaussian process
components. At present, the various process conditions (GP), a nonlinear statistical regression technique, and
are often selected by experienced engineers based on prior design of computer experiments to establish a surrogate
experience, resin supplier’s recommendations, and/or ref- model that can substitute tedious simulations while using
minimum computational resources and intelligently select-
ing new sampling points in the design space, so that a
Correspondence to: Lih-Sheng Turng; e-mail: turng@engr.wisc.edu large amount of process (or design) alternatives can be
Contract grant sponsor: National Science Foundation; contract grant num- evaluated in a reasonable timeframe during a system-level
ber: DMI-0323509. Contract grant sponsor: 3M Precision Optics, Inc.
DOI 10.1002/pen.20741
optimization process.
Published online in Wiley InterScience (www.interscience.wiley.com). The rest of this article is organized as follows. The
V
C 2007 Society of Plastics Engineers second section presents the literature reviews for the opti-

POLYMER ENGINEERING AND SCIENCE—-2007


mization of injection molding including a discussion of tives and the settings of design and process parameters.
the current optimization procedures for injection molding They also evaluated and compared various local and
and the desired improvements. The third section proposes global optimization algorithms in terms of computational
a new adaptive optimization procedure based on the GP efficiency and effectiveness for injection molding [8, 9].
surrogate modeling approach. A brief mathematical de- Sadeghi developed a neural network model for predicting
scription of GP and a flowchart of the new optimization the quality of injected molded parts based on important
procedure are introduced. A more thorough performance process and material variables. This approach used a
analysis is documented in fourth section where the pro- backpropagation artificial neural network (ANN) that was
posed procedure is applied to the optimization of process trained based on data taken from CAE simulation. The
conditions in injection molding. Finally, the last section system can greatly reduce the time required for optimiz-
gives the conclusions. ing process conditions [10]. Zhou and Turng proposed a
CAE optimization platform that can quickly determine the
optimal processing conditions for injection molding. This
ANALYSES OF CURRENT OPTIMIZATION
approach uses support vector regression (SVR) to estab-
PROCEDURES FOR INJECTION MOLDING
lish a surrogate model (or surrogate models in the case of
To date, quite a few researchers have developed and multiobjective optimization) to approximate the CPU-in-
employed different optimization schemes for selecting or tensive 3D simulations. GA is then implemented to evalu-
determining the optimal design and process parameters ate the surrogate model(s) for searching for the single or
for injection molding. These optimization schemes can be multiple optimal solutions, respectively. In addition, the
classified into two categories. The first is using an optimi- performance and capabilities of using different modeling
zation technique coupled with complete CAE simulation. approaches to establish the surrogate model, such as artifi-
For instance, Park and Kwon developed a computer-aided cial neural networks (ANNs), polynomial regression (PR),
optimal design system to improve the performance of a and SVR, were also investigated [11]. Castro et al. used
cooling system for injection molding by minimizing a an approach composed of computer simulation, ANN, and
weighted combination of the uniformity of the part tem- data envelopment analysis (DEA) to determine the proper
perature and the cooling time. They implemented the operating conditions for multiobjective optimization of
CONMIN algorithm to obtain the optimal injection mold- injection molding. They also discussed the identification
ing cooling system design with the special boundary inte- of robust efficient process and design settings for optimiz-
gral formulation and the corresponding design sensitivity ing the injection molding process [12].
analysis formulation, in which the constraints were dealt The above-mentioned surrogate model optimization
with via the augmented Lagrangian multiplier (ALM) procedure can be summarized and illustrated in Fig. 1. In
method, and the Davidon-Fletcher-Powell method was particular, given the optimization objectives for injection
used for successive unconstrained optimization [6]. Deng molding, the first step is to select a physical-based simula-
et al. applied the genetic algorithm (GA) to optimize tion program, and then assume a regression model, such
injection molding process conditions, such as melt tem- as RSM, ANN, or SVR, etc., which hopefully is suited to
perature, mold temperature, and injection time, with user- accurately represent the desired response. According to
definable objective functions. They also investigated the the selected model, the sampling points are typically gen-
effective GA attributes suitable for process conditions erated through a conventional design of experiments
optimization [7]. However, in these optimization strat- (DOE). Considering the large number of DOE cases, the
egies, every objective function fitness value was evaluated analysis that needs to be executed could be computation-
with a complete CAE simulation. Considering the large ally expensive. Once all of the cases are analyzed, proper
number of objective functional evaluations during the parameter settings for the preliminary model are then
optimization process, these strategies have been regarded determined. To validate the resulting assumed model,
as too time-consuming to be practically useful because of additional test cases need to be run to evaluate the predic-
the extensive runtime of simulations. tive performance of the model. On the basis of the error
Considering the limitations of the first category, the observed with the validation data, a decision is made as
other approach employs modeling methods to establish a to whether the initially assumed regression model is
surrogate model (also called metamodel) based on CAE appropriate. If it is, then the acceptable surrogate model
simulations, and thus substitute the simulations with the can be employed by an optimization algorithm to search
surrogate model in the optimization procedure. In particu- for the optimal solutions. If, however, the evidence shows
lar, Turng and Peic presented a CAE optimization tool that the model does not achieve the required predictive
that couples a process optimization program (OPTIMUS) performance, the proper way to improve the surrogate
with an injection molding simulation program (C-MOLD) model is to assume a new regression model and go
to determine the optimal design and process variables for through the entire process again.
injection molding. In OPTIMUS, a polynomial type of A critical assessment of the current surrogate model
response surface methodology (RSM) is used to determine optimization procedures reveals that there are four desired
and quantify the relationship between the design objec- main improvements:

DOI 10.1002/pen POLYMER ENGINEERING AND SCIENCE—-2007 685


FIG. 1. Current optimization procedure for injection molding using surrogate models.

1. From a statistical point of view, the present implemen- estimate of the confidence of the model prediction
tation for building the surrogate model is solely based along with the prediction itself.
on pure data. Essentially, there is no guidance for 4. The current optimization scheme is normally executed
selecting the regression model. Since selecting a model in a sequential fashion. The optimization algorithm can
with little or minimal information brings an increased be applied only after a surrogate model is generated
risk of making inaccurate regression model selection, with satisfactory prediction ability. Therefore, an
it is desirable to eliminate the requirements of assum- adaptive surrogate model approach that allows concur-
ing a regression model a priori. rent execution of the surrogate model building and
2. Typically speaking, computer simulation is a kind of optimization is desirable.
‘numerical’ experiment. More representative training
data give the surrogate model better approximation ac-
curacy. As training samples grow, however, the effi- OPTIMIZATION PROCEDURE WITH GP
ciency of sample data collection becomes very impor- SURROGATE MODELING APPROACH
tant because of extensive simulation runtime require-
ments. While suitable for classical physical experiments, In addressing the desired improvements to the current
conventional DOE may not perform well for highly optimization procedure pointed out previously, a novel
nonlinear regression models. Therefore, to generate optimization system for injection molding using the GP
proper samples for building a robust surrogate model, surrogate modeling approach was proposed and devel-
it is desirable to employ a special DOE scheme suita- oped. It has the following characteristics: (1) It does not
ble for computer simulation. require a specific model to be assumed a priori; (2) It uses
3. To verify the model selection, an additional randomly
specific DOE for computer simulation, which can per-
generated validation sample set is needed to assess the
predictive performance of the surrogate model at the
fectly capture the general behaviors of injection molding
end of model creation. If the current model fails to simulation with a relative small number of training sam-
yield sufficiently accurate representations, it needs to ples; (3) It is capable of giving both a prediction and an
be redefined, necessitating a repeat of the whole pro- estimate of the confidence (variance) for the prediction
cess, and an increase in the number of DOE cases to simultaneously, thus providing direction as to where addi-
be executed. Thus, it is desirable to have an inherent tional training points could be added to improve the sur-

686 POLYMER ENGINEERING AND SCIENCE—-2007 DOI 10.1002/pen


process, the P(tN|CN,{xN}) is assumed to follow a Gaus-
sian distribution as given by:

1
PðtN j CN ; fxN gÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð2pÞN jCN j
 
1 T 1
 exp  ðt  mÞ CN ðt  mÞ ð1Þ
2

where tN ¼ (t1(x1), t2(x2),. . ., tN (xN)), CN is the covari-


ance matrix for P(tN|{xN}), which composes of hyperpara-
meters needed to be optimized, and m is the mean which
will be zero for properly normalized data.
FIG. 2. Structure of the optimization procedure. Once the GP model is trained, the optimal values of the
hyperparameters are then used to construct the covariance
matrix, which, in turn, is inverted and used to give predic-
tions for both the predicted response and the predicted var-
rogate model; (4) It enables a continuous adaptive process iance simultaneously through the following formulas:
in which the procedure will optimize the current version
of the surrogate model, adaptively update this surrogate tNþ1 ¼ kT C1
^ N tN (2)
model whenever new data becomes available, and search
for the optimal solution with guidance.
s^2tNþ1 ¼ k  kT C1
N k (3)

where k ¼ (C(x1 xNþ1), C(x2 xNþ1),. . ., C(xN xNþ1)), and


GP for Regression k ¼ C(xNþ1, xNþ1). For the details of the GP method, the
The GP method is a highly nonlinear regression tech- reader is referred to Refs. 13–15.
nique based on the Bayesian probability and inference
approach. Taking into account prior information that is
Implementation of the Adaptive Surrogate Modeling
available from the data, and assuming the associated error
and Optimization Procedure
to be normally distributed, the GP regression methods find
the most likely value as well as the variance for the On the basis of the GP method, a novel adaptive opti-
desired response directly through the statistical approach. mization system is implemented (cf. Fig. 2). The proce-
A GP regression model is concerned with evaluating dure requires the user to specify the optimization criteria
the probability P(tNþ1|D,xNþ1), where the new input vec- in terms of an objective function and the constraints,
tor is xNþ1 and the corresponding output is tNþ1. In this which can include any input variables for the computer

FIG. 3. Training stage of the optimization procedure using LHS and GP.

DOI 10.1002/pen POLYMER ENGINEERING AND SCIENCE—-2007 687


FIG. 4. Iteration stage of the optimization procedure.

simulation. This allows a variety of optimization objectives, input design space with N dimensions. LHS is useful
such as a shorter cycle time, temperature uniformity, when a simulation scheme needs a sample set that is ran-
reduced residual stresses, minimal part shrinkage and warp- dom but relatively uniformly distributed over each dimen-
age, or some combination thereof, depending on the user’s sion [16, 17].
requirements. In the meantime, various constraints, such as To perform the optimization, a novel two-step hybrid
clamp force, shot volume, or maximum injection speed and optimization method that incorporates the GA with a local
pressure, can be specified [6]. It should be pointed out that, optimization technique, the Hooke-Jeeve pattern search
due to the generic nature of this optimization scheme, any algorithm, is employed to evaluate the surrogate model to
of the commercial or research-oriented simulation programs search for the global optimal solutions concurrently with
can be incorporated into this procedure. the GP surrogate model. The Hooke-Jeeve pattern search
In this proposed procedure, the Latin hypercube sam- algorithm is implemented to improve the convergence
pling (LHS) method, which is a stratified random sam- performance [11, 18–20]. Because of an adaptive approxi-
pling technique, is used for generating the samples needed mation of the optimization problem, by starting with a
for initial training of the surrogate models. LHS allows coarsely approximated surrogate model, the hybrid GA
exceedingly sparse M point samples to be selected for an will proceed in the right direction to the near-optimal area

688 POLYMER ENGINEERING AND SCIENCE—-2007 DOI 10.1002/pen


and approach the optimal solution with generations using
the successively updated surrogate model.
The new optimization procedure with the GP modeling
approach will intelligently identify the points with a large
variance in the design space and select those as the next
sampling points to adaptively update and improve the sur-
rogate model. The optimization procedure includes the
following two stages: the training stage (cf. Fig. 3) and
the iteration stage (cf. Fig. 4). Either the convergence of
the surrogate model or the convergence of the optimal so-
lution can be used as the termination criteria. FIG. 5. Setup of the threshold for the predicted variance.
Training Stage (cf. Fig. 3)
dation index, rmajor, and then set up a time variant ran-
dom major validation criterion, pmajor. If rmajor , pmajor,
1. Determine the inputs (process conditions)/outputs (objec- the corresponding samples will be selected to be eval-
tive values). uated with the simulation and will act as the additional
2. Perform the DOE for the computer simulation over the training samples to update the surrogate model. The
expected feasible design space. Each set of input pa- other samples in the temporary dataset will be moved
rameters generated by the LHS method will be ana- to the candidate sampling dataset (cf. Fig. 6).
lyzed by CAE simulation. d. Additional simulation will be run with the remaining
3. Execute simulation programs for all of the cases gener- samples from the candidate sampling dataset. This is
ated by LHS; collect the results and extract useful data done by allocating a random number between [0, 1]
for building the preliminary surrogate model. to each of the samples in the candidate sampling
4. Fit a GP surrogate model to the simulation results. dataset as the minor validation index rminor, and then
setting up a time variant random minor validation
The GP model generated from the current stage pro- criterion, pminor. If rminor < pminor, the corresponding
duces predictions for both the response and the variance samples are also added to the training dataset to
of the response simultaneously. This predicted variance improve the surrogate model (cf. Fig. 7).
could be used as an estimate of the accuracy of the pre- e. Perform the optimization on the current surrogate
diction and for validating and updating the preliminary model. The optimal sample will also act as the addi-
surrogate model. tional training sample and be used to improve the
surrogate model.
Iteration Stage (cf. Fig. 4) A major aspect of the new 3. Update the surrogate model with the new training samples.
modeling approach in this iteration stage is the inclusion 4. Evaluate the termination criteria. If the criteria are not
of a runtime surrogate model improvement capability met, repeat the process or stop after a predetermined
derived from the Bayesian nature of the GP. This is done number of iterations.
by using the following steps:
OPTIMIZATION APPLICATIONS AND RESULTS
1. Access the current surrogate model with the randomly ANALYSIS
generated samples within the design space.
2. Determine the data set to update the surrogate model:
a. Perform a feasibility check of the randomly gener- Process Conditions Optimization for an Optical Precision Lens
ated samples according to the constraints. If the In this example, the injection molding process of com-
sample is known to be infeasible, discard this one
plex optical precision lenses (cf. Fig. 8) is optimized
from the current dataset.
b. Setup a time variant random threshold. If the pre-
dicted variance of the sample is greater than the
threshold value, i.e., s2 > threshold, the correspond-
ing sample is extracted into a temporary dataset.
This temporary dataset contains the samples that,
once simulated, would greatly improve the surrogate
model. However, due to the large number of sam-
ples in the temporary dataset and the CPU-intensive
simulation, not all of the samples in this dataset are
simulated. Only a portion of the samples will be
selected as described in the next step. The other
samples with smaller variances are grouped into a
candidate sampling dataset (cf. Fig. 5).
c. Allocate a random number between [0, 1] to each of
the samples in that temporary dataset as the major vali- FIG. 6. Setup of the major validation percentage.

DOI 10.1002/pen POLYMER ENGINEERING AND SCIENCE—-2007 689


Minimize maximum volumetric shrinkage ð%Þ
Subject to 220 C  Tm  250 C
60 C  Tw  80 C
6 s  tinj  10 s
(4)
15 s  tpack  30 s
70%  ppack  100%
10 s  tcool  40 s
93%  V=P by volume  99%

The packing pressure values in percentages are speci-


FIG. 7. Setup of the minor validation percentage. fied in relation to the maximum capacity of the injection
pressure. The convergence of the optimal solutions (e <
0.0005) is used as the termination criteria.
To train the GP model, 15 process condition combina-
through the proposed GP optimization scheme. As shown tions generated by the LHS technique are used initially
in the figure, the outside diameter of the lens part is for building the surrogate models. The GP model is
96.19 mm and the height of part at the center is 19.87 trained using the covariance function shown below [15]:
mm. The thickest part of the lens is at the outer rim of
the lens that is 10.50 mm, while the thickness at the cen- " #
L ðxðlÞ  xðlÞ Þ2
ter of the part is 6.00 mm. The weight of the full shot 1X i j
Cðxi ; xj Þ ¼ y1 exp  þ y2 (5)
lens part is 69.8 g. 2 l¼1 rl2
In this application, the maximum volume shrinkage is
minimized by optimizing seven independent process pa- where the hyperparameters y ¼ (y1, y2, r) correspond to
rameters, namely, melt temperature Tm, mold temperature the following: y1 controls the overall vertical scale of the
Tw, injection time tinj, packing pressure ppack, packing variance relative to the mean of the GP; y2 sets the bias
time tpack, cooling time tcool, and the velocity/pressure of the correlation as in an intercept; and r allows a differ-
switch-over (V/P) by volume. The Moldflow software, ent distance measure for each input dimension. Perform-
which can quickly provide volumetric shrinkage predic- ing the hybrid GA to maximize the logarithmic likelihood
tion [21], is selected for simulation to obtain estimates estimation, the optimal values of the hyperparameters are
for the objective values. The output maximum volumet- achieved and given in Table 1.
ric shrinkage as a function of the seven inputs is mod- After the training stage, the preliminary surrogate
eled using GP. The material used in this application is model will be adaptively updated and improved during
Rohm & Haas PMMA Plexiglass V-825. The optimi- the iteration stage. Using the GP surrogate model, each
zation problem in this application is now defined as objective function evaluation takes less than 3 s on a Dual
follows: 2.66 GHz Xeon workstation. In the first iteration stage,
500 validation samples that could potentially be evaluated
by the GP model are randomly generated within the
design space, and the variance threshold is set as s2 >
0.5. Accordingly, 14 points among the 500 points are
qualified as the additional samples for updating the model
due to their large predicted variance, and each point is
given a random number as the index rmajor. After choos-
ing the major validation percent as pmajor ¼ 5%, 3 of the

TABLE 1. Optimal values for the hyperparameters.

y1 2.911
y2 3.786
r1 0.941
r2 0.992
r3 0.212
r4 0.238
r5 1.008
r6 0.742
r7 1.082
FIG. 8. Optical lens part molded with injection molding.

690 POLYMER ENGINEERING AND SCIENCE—-2007 DOI 10.1002/pen


TABLE 2. Number of training samples, parameters, and the resulting objective values for each of the GP iteration step.

Iteration parameters Additional


Training (threshold variance, training samples Objective
samples pmajor, and pminor) (major/minor) values

Initial step 15 0.5, 5%, 0.5% 3/1 8.357


Step 1 20 0.1, 5%, 0.5% 3/1 8.255
Step 2 25 0.05, 2%, 0.2% 2/1 8.179
Step 3 29 0.01, 1%, 0.1% 1/1 8.134
Step 4 32 0.005, 1%, 0.1% 1/1 8.116
Step 5 35 — — 8.112

14 points are selected and will act as the additional train- tive capability with a reasonably small amount of compu-
ing samples, while the remaining 497 points compose the tational resources.
candidate sampling set, in which the minor validation per-
cent is set as pminor ¼ 0.5%. As a result, one point is
extracted from the candidate sampling set into the addi- Shrinkage and Warpage Optimization for a Box Part
tional training set. At the same time, the hybrid GA is
used to evaluate this preliminary GP model to find the In this application, the shrinkage and warpage (S&W)
current ‘optimal’ process condition, and this will also act of a polypropylene (PP) box part (cf. Fig. 10) in four
as the additional training sample. Thus, after the first-iter- directions will be optimized with the proposed adaptive
ation step, additional 5 simulations will be performed and procedure. The walls perpendicular to the x-direction are
the corresponding data will be collected to improve the 1.3 mm thick and the walls perpendicular to the y-direc-
overall predictive capability of the GP surrogate model. tion have four stepwise changes in thickness of 3.5, 3,
Following the above procedure, after five iterations, the 2.5, and 2 mm. The overall outside dimensions of the part
iteration stage stops as the termination criterion is reached. are 7.4 cm by 9.8 cm by 4.5 cm. Considering such a com-
The GP model that satisfied the termination criterion is built plex part design with thickness difference in x and y
with overall 35 training samples. Table 2 shows the corre- directions and variations in the y-direction and the semi-
sponding model samples in each of the iteration step. crystalline PP resin, noticeable and statistically significant
Table 3 tabulates the optimal process conditions S&W is produced over a wide range of process condi-
searched by the adaptive optimization scheme, and Fig. 9 tions. The optimization results will be compared with
shows the convergence processes using the hybrid GA to conventional injection molding experiment data to demon-
evaluate the final surrogate model. strate the capabilities of the optimization procedure. For
To verify the optimal process conditions, simulation is the details of the experiment setup, data, and result analy-
performed using the optimal conditions. The maximum sis, the reader is referred to [22]. The objective function
volumetric shrinkage, obtained through simulation under in this application can be expressed as:
the optimized process conditions, is 8.119 (cf. Table 3). Minimize shrinkage & warpage S&W
The error between the predictions of the surrogate model
ðj  xj þ j þ xj þ j  yj þ j þ yjÞ
and the simulation is relative small (0.086%), which ¼
confirms that the trained surrogate model can predict the 4
simulation results quite well. With the help of this auto- Subject to : 3 s  tpack  6 s
mated and adaptive integrated optimization procedure, the 20 s  tcool  35 s (6)
overall optimization task can be accomplished in hours 40 mm=s  vinj  100 mm=s
(7–8 h in this application) instead of days as required
40 MPa  ppack  60 MPa
by the conventional optimization approaches [11]. There-
fore, it shows that the trained GP model, as a substitution 205 C  Tm  230 C
for complete simulations, can achieve satisfactory predic- 20 C  Tw  40 C

TABLE 3. The optimal process conditions and objective values predicted by the surrogate model and simulation at the end of five iteration steps.

Process conditions Optimum solution


thold tcool tinj ppack Tmelt Tmold V/P Objective value Average error %

GP surrogate model þ hybrid GA 17.4 s 12.7 s 10 s 75.2% 221.48C 72.58C 94.5%


GP surrogate model þ hybrid GA 8.112 —
Moldflow simulation result 8.119 0.086%

DOI 10.1002/pen POLYMER ENGINEERING AND SCIENCE—-2007 691


the successive optimization studies based on CAE simula-
tions.
To establish the initial surrogate model, 40 sets of pro-
cess conditions are generated with the LHS technique. Af-
ter executing all of the simulations, the sample data are
extracted to establish the initial GP model, in which Eq. 5
is used as the covariance function. The convergence of
the surrogate model (e < 0.005, not the convergence of
the optimal solutions as used in the previous application
example) is used as the termination criteria. Following the
proposed procedure, after seven iterations, the termination
criterion is reached; and the GP model that satisfies the
termination criterion is built with 73 training samples
overall. Table 4 shows the corresponding GP model data
for each of the iteration steps.
This adaptive model is further validated over the previ-
ously mentioned 32 verification simulation results to esti-
FIG. 9. The convergence processes of optimization using hybrid GA. mate the prediction accuracy. The model predicted S&W
values (in terms of distance deviated from the designed
straight walls) of the four directions are compared with
A commercial injection molding simulation program, the simulation results of verification samples in Fig. 11.
Moldex3D [23], is used in this application to evaluate the In this figure, the solid line represents the best fit whereas
objective functions. The validation of the simulation pro- the dotted line corresponds to the unity slope (where the
gram by comparing the numerical predictions of S&W for model matches with the sampling points). Good agree-
the test box part with the experimental measurements can ment is seen between the model predictions and the simu-
be found in Ref. 11. After executing 32 (261) simula- lation results (samples), which confirms that the predictive
tions under the various process conditions that correspond ability of model construction is adequate. When achieving
to the fractional factorial DOE molding trials, it is found the adaptive GP model, the hybrid GA is used to find the
that the averaged overall error percentage between the optimal process conditions that can minimize the objec-
simulation and experiment results is 14.87%, suggesting tive in Eq. 6. To evaluate the accuracy of those optimal
that the simulation results agree fairly well with the points, simulation is executed using those optimal condi-
experiment data. Therefore, it is adequate for performing tions. The simulation and optimization results generated

FIG. 10. Box part (polypropylene) molded with the injection molding process: (a) molded box part; (b)
CAD model of box part.

692 POLYMER ENGINEERING AND SCIENCE—-2007 DOI 10.1002/pen


TABLE 4. Number of training samples, parameters, and the convergence from the GP model are compared in Table 5. Table 5 also
criteria for each of the GP iteration steps. lists the optimal process conditions obtained from experi-
Additional
mental regression for minimizing the S&W, as detailed in
Iteration parameters training Ref. 22. From the comparison, the average error percent-
Training (threshold variance, samples Convergence age between the results of the GP model and the simula-
samples pmajor, and pminor) (major/minor) criteria tion in all four directions is quite small (6.67%), which
suggests that the adaptive GP model can approximate the
Initial step 40 0.5, 5%, 0.5% 6/2 4.744
Step 1 48 0.1, 5%, 0.5% 3/2 2.261
simulation results well. Compared with the minimum
Step 2 53 0.05, 2%, 0.5% 3/2 1.237 S&W achieved from experimental regression, the optimal
Step 3 58 0.01, 1%, 0.2% 3/1 0.921 results from the GP model are much smaller. The reason
Step 4 62 0.01, 1%, 0.1% 4/0 0.797 is that using the surrogate modeling approach is more
Step 5 66 0.01, 1%, 0.1% 4/0 0.725 flexible in determining the optimal conditions than the
Step 6 70 0.01, 1%, 0.1% 3/0 0.679
Step 7 73 — — 0.674
conventional factorial DOE, as those optima need not to
be restricted to the fixed process condition setting levels
as used in the physical molding trials. Since the new

FIG. 11. S&W data (in mm) comparison between the adaptive GP model prediction and the 32 verification simula-
tion results (samples) in the x, þx, y, and þy directions. Solid line represents the best fit whereas the dotted line cor-
responds to the unity slope (where the model prediction matches with the simulation result). (a) S&W in x direction
[mm], (b) S&W in þx direction [mm], (c) S&W in y direction [mm], (d) S&W in þy direction [mm].
TABLE 5. The optimal process conditions and objective values predicted by the surrogate model, simulation, and experiment regression after seven
iteration steps.

Process conditions Optimum S&W


Objective Ave
thold tcool vinj ppack Tmelt Tmold |x| |þx| |y| |þy| value error %

GP surrogate
model þ hybrid GA 5.93 32.57 78.68 59.17 214.76 37.45 0.151 0.794 0.802 0.809 0.639 —
Experiment regression
optimum 1 6 20 — 60 — 40 0.322 0.918 0.836 0.823 0.727 —
Experiment regression
optimum 2 6 35 — 60 — 20 0.062 1.324 0.836 0.823 0.760 —
Experiment regression —
optimum 3 6 20 — 60 20 0.117 1.303 0.836 0.823 0.768 —
Moldex3D simulation result 0.176 0.859 0.828 0.825 0.672 6.67%

adaptive optimization system does not require complete 3. C.L. Tucker, Fundamentals of Computer Modeling for Poly-
CAE simulation for every objective function evaluation mer Processing, Hanser, Munich (1989).
[7, 11], the overall design cycle time is greatly reduced. 4. R.Y. Chang and W.H. Yang, Int. J. Numer. Methods Fluids,
This intelligent system can provide design space explora- 37, 125 (2001).
tion to ensure that the optimal solutions can be quickly 5. F. Illinca and J.F. Hetu, Int. J. Numer. Methods Eng, 53,
achieved that meets or exceeds all design requirements. 2003 (2002).
6. S.J. Park and T.H. Kwon, Polym. Eng. Sci., 38, 1450 (1998).
CONCLUSIONS 7. Y.M. Deng, Y.C. Lam, and G.A. Britton, Int. J. Prod. Res.,
42, 1365 (2004).
In this study, a novel optimization procedure based on
8. L.S. Turng and M. Peic, J. Proc. of the I MECH E, Part B,
the GP surrogate modeling approach and DOE for com-
J. Eng. M., 216, 1523 (2002).
puter simulation are presented and shown to be theoreti-
9. M. Peic and L.S. Turng, SPE ANTEC Tech. Papers, 624
cally sound and practically applicable to the optimization
(2002).
of the injection molding process. Based on the Bayesian
10. B.H.M. Sadeghi, J. Mater. Process Tech., 103, 411 (2000).
probability and inference approach, the GP surrogate
model can provide predictions and estimations of the con- 11. J. Zhou and L.S. Turng, Int. Polym. Proc., 21, 509 (2006).
fidence (in terms of variance) simultaneously. The new 12. C. Castro, M.C. Rios, B. Lilly, and J.M. Castro. J. Polym.
optimization procedure does not need to assume a specific Eng., 25, 176, 459 (2005).
regression model in advance; it can intelligently deter- 13. M.N. Gibbs and D.J.C. MacKay, Efficient Implementation of
mine the optimization direction while adaptively selecting Gaussian Processes, University of Cambridge, England
(1997).
the additional samples needed to improve the surrogate
model. On the basis of the results of the illustrative appli- 14. C.A.L. Bailer-Jones, A summary of Gaussian Processes,
Canvedish Laboratory, University of Cambridge, England.
cation examples, the new procedure proposed here can
effectively establish the surrogate model with minimum 15. D.J.C. Mackay, Introduction to Gaussian Processes, Univer-
sity of Cambridge, England (1998).
computational resources and adaptively search for the
optimal design and process conditions for injection mold- 16. D.C. Montgomery, Design and Analysis of Experiments, Wiley,
New York (1991).
ing within a reasonable timeframe.
17. J.P.C. Kleijnen, S.M. Sanchez, T.W. Lucas, and T.M.
Cioppa, INFORMS J. Computing, 17, 263 (2005).
ACKNOWLEDGMENTS 18. Y.G. Xu, G.R. Li, and Z.P. Wu, Appl. Artif. Intell., 15, 601
The authors thank Moldex3D and Moldflow for mak- (2001).
ing their simulation software available for this study. 19. R. Hooke and T.A. Jeeves, J ACM, 8, 212 (1961).
20. D.E. Goldberg, Genetic Algorithms in Search, Optimization,
REFERENCES and Machine Learning, Addison-Wesley, New York (1989).
21. Moldflow Corporation, Available at http://www.moldflow.com.
1. D.V. Rosato and M.G. Rosato, Injection Molding Handbook, 22. A. Kramschuster, R. Cavitt, D. Ermer, Z.B. Chen, and L.S.
Chapman & Hall, New York (1995). Turng, Polym. Eng. Sci., 45, 1408 (2005).
2. S.W. Kim and L.S. Turng, Modelling Simul. Mater. Sci. 23. CoreTech System Co., Ltd, Available at http://www.moldex3d.
Eng., 12, S151 (2004). com.

694 POLYMER ENGINEERING AND SCIENCE—-2007 DOI 10.1002/pen

You might also like