Applied Soft Computing 12 (2012) 2601–2613

Contents lists available at SciVerse ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Thermal parameters optimization of a reflow soldering profile in printed

circuit board assembly: A comparative study
Tsung-Nan Tsai ∗
Department of Logistics Management, Shu-Te University, Kaohsiung, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a comparative study for optimizing the thermal parameters of the reflow soldering
Received 16 July 2011 process using traditional and artificial intelligence (AI) approaches. High yields in the reflow soldering
Received in revised form 13 January 2012 process are essential to a profitable printed circuit board (PCB) assembly operation. A reflow thermal pro-
Accepted 18 March 2012
file is a time-temperature graph which is used to properly control the thermal mass and heat distribution
Available online 20 April 2012
to form robust solder joints between the PCB and electronics components during reflow soldering. An
inhomogeneous temperature distribution for a reflow thermal profile can cause various soldering defects,
Keywords:
which can jeopardize product reliability and lead to significant productivity loss. In the multi-objective
Surface mount technology
Reflow soldering
optimization problem, three alternative optimization methods are discussed and compared: response
Response surface methodology surface methodology (RSM), nonlinear programming (NLP), and a hybrid AI technique. A dataset was
Nonlinear programming gathered using a 38-4 experimental design for the development of meta-models through response sur-
Neural network face quadratic modeling. In the first method, RSM is used to acquire the optimal heating parameters,
Genetic algorithm while in the second method NLP is used to derive a global solution based on the meta-models. The back-
propagation neural network (BPN) is used in the third method to formulate the nonlinear relationship
between the heating inputs and responses. A genetic algorithm (GA) is then used to elicit the optimal
heating parameters from the established BPN model. The evaluation results show that all three methods
provide satisfactory soldering performance in terms of the process capability, sigma level, and process
window indices (PWIs). Particularly, the hybrid AI approach provides superior nonlinear formulation
capability and optimization performance.
© 2012 Elsevier B.V. All rights reserved.

1. Introduction between the SMDs and the PCB without altering the original char-
acteristics of the electrical components [1].
Printed circuit boards (PCBs) are found in most electronic In the reflow soldering process, a reflow oven is used to heat
products. Modern electronics products increasingly require the the assembled boards to an appropriate temperature for the speci-
features of lighter weight, smaller size, and higher quality, high fied period of time at the proper heating rate [1]. A thermal profile
pin-count, and various other product configurations. This is par- time-temperature graph is commonly generated to adjust the heat-
ticularly evident in personal computers and telecommunication ing parameters and control the effects of heating on the assembled
devices. Surface mount technology (SMT) is the primary method boards. Fig. 1 illustrates a typical reflow thermal profile from a con-
used to attach and connect surface mount devices (SMDs) perma- vection reflow oven. Tr (r = 1, 2, . . ., 6) represents the timestamp of
nently onto the PCB. A typical SMT production line consists of the a reflow thermal profile. The time needed to complete a heating
stencil printing, component placement, and solder reflow process cycle is about five minutes.
steps. Stencil printing is a preliminary process where the desired There are four independently controllable heating zones in a
amount of solder paste is first deposited onto the pads of the PCB. reflow thermal profile, the preheating (the duration is from 0 to T1),
The SMDs are then placed on the pasted pads. The reflow solder- soaking (the duration is from T1 to T2), reflow (the duration is from
ing process heats the entire assembled board to welt the adjoining T3 to T5), and cooling (the duration is from T5 to T6) zones. Each
surfaces and to melt the solder paste in order to form solder joints heating zone has its associated heat absorption rate and elapsed
time. For the instance of the preheating zone, the time elapsed
to quickly heat both the board and SMDs from room tempera-
ture to about 150 ◦ C is D1 s (T1 − 0). The duration of the soaking
∗ Correspondence address: No. 59, Hun Shan Rd., Yen Chau, Kaohsiung, 82445, zone is from T1 to T2 (T2 − T1 = D2 s). In this zone, the board tem-
Taiwan. Tel.: +886 7 6158000x4511; fax: +886 7 7801946. perature increases to around 180 ◦ C which activates the flux leading
E-mail address: [email protected] to welting between the metallized surfaces of the solder pads and

1568-4946/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.asoc.2012.03.066
2602 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613

Fig. 1. Typical reflow thermal profile.

component leads to reduce metal oxidation. The solder particles and relationship in the underlying system for prediction of the
melt and turn into liquid solder in the reflow zone. Finally, the sol- heating outputs [12,13].
der joints are formed between the component leads and solder pads
in the cooling zone.
Most prior studies have concentrated on finding the process
Soldering quality is extremely important in order to assure the
conditions to match the desired shape of the reflow thermal profile.
reliability of electronic products in the electronics assembly [1–3].
These methods make a contribution towards the design of thermal
The reflow soldering process is recognized as one of the major
recipes for producing the specified target profile. However, it is
sources of soldering defects [4]. A major setup usually occurs when
complicated and difficult to match all the targets or specifications
the coming PCB configuration is different from the current board,
because of the complexity of the product configuration. In other
since an incompatible temperature distribution in the components
words, some soldering defects may still occur and the reliability
and/or PCBs during the reflow soldering process has the potential
of the solder joints cannot always be ascertained [14]. Meanwhile,
to produce defective products [1,5]. In practice, defective products
the use of a traditional SPC system may lead to erratic recipes since
require additional repairs and reworking that can cause a loss of
the mechanism is insufficient to formulate the nonlinear nature of
productivity of roughly 30–50% of the total manufacturing costs
reflow soldering. Additionally, most studies have been limited to
[6]. Accordingly, defining, measuring, and optimizing the reflow
manipulating a few reflow factors, and a certain degree of improve-
soldering process is the primary goal towards assuring soldering
ment is considered significant.
quality and solder joint reliability, and cutting down on manufac-
In this study, we considered a wider range of process vari-
turing costs for the electronics assembly industry.
ables and responses, comparing three methods to resolve the
Over the past few years, there have been many studies carried
multi-objective optimization problem for reflow thermal profiling
out on the monitoring of the reflow soldering process and improve-
operations while deriving the desired shape of the reflow ther-
ment of soldering quality. Most of these studies have adopted
mal profile and controlling the heating factor in order to achieve
qualitative defect mechanism analysis, statistical process control
higher reliability of solder joints. The response surface method-
(SPC) systems, and computational simulation models for upgrad-
ology (RSM) is an important branch of experimental design and a
ing the soldering quality through keeping the actual profile at the
useful technique in developing new processes and optimizing their
specified target level [7]. Three commonly used methods have been
performance [15]. Therefore, the traditional RSM is considered as
implemented to achieve this objective:
an optimization method in this study. Mathematical programming
provides the basic algorithms to develop engineering applications
(1) Simulation modeling: a computational model with process including process design, nonlinear control systems, and a variety
specifications is used to simulate the shape of the reflow ther- of others [16]. Considering the nonlinear behavior of the reflow
mal profile. The simulations are performed iteratively until the soldering process, nonlinear programming (NLP) is an alternative
results agree with the predefined specifications or the shape method employed to solve the given optimization problem. The
of the target thermal profile. The final recipe(s) can then be artificial neural network (ANN) and genetic algorithm (GA) are two
utilized in the reflow oven setup [7–10]. of the most promising natural computation techniques. These two
(2) SPC method: the ingredients of a reflow thermal profile (e.g., AI methods have proven to be useful in solving many engineering
peak temperature, soaking temperature, preheating slope, and problems and in optimizing nonlinear systems [17]. The hybrid AI
so on) are monitored by traditional control charts (i.e., X̄ − R and algorithm which combines ANN with GA has proved to be effective
EWMA charts) with predefined limits to diagnose and identify at solving numerous prediction and process optimization problems
any variability in the reflow soldering process and reduce the for complex and nonlinear systems [18–20]. The hybrid AI tech-
need for post-reflow inspections [11]. nique is suitably utilized to optimize the thermal parameters of a
(3) Artificial intelligence (AI) approach: the AI models are des- reflow thermal profile in this research as an alternative.
ignated to acquire knowledge of the essential factors that The remainder of this paper is organized as follows. In Section 2
influence the performance of the reflow soldering process. the literature is reviewed and the methodology used in this study
Specifications are considered to deduce the most acceptable introduced. Section 3 outlines the proposed optimization proce-
recipes for achieving homogeneous temperature distribution dure and its implementation. Section 4 compares the performance
in the SMDs and PCBs as well as high solderability. A computa- of the three optimization methods. Finally, some conclusions are
tional algorithm is developed based on the process knowledge offered in Section 5.
 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613 2603

2. Literature review The output signal is given by Eq. (3) through a transfer function
(f).
2.1. Heating factor n

o = f( wh xh ), (3)
Recently, a new reflow parameter has been proposed for charac-
h=1
terizing the combined effects of the temperature and the duration
above the liquids in the reflow zone. This parameter, the heating where w denotes the weight vectors, w ≡ [w1 , w2 , · · ·, wn ]; and x
factor (Q ), is important for adequately controlling the reliability represents the input vectors, x ≡ [x1 , x2 , · · ·, xn ].
of the solder joints [7,14,21,22]. The heating factor is shown in the The selection of the transfer function typically depends on the
area of the reflow temperature profile (see the top of Fig. 1). Robust nature of the network output. The most typical continuous transfer
solder joints may not form if the heating factor is far from the 200 function is sigmoidal, as shown in Eq. (4)
and 700 range (s ◦ C) [7,14]. The heating factor is defined below. ′
f (w′ x) ≡ 1/(1 + e(−w x) ). (4)
The liquid temperature Tm (Fig. 1) is the melting point of the
eutectic solder paste being used. In practice, the Tm is a constant The back-propagation neural network (BPN) is the best-known
(183 ◦ C) if a 63Sn37Pb alloy solder paste is used. The heating factor supervised technique for adjusting randomized weights during the
can be calculated as the integral of the measured temperature (T(t)) data-learning phase according to the steepest gradient along the
over the Tm across the dwelling time T5–T2, which is formulated as error surface [27]. The BPN has input, hidden, and output layers. In
 T5 order to design a BPN model, user-defined parameters include: (1)
Q = (T (t) − Tm )dt, (1) the number of data-learning iterations; (2) the learning rate; (3)
T2 momentum; (4) the number of hidden layers; and (5) the number
of processing elements in the hidden layer. The general design rules
where Q denotes the heating factor; Tm is the melting point.
have been reported by Lippmann [28]. Readers are referred to Hou
et al. [29] for a more detailed discussion of the BPN. Due to the
2.2. Response surface methodology complex and nonlinear behavior of the reflow thermal profiling
process, the BPN algorithm is suitable for mapping the cause-and-
The RSM is a series of mathematical and statistical techniques effect relationships of inputs and responses, and to improve the
useful for the modeling and analysis of problems in which the convergence rate and prediction performance. BPN is utilized here
responses of interest are influenced by several influential factors, as one of the modeling tools in this study.
and the main objective is to optimize the responses [15]. The main
idea of RSM is to use a sequence of physical experiments to obtain 2.4. Genetic algorithm
optimal responses with respect to some statistical criteria [23]. The
step-by-step design procedure of RSM consists of: (1) measuring The GA is an adaptive heuristic search algorithm that can escape
and collecting the responses of interest using a series of exper- from local optima traps to find the global optimal solutions. This
imental designs; (2) developing the approximation meta-models method is inspired by the mechanisms of natural selection and
with the best fittings; (3) acquiring the optimal set of input factors genetics in human biological systems [30]. In contrast to the point-
to maximize or minimize response through a two or three dimen- to-point searching method, the GA concurrently evaluates many
sional hyper surface; (4) using the models to predict the values of points across the solution space, thus making convergence towards
improved response. the global solution more likely [30]. The GA method initially begins
The second-order model (quadratic polynomials) is flexible with a set of solution candidates called a population. A set of popu-
enough to be frequently utilized to approximate a portion of lations is generated according to a predefined number of modeling
the true response surface with parabolic curvature. Eq. (2) is parameters encoded as binary strings or real numbers. Successive
a response-surface approximation function, which is one of the generations are generated from the parent population after evalu-
methods used to optimize the problem presented in this study ating the parameter sets with higher fitness values. Offspring are
k
 k
  bred through searching the solution space for which there is a high
Y = ˇ0 + ˇi xi + ˇii xi 2 + ˇi·j xi xj + ε, (2) probability of improvement being achieved through the operations
of selection, crossover and mutation. The GA selection operator
i=1 i=1 i<j
selects individuals for the next generation. The crossover operation
where ˇ0 is a constant; ˇi is linear term coefficient; ˇii stands for can breed new offspring by exchanging the chromosomes of two
the quadratic term coefficient; ˇi·j denotes the interaction term parents. The mutation process increases variability to the selected
coefficient; and ε denotes a random error component. parents acquired from the crossover stage. Interested readers are
referred to Coley [31] for a more detailed discussion of the GA
2.3. Artificial neural network algorithm.

An ANN is a computational model that mimics the decision- 3. Multi-objective optimization and implementation
making process of the human neural system through the data
learning process. The ANN is an effective modeling tool for approx- The proposed methodologies for solving a multi-objective opti-
imate mapping of the nonlinear relationship between multi-inputs mization problem are illustrated using a case study. Fig. 2 shows a
and multi-outputs (MIMO) without making any distributional flowchart of the optimization process. A set of experimental data
assumptions regarding the underlying data-learning process [24]. was collected from company V, a well-known American electronics
The ANN has gained considerable attention for solving multi- manufacturer, located in Kaohsiung, Taiwan, which fabricates and
objective optimization problems [25] and process controls [26]. designs many types of electronics transaction and payment prod-
The basic idea of a neural network is a neuron with several input ucts. The experimental data were investigated in order to derive
connections and a single output. The signal flow of the inputs and the meta-model for each output of a reflow thermal profile. Based
outputs are considered unidirectional. An ANN model consists of a on the meta-models, the first method employs the RSM to acquire
number of processing elements (neurons), which are distributed in the optimal heating parameters, while the second method the NLP
a number of hierarchical layers. is used to find the global solution. The BPN algorithm is utilized to
2604 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613

Fig. 2. Flowchart of the reflow thermal profiling optimization process.

formulate nonlinear cause-and-effect relationships of the process the connection of the CAD Gerber file and the bills of material
factors and responses by learning from the experimental dataset (BOM)
followed by GA searching for the optimal heating parameter sets Cn
in the third method. The optimization performances of these three 
Ni Vi
solutions were confirmed and compared using the process win-
i=1
dow index (PWI), defects per million opportunities (DPMO), and Component density = , (5)
Cpk . The development flow is described in detail in the sections VBoard
below. where Cn represents the number of component types; Ni is the
quantity of component type i; Vi indicates the volume of component
type i; VBoard stands for the volume of the bare boards.
3.1. Process variables and responses The reflow thermal profile is extensively divided into six time
zones, D1, D2. . . and D6 (see Fig. 1 and Fig. 3) in this study. The
Earlier studies confirmed that there must be agreement heating factor and the time above the liquids are critical for achiev-
between the reflow methods, design of the assembly, component ing good solderability [7,14,21,22] and increasing the resolution
density, settings of the reflow oven, and the heating factor for of a thermal profile in the analysis of the optimization problem.
obtaining an acceptable reflow thermal profile [5,7,14,32]. For the Therefore, the responses of a thermal profile are represented
sake of efficiency and accuracy, domain experts evaluated the col- by eleven outputs, including preheating slope (◦ C/s), soaking
lection of variables and chose ones that would be fruitful to an
experimental design. Therefore, the factor selection of the reflow
soldering process was based on a literature review [1,7,32–34]
and the field knowledge of veteran engineers. The selected fac-
tors contain three major elements in the experiment, including the
types of SMDs, PCBs, and the reflow oven settings, as illustrated in
Fig. 3.
There are eight inputs and eleven responses of a reflow thermal
profile. The input is comprised of the settings of the reflow oven and
PCB configuration (component density). The reflow oven settings
include temperature settings for four oven heaters (Heater #1 to
Heater #4), the conveyor speed, and the air exhaust rate. The major
source of PCBs used by subject company is 1.6 mm thick FR-4 PCBs
(above 95% of the total production volume). Therefore, this type of
PCB was selected for the experiments.
Given the same oven settings, the temperature profile is a func-
tion of the component density and the board flow gap, which affects
heat rising and absorption rates [32]. The board gap is the des-
ignated distance between two consecutive boards as they move
through the oven. Furthermore, the reflow thermal profile is depen-
dent on the configuration of the designated PCB, PCB thickness,
component types, and heating requirements [3,4]. The compo-
nent density is defined in Eq. (5) and can be calculated through Fig. 3. Map of process variables and responses for reflow thermal profiling.
 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613 2605

Table 1
Factor levels for the reflow thermal profiling.

Factors Symbol Level 1 Level 2 Level 3

x1 : Component density (mm3 /cm2 ) N/A 0.1 0.17 0.33

x2 : Exhaust rate (m3 /min) Exhaust 2.5 4.0 5.5
x3 : PCB gap (cm) Pitch 1 11 21
x4 : Conveyor speed (m/min) Con speed 0.55 0.70 0.85
x5 : Heater #1 temperature (◦ C) Heater 1 140 165 190
x6 : Heater #2 temperature (◦ C) Heater 2 150 165 180
x7 : Heater #3 temperature (◦ C) Heater 3 145 165 185
x8 : Heater #4 temperature (◦ C) Heater 4 230 255 280

Table 2
Reflow thermal profile responses.

Responses Symbol Lower limit Target Upper limit

y1 : Peak temperature (◦ C) Peak 205 210 230

y2 : Preheating slope (◦ C/s) Pre slope 1.5 3.0 5.0
y3 : Soaking temperature (◦ C) Soaking 150 160 170
y4 : Ramp-up slope (◦ C/s) Ramp up 1 2 3
y5 : D1 (s) D1 90 110 125
y6 : D2 (s) D2 60 70 80
y7 : D3 (s) D3 20 30 40
y8 : D4 (s) D4 10 20 25
y9 : D5 (s) D5 20 30 40
y10 : D6 (s) D6 35 45 60
y11 : Heating factor (s ◦ C) H factor 200 N/A 700

temperature (◦ C), ramp-up slope (◦ C/s), peak temperature (◦ C), temperature being 2 ◦ C/s for the development of robust intercon-
durations of the six time zones, and heating factor (Q ) (s ◦ C). nected solder joints, with minimization of solder spattering and
balling failures.
Following Fig. 1, the preheating time (D1) = T1 − 0, and the soak-
3.2. Design of the experiment
ing time (D2) stands for the duration from T1 to T2, respectively. The
other time zones (D3–D6) can be computed in the same fashion. The
To minimize the design aberrations, Franklin [35] has provided
reflow zone takes place between T2 and T5. We purposely divide
an introduction to generate a 3K-P orthogonal array. Readers are
the monitored reflow duration into three zone times (D3–D5). The
also referred to Franklin [35] for a detailed discussion of the 3K-P
D4 duration time is more critical to achieve good solderability. D6
fractional factorial experimental design. For the proposed problem,
indicates the cooling phase that starts at T5–T6. The heating absorp-
there are eight design parameters, so the value of K is therefore
tion rate varies from component to component. The heating rate for
equal to eight. There is consideration of the trade-off between data
a larger component (such as a plastic leaded chip carrier (PLCC) or a
collection cost and sample size. If the value of P is changed from
quad flat package QFP component) is slower than those for a smaller
4 to 3 and to 5, the resulting designs will have 243 and 27 exam-
sized one (i.e., resistor chips/capacitor chips or small outline tran-
ples, respectively. Collecting 243 tuples of experimental data is too
sistor (SOT)). Four thermocouples are attached to the component
expensive, while 27 tuples are too few to formulate the nonlinear
leads for the purpose of collecting thermal data for each board.
relationship of process inputs and outputs. That is, the best compro-
The temperature contour for each experimental run was gener-
mise generates the resulting 38-4 (81 experiment runs) orthogonal
ated using the thermal tracker and DataPaq® software package. The
array.
thermocouples were articulated to measure the components with
Each experimental run has different factor values, namely factor
the highest, lowest, and two middle heating rates. The responses
levels. In the experiments, the number of levels of an indepen-
are recorded as the average of the four values. The heating factor
dent variable was equal to the number of variations of that factor.
is approximated by calculating the sum of the rectangular area as
We defined three levels associated with each factor and the pro-
shown in Fig. 4.
cess specifications were determined from discussions with domain
experts, as shown in Table 1. The three levels of each design param-
eter were represented by ‘Level 1’, ‘Level 2’ or ‘Level 3’ in the matrix.
For instance, the conveyor speed of the reflow oven differentiates
from 0.55 (m/min) (Level 1) to 0.85 (m/min) (Level 3) for the pur-
pose of investigating its main effect and interactions with other
factors.
The responses of a reflow thermal profile are summarized in
Table 2. The target values and ranges of responses are defined as
the process specifications. For example, the appropriate peak tem-
perature for the formation of strong solder joints in the reflow zone
should fall into the range from 205 ◦ C to 230 ◦ C, more appropri-
ately 210 ◦ C. The preferred soaking temperature should be between
150 ◦ C and 170 ◦ C to ensure proper melting of the solder paste
and wetting of the surfaces of the component leads and PCB pads.
The desired soaking temperature is 160 ◦ C. The ramp-up slope
should range from 1 ◦ C/s to 3 ◦ C/s with the appropriate target Fig. 4. Heating factor approximation.
2606 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613

Fig. 5. Normality testing of residual plots.

3.3. Preliminary data analysis and meta-model development quadratic modeling to generate a meta-model (fitted response-
surface model) for each output. A total of eleven meta-models were
The 81 tuples of experimental data were analyzed using derived to formulate the nonlinear relationship between process
the Design-Expert® [36] software package for response surface inputs and responses. Fig. 5 shows the normal probability plots of

Table 3
Summary of ANOVA results.

Responses Factors

Density Exhaust Pitch Con speed Heater #1 Heater #2 Heater #3 Heater #4 R2 Adj R2

** * ** * ** **
Peak 0.9961 0.991
** ** ** ** ** **
Pre slope 0.9162 0.9065
** ** ** ** **
Soaking 0.9837 0.9621
** ** ** * ** **
Ramp up 0.9255 0.9168
** ** ** ** **
D1 0.9609 0.9266
** **
D2 0.9807 0.955
** ** ** **
D3 0.9556 0.8964
* ** ** *
D4 0.8570 0.8262
* ** * * ** ** **
D5 0.8533 0.8363
* * ** ** ** **
D6 0.9597 0.906
** ** ** * **
H factor 0.9764 0.9481
*
p-value < 0.05.
**
p-value < 0.01.
 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613 2607

Fig. 6. 3D surface of peak temperature and heating factor for component density = 0.1 mm3 /cm2 .

the residuals (for conciseness, only 4 out of 11 models are par- surfaces. First of all, the component density (x1 ) (defined in Eq. (5))
tially demonstrated); the data points are approximately linearly is fixed at three numbers (i.e., 0.1, 0.17, and 0.33 mm3 /cm2 ), since
distributed, thereby confirming the acceptability of the assump- the product families built by the subject company are approxi-
tion of normality. The analysis of variance (ANOVA) results are mately distributed according to these three types of component
summarized in Table 3. density. The rest of the input variables are bounded with the values
As a rule of thumb, if the R2 value is more than 0.90 and the of factor level 1 to factor level 3 as listed in Table 1. The desired value
value of the adjusted R2 is more than 0.80, the fitted model will for each dependent variable (response) is set as the target listed
be considered adequate and each model provides sufficient expla- in Column 2 of Table 2. For example, the process input component
nation capability. The analysis results prove that the experimental density (x1 ) is fixed at 0.1 (mm3 /cm2 ), the boundary of the exhaust
data are adequate for modeling the thermal profiling operation. The rate (x2 ) is set between 2.5 and 5.5 (m3 /min), the PCB gap (x3 ) ranges
independent variables or combinations thereof affect each process between 1 and 21 (cm), etc., in order to obtain the desired outputs,
response (dependent variable). The significant variables influenc- such as the targets of peak temperature (y1 ) is 210 (◦ C), preheating
ing each response are marked with asterisk(s). For example, the slope (y2 ) of 3.0 (◦ C/s), soaking temperature (y3 ) of 160 (◦ C), etc. The
heating factor is controlled and adjusted based on the exhaust rate, RSM method can then be performed to extract the optimal param-
board gap, conveyor speed, and temperature settings of heater #3 eter settings and predictions through the developed meta-models
and heater #4, as illustrated in the last row of Table 3. Notably, the (discussed in Section 3.3) for each type of component density.
conveyor speed is an important influential variable which affects The 3D surface graphs of the first optimal parameter set
all the responses of the reflow soldering process. (0.1 mm3 /cm2 of the component density) are partially demon-
For the sake of conciseness and due to the complexity of the strated for the sake of conciseness. Only the responses of the peak
meta-models, only the two models for peak temperature and the temperature (y1 ) and heating factor (y11 ) are shown in Fig. 6(a)
heating factor are depicted as in Eqs. (6) and (7). The symbols (x1 , x2 , and (b), respectively. The surface plots are employed to predict the
. . ., x8 ) and (y1 , y2 , . . ., y11 ) are defined in Tables 1 and 2, respectively values of the dependent variables for the preparation and opti-
denoting the inputs and outputs of the reflow thermal profile listed mization of the reflow thermal parameters. Both graphs show a
in the equations. curvilinear contour in conformity to the meta-model (quadratic

y1 = −8.99 + 369.81x1 + 4.7x2 + 0.35x3 − 8.12x4 − 0.15x5 − 0.18x6 + 0.15x7 + 0.74x8 + 2.86x1 x2
−0.23x1 x3 − 6.9x1 x4 + 0.03x1 x5 + 0.09x1 x6 + 0.09x1 x7 − 0.23x1 x8 − 7.11(E − 003)x2 x3
+2.67x2 x4 − 0.01x2 x5 − 8.48(E − 003)x2 x6 + 1.03(E − 003)x2 x7 + 0.02x2 x8 + 0.32x3 x4
−3.12(E − 004)x3 x5 − 1.47(E − 003)x3 x6 − 2.5(E − 004)x3 x7 + 1.01(E − 003)x3 x8 + 0.04x4 x5 (6)
+0.2x4 x6 − 0.04x4 x7 − 0.44x4 x8 + 1.32(E − 003)x5 x6 − 4.83(E − 004)x5 x7 + 9.57(E − 005)x5 x8
−0.95(E − 004)x6 x7 + 1.12(E − 003)x6 x8 + 3.29(E − 004)x7 x8 − 803.54x12 − 0.69x22 − 0.01x32
+14.40x42 + 1.19(E − 004)x52 − 4.73(E − 004)x62 + 7.75(E − 004)x72 − 2.59(E − 004)x82

y11 = −11709.11 + 10161.31x1 + 658.13x2 + 48.23x3 + 3882.4x4 + 2.05x5 + 1.02x6 + 18.86x7

+30.33x8 + 100.97x1 x2 − 14.65x1 x3 − 764.04x1 x4 − 2.94x1 x5 + 17.44x1 x6 + 11.85x1 x7
−1.5x1 x8 − 0.44x2 x3 − 217.63x2 x4 − 0.29x2 x5 − 0.43x2 x6 − 2.0x2 x7 + 1.27x2 x8 − 15.25x3 x4
(7)
+0.06x3 x5 − 0.08x3 x6 − 0.08x3 x7 + 0.04x3 x8 + 0.15x4 x5 + 5.74x4 x6 + 2.1x4 x7 − 55.78x4 x8
+0.01x5 x6 − 0.001x5 x7 + 0.01x5 x8 + 0.04x6 x7 + 0.02x6 x8 − 0.003x7 x8 − 31630.99x12
−28.31x22 − 0.71x32 + 4832.1x42 − 0.03x52 − 0.08x62 − 0.04x72 + 0.05x82

3.4. Optimization using RSM model). The two demonstrated response surfaces of y1 and y11
are shown in Fig. 6. These two response contours are similar
In the optimization process to find the reflow thermal parame- which can be explained by the two responses being affected by
ters, RSM quantifies the relationship between the heating variables the component density and the temperature setting of heater
and the response surfaces, and further, finds the optimal solution #4. In this instance, the optimal parameter set (x2 = 3.9, x3 = 5,
sets with the highest desirability according to all the response x4 = 0.74, x5 = 142, x6 = 170, x7 = 182, and x8 = 245) is derived to
2608 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613

acquire the responses for peak temperature (y1 = 208 ◦ C), pre-

10%
20%
20%
0%
14%
10%
0%
13%
10%
12%
30%
Hybrid AI

PWI
heating slope (y2 = 2.9 ◦ C/s), soaking temperature (y3 = 160 ◦ C), and

0.72
4.4
heating factor (y11 = 498 s ◦ C), as illustrated in Table 4. Partial com-

168
160
183
248
parison to the targets of a reflow thermal profile show that the

3.1
Pred.
peak temperature (y1 ) and preheating slope (y2 ) are slightly lower

30%
211

158
2
110
69
30
21
31
46
525
than the target values, while the soaking temperature (y3 ) meets
the target, and the heating factor (y11 ) lies within the process

50%
49%
70%
20%
9%
40%
20%
27%
50%
36%
55%
specifications.

PWI
0.71
3.6
The optimal parameter sets for the other two types of compo-

NLP

150
155
185
245
nent density were drawn in the same fashion and are summarized
in Table 4. The predictions provided by the RSM method all confirm

3.6

2.2
Pred.
0.33 (mm3 /cm2 )

70%
205

153

109
66
28
22
25
43
588
the pre-defined process specifications listed in Table 2.

3.5. Optimization using NLP

0.74

30%
31%
40%
10%
20%
20%
20%
0%
10%
36%
41%
3.5

PWI
RSM

140
151
185
235
NLP is a mathematical programming technique where the objec-
tive function is nonlinear, or at least one or more of the constraints

3.3

1.9
Pred.

41%
207

156

104
68
28
20
29
52
552
have a nonlinear relationship or both. NLP is employed routinely
and is particularly efficient in solving optimal control problems. An
NLP model can be generally formulated as follows:

Hybrid AI

0%
26%
10%
10%
26%
0%
20%
0%
30%
20%
28%
PWI
0.73
Min

4.2
f (x)

11

172
160
178
247
x
s.t.

3.2

1.9
Pred.
gj (x) ≤ 0

30%
210

161

112
70
28
20
27
50
520
hj (x) = 0
x∈X

50%
31%
30%
20%
31%
20%
20%
27%
20%
20%
49%
0.76

PWI
3.5
where X is a subset of Rn ; x is a vector of the n component x1 , x2 , NLP

179
155
185
245
. . ., xn , and f : X → R; the function f is the objective function; gj (x)
is an inequality constraint; and hj (x)is an equality constraint.

3.3

1.8
0.17 (mm3 /cm2 )

Pred.
The optimization of the reflow thermal parameters can be

50%
205

157

113
72
28
18
32
50
572
formed as a constrained regression problem using NLP [37]. A
reasonable performance measure of the model’s accuracy is the
0.76
4.2

30%
26%
10%
10%
26%
50%
20%
0%
30%
20%
29%
difference between the actual numbers of responses (yi ) and the
RSM

PWI
2

142
155
180
233
predicted numbers (ŷi ), namely, the constrained minimum sum of
absolute errors (CMAE).
The objective function is to minimize the CMAE, expressed as

3.2

1.9
Pred.

30%
207

161

112
65
28
20
27
50
523
11
  
Minimize z = yi − ŷi  (8)
Hybrid AI

10%
0%
10%
10%
14%
10%
0%
13%
0%
4%
13%
PWI
0.71

i=1
4
10

160
173
175
240

s.t.
2.1
Pred.

14%
209
3
161

110
69
30
19
30
47
483
g1 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of
peak temperature (Eq. (6))
Summary of the optimization results for different component densities.

g2 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of
0.77

50%
43%
0%
10%
77%
80%
20%
13%
50%
12%
44%
3.5

preheating slope
PWI
NLP

155
180
180
245

g3 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of
soaking temperature
3.5

2.1

g4 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of
0.1 (mm3 /cm2 )

Pred.

50%
205

160

121
62
28
19
25
49
560

ramp-up slope
g5 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of D1
0.74
3.9

g6 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of D2
RSM

20%
9%
0%
0%
31%
10%
10%
0%
20%
12%
19%
5

142
170
182
245

PWI

g7 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of D3
g8 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of D4
g9 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of D5
2.9
Pred.

20%
208

160
2
102
71
31
20
28
49
498

g10 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of D6
Response prediction/PWI (%)

g11 (x1 ,x2 ,x3 ,x4 ,x5 ,x6 ,x7 ,x8 ) = the meta-model for the response of
heating factor (Eq. (7))
Component density (x1 )

Heater #1 Temp. (x5 )

Heater #2 Temp. (x6 )
Heater #3 Temp. (x7 )
Heater #4 Temp. (x8 )

Preheating slope (y2 )

Conveyor speed (x4 )
Exhausting rate (x2 )

Ramp-up slope (y4 )

Heating factor (y11 )

Soaking Temp. (y3 )

g12 (x2 ) = 2.5 ≤ x2 ≤ 5.5 (9)

Peak Temp. (y1 )
PCB gap (x3 )

g13 (x3 ) = 1 ≤ x3 ≤ 21 (10)

MAX (PWI)
Methods

D6 (y10 )
D1 (y5 )
D2 (y6 )
D3 (y7 )
D4 (y8 )
D5 (y9 )

g14 (x4 ) = 0.55 ≤ x4 ≤ 0.85 (11)

Table 4

g15 (x5 ) = 140 ≤ x5 ≤ 190 (12)

 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613 2609

Fig. 7. Diagram of an ANN model for reflow soldering process.

g16 (x6 ) = 150 ≤ x6 ≤ 180 (13) 3.6. Optimization using hybrid artificial intelligence

A synthetic method combining ANN and GA is utilized to model

g17 (x7 ) = 145 ≤ x7 ≤ 185 (14)
and optimize the reflow thermal profiling operations. The proce-
dure is depicted on the left-hand side of Fig. 2. The BPN neural
g18 (x8 ) = 230 ≤ x8 ≤ 280 (15) network is used in order to nonlinearly map the relationship
between process factors and responses, while the trained neural
network serves as the fitness function evaluated by a GA for deriv-
x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 ≥ 0. (16)
ing the optimal parameter settings.
Prior to developing the ANN model, the experimental data
Three mathematical models for the three types of compo- was normalized (i.e., the process inputs and outputs were scaled
nent density (i.e., x1 = 0.1, 0.17, 0.33 mm3 /cm2 ) were formulated into the range of 0–1) in the data-preprocessing stage. The neural
according to the objective function and constraints in terms of the network modeling requires several parameters (discussed in
customized programming codes. The LINGO® (V.11) software pack- Section 2.3) to tune the network topology. However, there is
age was employed to solve the models and elicit global optimal no universally accepted criterion to determine an appropriate
heating parameters with respect to the three types of component network topology. An effective ANN design is not straightforward
density. and depends on the appropriate selections of network topology
The optimal solution and predictions obtained using the NLP and parameters, and there is the tradeoff between mapping
approach are given in Table 4. For instance, for a component performance and computational complexity. The NeuralPower®
density of 0.1 (mm3 /cm2 ), the optimal parameter set (x2 = 3.5, [38] software package was utilized and followed by a two-stage
x3 = 5, x4 = 0.77, x5 = 155, x6 = 180, x7 = 180, and x8 = 245) achieves design procedure to develop the ANN model used in this study. The
the responses of peak temperature (y1 = 205 ◦ C), preheating slope default settings for the learning rate (0.5) and momentum (0.4),
(y2 = 3.5 ◦ C/s), soaking temperature (y3 = 160 ◦ C), . . ., and heating one hidden layer and an output layer with the Sigmoid transfer
factor (y11 = 560 s ◦ C). A partial comparison to the targets of a reflow function were designated to train and test the experimental data
thermal profile, show that the predicted peak temperature (205 ◦ C) with a limit of 10,000 epochs. The 81 exemplars were randomly
is lower than the desired temperature (210 ◦ C), while the preheat- separated into two groups of 65 (about 80% of the experimental
ing slope is higher than expected, and the soaking temperature hits data) and 16 (about 20% of experimental data) as the training and
the target. However, the predictions either match the targets or lie testing samples, respectively. The modeling errors were measured
within the specifications. via the root-mean-square-error (RMSE). A final ANN topology

Table 5
Verification of the ANN topology with RMSE.

Response Peak temp. Preheating slope Soaking Temp. D3 duration Heating factor

y1 y2 y3 y7 y11

Training/testing Training Testing Training Testing Training Testing Training Testing Training Testing

65/16 (#1) 2.939 3.233 0.226 0.329 3.121 3.723 1.33 2.13 40.16 45.85
65/16 (#2) 2.898 3.811 0.254 0.384 3.111 3.778 1.86 2.68 45.11 48.99
65/16 (#3) 2.893 3.778 0.206 0.327 2.873 3.406 1.44 2.24 50.21 55.21
65/16 (#4) 2.567 2.784 0.224 0.299 2.954 3.798 1.56 2.89 52.37 57.56
65/16 (#5) 2.712 3.467 0.213 0.265 3.587 3.997 1.27 2.05 43.87 47.12
2610 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613

Fig. 8. Average RMSE and R for ANN modeling.

“8-17-11” was derived through trial-and-error experiments by In the GA searching process, the established BPN model (8-17-
minimizing the training errors. The resulting ANN topology has 8 11) is the objective function for GA evolutionary operations. The
neurons for the input layer, 17 neurons for the hidden layer, and desired response values are set as the targets (listed in Table 2).
11 neurons for the output layer, as depicted in Fig. 7. That is, the outputs of the GA should be an optimal combination of
In the second stage, the learning rate and momentum were process parameters (x2 , x3 , x4 , x5 , x6 , x7 , x8 ) needed to achieve the
systematically tested and optimized based on the predetermined desired outputs (y1 , y2 , y3 , y4 , y5 , y6 , y7 , y8 , y9 , y10 , y11 ) for each com-
network topology. The resulting parameters were the learning rate ponent density (x1 ) (i.e., 0.1, 0.17, 0.33 mm3 /cm2 ). A chromosome
(set at 0.6) and momentum (set at 0.5) from which a lower RMSE is represented by a vector of floating point numbers with values
and a higher convergence rate were produced. Additionally, to within the upper bound and lower bound of each parameter. The
ensure that there was no bias in the model, the modeling process tournament selection scheme is adopted to find the one with the
was tested through a non-repeating 20% of the testing samples. best fitness for crossover purposes according to a preliminary test.
The modeling errors were evaluated using the RMSE values. The The uniform crossover and real-value coding are applied in the GA
verification results for the proposed ANN model are partially evolutionary process. The GA optimization process can be divided
summarized in Table 5 (only partial responses are listed). They into four key steps:
indicate the consistency in selecting the testing samples and a
high mapping accuracy. 1. The size of the initial population is 80, which ensures the effi-
The average RMSE and correlation coefficient (R) for the ANN ciency of the searching process and premature convergence.
modeling error and correlation coefficient are depicted in Fig. 8. 2. Each chromosome is evaluated by applying the trained BPN as
The correlation coefficient is defined as follows: the fitness function to the decoded sequences of the factors. The
results are obtained for the entire population and compared to
n
 obtain the ranked fitness values.
(Qi − Q̄ )(qi − q̄) 3. The GA evolutionary operators are used to rebreed the consti-
i=1 tutions of the offspring for the next generation based on the
R=  , (17)
probability distribution to prevent it from being trapped in local
 n

 (Qi − Q̄ )2 (qi − q̄)2 optima. A preliminary DOE was conducted to tune the GA evolu-
tionary parameters. The GA parameters consist of the crossover
i=1
rate (Pc ) and the mutation rate (Pm ). Three replications of a 23
n
 n
 experimental design were performed and each parameter is set
where Q̄ = 1
Qi , q̄ = 1
qi ; Qi denotes the ith training sample at three levels (low, middle, and high). The level settings of
n n
i=1 i=1
the two parameters are: crossover probability (0.3, 0.5, 0.7) and
data; and qi represents the predictions. mutation probability (0, 0.2, 0.4). The optimal setting (Pc = 0.7,
A lower RMSE indicates that the estimates of the reflow out- Pm = 0.2) achieves the least sum of absolute error (the differ-
puts derived from the proposed ANN model are close to actual ence between the measured value of a quantity and its actual
observed values. The correlation coefficient approaches 1 which value). Therefore, Pc is set to 0.7, Pm is set to 0.2, and tourna-
indicates a high correlation between the process inputs and out- ment selection is utilized to carry new parameter sets into the
puts. Accordingly, both the average RMSE and the average R exhibit GA evolutionary process.
high modeling accuracy. The final results from the proposed model 4. Repeat steps 2 and 3 above. The algorithm runs only once and ter-
are illustrated in Table 6. minates when there is no improvement on the best chromosome
for 10,000 iterations.

Table 6 The optimum parameter settings obtained after the GA search-

Learning results for process responses. ing process are illustrated in Table 4. For instance if the component
Responses RMSE R
density is 0.1 (mm3 /cm2 ), the optimal parameter set (x2 = 4.0,
x3 = 10, x4 = 0.71, x5 = 160, x6 = 173, x7 = 176, and x8 = 240) is dis-
Peak temp. – y1 2.520 0.983
covered to attain the responses of peak temperature (y1 = 209 ◦ C),
Preheating slope – y2 0.171 0.976
Soaking temp. – y3 2.317 0.973 preheating slope (y2 = 3.0 ◦ C/s), soaking temperature (y3 = 161 ◦ C)
Ramp-up slope – y4 0.145 0.963 and heating factor (y11 = 483 s ◦ C), as illustrated in Table 4. A par-
D1 – y5 5.126 0.960 tial comparison to the targets of a reflow thermal profile show that
D2 – y6 5.323 0.961
the peak temperature (y1 ) is the closest to the target (210 ◦ C), the
D3 – y7 1.04 0.956
D4 – y8 0.467 0.989
preheating slope (y2 ) meets the target, the soaking temperature
D5 – y9 1.872 0.971 (y3 ) is slightly higher than the target at 1 ◦ C, and the heating factor
D6 – y10 3.151 0.948 (y11 ) lies within the process specifications. Also, all predictions stay
Heating factor – y11 39.65 0.991 within the process specifications with small differences.
 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613 2611

Fig. 9. Simulated profiles for the three methods.

4. Performance evaluations and confirmation experiments (1) The optimization performance provided by the hybrid AI tech-
nique is superior to that of the other methods. The predictions
4.1. Comparison of optimization performance through PWI are mostly close to the targets with the lowest PWIs.
(2) The Hybrid AI model provided less complexity in the deploy-
The optimized thermal profiles derived from the three meth- ment of optimization process and miner computational burden
ods were simulated using the Datapaq® software, as depicted in than others. The NLP approach requires much effort in modeling
Fig. 9. The simulated thermal contour acquired from the hybrid and deriving the optimum solutions.
AI model (blue) is closest to the target profile (green). The time-
(3) All three methods predicted a slightly lower peak temperature
temperature profile derived from the NLP model produces a much than the target. However, the NLP approach created a steeper
larger area (heating factor) than the other methods. Obviously, the preheating slope and ramp-up slope that could result in solder
hybrid AI technique provides more satisfactory performance in the balling defects.
case study. (4) The areas (heating factor) produced by the three methods com-
The process window index (PWI) is a widely accepted statistic
ply with the specified values and can assure the reliability of
used to qualify the robustness of a thermal process and the good- solder joints.
ness of the thermal profile fits for the predefined process limits in (5) The ANN approximation can reduce the numerical noise inher-
the electronics assembly industry [39,40]. The center of the process ent in solving the optimization problem for the reflow soldering
window is defined as zero, while the extreme edge of the process operations that is evidenced from the evaluation results.
window has a PWI of 99% [40]. A PWI greater than or equal to 100%
indicates quality characteristics that do not satisfy the specifica-
4.2. Confirmation experiments and performance
tions. A PWI of 30% indicates that a profile is using 30% of the process
evaluation/comparison
specification. The lower the PWI value, the more robust the process
achieved [39]. The optimization performance of the three methods is evaluated
The PWI is defined as in Eq. (18) [40]. For example, if the process through confirmation experiments. The stencil printing quality and
mean (target) for the peak temperature is set at 210 ◦ C, with the placement accuracy were maintained at a sufficient level prior to
process window calibrated from 200 ◦ C to 220 ◦ C, then a measured the confirmation experiments. In particular, the desired amount of
value of 205 ◦ C can be calculated to have a PWI of 50%, solder paste and the component offset rate were examined to pre-
  serve the “in-control” status. 270 boards with different component
i,j−1  measured value|[i,j|] − average limits|[i,j|]  densities (30 for each component density for each optimization
PWI = 100×MAX   ,
N,M  (range[i,j] /2)  method) were assembled on the same production line. Based on
(18) the principle of Six Sigma management, the process capability
index (PCI) is a statistical measure of process capability used to
where N stands for the number of thermo couples; i = 1,. . .,N; M
represents the number of statistics per thermocouple; j = 1,. . .,M;
Table 7
measured value [i, j] is the [i,j]th statistic; average limits[i,j] is the Relationships of Sigma level, yield, and DPMO.
average of the [i,j] high and low limits; range[i,j] is the [i,j] high limit
Cpk Sigma level Yield rate Process fallout
minus the low limit.
(DPMO/PPM)
To precisely evaluate the thermal performance, obtain maxi-
0.33 1 30.90% 691,462
mum efficiency, avoid thermal shock, and overheating or rapid
0.67 2 69.10% 308,538
cooling, the process mean (average limits) was replaced with the 1.00 3 93.30% 66,807
targets listed in Table 2 in this study. The PWIs were calculated for 1.33 4 99.40% 6210
the eleven outputs are illustrated in Table 4. Notably, eight signifi- 1.43 4.3 99.73% 2555
cant outputs were evaluated using PWI and the limits suggested in 1.67 5 99.98% 233
2.00 6 99.99% 2
the literatures [1,2,9,12,33].
2612 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613

Table 8
Component densities and defect opportunities.

PCBA No. Component Number of IC Number of Defect

density packages passive devices opportunities/unit

03916-02 0.10 12 97 2536

04009-02 0.17 36 156 3910
25074-02 0.33 65 224 5166

Table 9
Summary of confirmation experimental results.

Methods RSM NLP Hybrid AI

PCB density 0.1 0.17 0.33 0.1 0.17 0.33 0.1 0.17 0.33

Opportunities/30 units 76,080 117,300 154,980 76,080 117,300 154,980 76,080 117,300 154,980
Open 18 37 55 34 57 64 10 35 48
Short 26 48 77 38 76 85 18 45 65
Tombstone 7 25 75 18 47 54 9 25 58
Solder void 11 46 66 26 45 98 13 37 59
Poor wetting 8 22 50 19 23 74 7 18 54
Solder charring 6 28 42 25 21 42 3 25 33
Sub-total defects 76 206 365 160 269 417 60 185 317
Yield rate 99.90% 99.82% 99.76% 99.79% 99.77% 99.73% 99.92% 99.84% 99.80%
DPMO 999 1756 2355 2103 2293 2691 789 1577 2045
Sigma level (Short-term) 4.59 4.42 4.33 4.36 4.33 4.28 4.66 4.45 4.37
Cpk 1.53 1.47 1.44 1.45 1.44 1.43 1.55 1.48 1.46
PWI 20% 30% 41% 50% 50% 70% 14% 30% 30%

evaluate whether the process output is within specification limits flat packs (PQFPs) with 208 leads, small outlined integrated cir-
[41,42]. The Cpk index is used to measure process fallout (mea- cuits (SOIC) (i.e., SOIC 08, SOIC 16, SOIC 28, SOIC 32), thin small
sured by DPMO or parts-per-million (PPM)) which quantifies how outlined packages (TSOPs) with 32 leads, and so on. The number of
many defects a process produces. Process yield is the complement opportunities can be gathered and calculated through the relational
of process fallout and is approximately equal to the area under connections of the BOM and the CAD system. The defect opportu-
the probability density function. The sigma level (adjusted by 1.5 nities for the three component densities are illustrated in Table 8.
sigma) has a relationship with the yield and process fallout, as The solder opening and short defects in the assembled boards
shown in Table 7. For example, the yield is 99.40% and process after reflow soldering were examined by an automated in-circuit
fallout is 6210 DPMO (or PPM) if Cpk reaches 1.33. tester (ATE). Soldering defects were also identified by human
The subject company uses the DPMO metric to trace assembly inspectors with the aid of an automated visual inspection (AVI)
yield. The performance of the three methods is compared using the or an automated optics inspection (AOI) system. The results of
three metrics, the DPMO, the short-term process sigma, and the PCI the optimization performance for the three methods are summa-
(Cpk ) in the confirmation experiments. The Cpk is transformed by rized in Table 9 and depicted in Fig. 10. The process capability
dividing process sigma by 3. The DPMO value is defined as (Cpk ) derived from all three methods exceeds 1.43. That is, the pro-
cess capability derived from the three methods is satisfactory and
106 × number of defects
DPMO = . (19) is capable of producing soldering defect less than 2555 DPMO so
number of opportunities the yield rate reaches to 99.73%. However, the hybrid AI approach
Each type of component density is designated with different obviously provides the most satisfactory soldering quality, a higher
numbers of passive devices (i.e., Resistor and capacitor 0201, 0402, sigma level and better process capability.
0603, 0805, and 1206), transistors, and integrated circuit (IC) pack-
ages. The IC packages consist of thin quad flat packs (TQFPs) with
5. Conclusions
different numbers of leads (i.e., 100, 128, 144 leads), plastic quad

Reflow soldering is the dominant method for the fabrication of

reliable electronic products in the electronics assembly industry.
The improvement and optimization of the reflow soldering pro-
cess is critical in order to achieve high product quality and reliable
solder joints, which in turn increases the yield of the downstream
stages of the PCB assembly. The reflow soldering process involves
multiple performance characteristics and is nonlinear in nature.
The thermal reflow profile is the usual method used to adequately
adjust the process parameters and control the thermal effects for
board assembly. In this study, three alternative methods, including
traditional RSM and NLP, and a hybrid AI technique, were imple-
mented in order to model and solve the thermal factor optimization
problems for the thermal profile operations.
The performance evaluation results show that the hybrid AI
approach provides better performance than the traditional RSM
or NLP methods and is better at solving the given optimization
problem. However, the process capability (Cpk ) derived from all
Fig. 10. Comparison of the performance of the three optimization methods. three methods exceeds 1.43. That is, the process capability derived
 T.-N. Tsai / Applied Soft Computing 12 (2012) 2601–2613 2613

from the three methods is satisfactory and is capable of producing [14] B. Tao, Y. Wu, H. Ding, Y.L. Xiong, A quantitative method of reliability estimation
soldering defects of less than 2555 DPMO. Moreover, to confirm for surface mount solder joints based on heating factor Q␩ , Microelectronics
Reliability 46 (2006) 64–72.
the optimization performance in practice, 270 boards were assem- [15] D.C. Montgomery, Design and Analysis of Experiments, 2000, John Wiley &
bled, tested and inspected. This was done to verify the feasibility Sons, 2000.
of the three solutions sets. The results show that (1) the hybrid [16] M.S. Bazaraa, H.D. Sherali, Nonlinear Programming: Theory and Algorithms,
John Wiley & Sons Inc, 2006.
AI model provided a more feasible solution to improve the perfor- [17] J. Sanchis, M.A. Martínez, X. Blasco, Integrated multiobjective optimization and
mance of the reflow soldering process. (2) The solution sets derived a priori preferences using genetic algorithms, Information Sciences 178 (2008)
from all three methods can in practice enhance short-term pro- 931–951.
[18] M. Abouhamze, M. Shakeri, Multi-objective stacking sequence optimization of
cess capability (Cpk ) to exceed 1.43 and extend the tolerance of the
laminated cylindrical panels using a genetic algorithm and neural networks,
sigma level. (3) The AI model exhibits higher capability in terms Composite Structures 81 (2007) 253–263.
of nonlinear mapping of the nonlinear MIMO problem for reflow [19] H.-J. Kim, K.-S. Shin, A hybrid approach based on neural networks and genetic
algorithms for detecting temporal patterns in stock markets, Applied Soft Com-
thermal profiling operations. (4) The research results can help the
puting 7 (2007) 569–576.
subject company to achieve both the objectives of improving sol- [20] H.-H. Chang, Y.-K. Chen, Neuro-genetic approach to optimize parameter design
dering quality and process capability as specified in the Six Sigma of dynamic multiresponse experiments, Applied Soft Computing 11 (2011)
management paradigm and continuous improvement program. 436–442.
[21] P.L. Tu, Y.C. Chan, K.C. Hung, J.L.K. Lai, Study of micro-BGA solder joint reliability,
Microelectronics Reliability 41 (2001) 287–293.
Acknowledgements [22] F. Wu, Effects of heating factors on the geometry size of unrestricted lead-
free joints, in: IEEE International Conference on Asian green electronics, Berlin,
Germany, 2004.
The author would like to thank company V for providing related
[23] R.H. Myers, D.C. Montgomery, Response Surface Methodology – Process
support and the National Science Council of the Republic of China, and Product Optimization Using Designed Experiments, John Wiley & Sons,
Taiwan for partially supporting this research under Contract no.: 1995.
NSC 98-2221-E-366-005. [24] R.P. Cherian, L.N. Smith, P.S. Midha, A neural network approach for selection of
power metallurgy materials and process parameters, Artificial Intelligence in
Engineering 14 (2000) 29–44.
References [25] H.-Z. Huang, Z. Tian, M. Zuo, Intelligent interactive multiobjective optimiza-
tion method and its application to reliability optimization, IIE Transactions 37
[1] N.C. Lee, Optimizing the reflow profile via defect mechanism analysis, Soldering (2005) 983–993.
& Surface Mount Technology 11 (1999) 13–20. [26] M. Pacella, O. Semeraro, Using recurrent neural networks to detect changes
[2] C.-N. Chuvej, K. Jumroon, Optimization of reflow-thermal profile by design of in autocorrelated process for quality monitoring, Computers and Industrial
experiments with response surface methodology for minimizing solder-ball Engineering 52 (2007) 502–520.
defects, in: 4th IEEE International Conference on Management of Innovation [27] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning internal representations
and Technology, ICMIT, Bangkok, Thailand, 2008. by error propagation, in: Parallel Distribution Processing: Explorations in
[3] L. Zhang, C. Zhang, T. Shi, Forecasting and parameters optimization of reflow Microstructure of Cognition, MIT Press, Cambridge, MA, 1986.
soldering profile based on BPNN and GA, Advanced Materials Research 139–141 [28] R.P. Lippmann, An introduction to computing with neural nets, IEEE, Acoustics,
(2010) 990–995. Speech, and Signal Processing Magazine 4 (1987) 4–22.
[4] R.P. Prasad, One solder profile for all board? Surface Mount Technology Maga- [29] T.-H. Hou, L. Lin, P.D Scott, A neural network-based automated inspection sys-
zine (2000 May) 11–14. tem with an application to surface mount devices, International Journal of
[5] B. Illés, Distribution of the heat transfer coefficient in convection reflow oven, Production Research 31 (1994) 1171–1187.
Applied Thermal Engineering 30 (2010) 1523–1530. [30] E.D. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learn-
[6] L. Feldmann, J. Sturm, Closed loop quality control in printed circuit assembly, ing, Addison-Wesley, Reading, MA, 1989.
IEEE Transactions on Components, Hybrids and Manufacturing Technology 17 [31] D.A. Coley, An introduction to genetic algorithms for scientists and engineers,
(1994) 270–276. World Scientific Singapore (1999).
[7] J. Gao, Y. Wu, H. Ding, Optimization of a reflow soldering process based on the [32] Y.Y. Su, K. Srihari, C.R. Emerson, A profile identification system for surface
heating factor, Soldering & Surface Mount Technology 19 (2007) 28–33. mount printed circuit board assembly, Computers and Industrial Engineering
[8] F. Sarvar, P.P. Conway, A modeling tool for the thermal optimization of the 33 (1997) 377–380.
reflow soldering of printed circuit assemblies, Finite Elements in Analysis and [33] D. Scheiner, Reflow profile optimization, Surface Mount Technology (2003)
Design 30 (1998) 47–63. 60–62.
[9] D.C. Whalley, A simplified model of the reflow soldering process, in: 8th Interso- [34] D. Barbini, G. Diepstraten, U. Marquez, Process considerations for optimizing a
ciety Conference on Thermal and Thermomechanical Phenomena in Electronic reflow profile, Surface Mount Technology (2005) 28–31.
Systems, San Diego, California, 2002. [35] M.F. Franklin, Constructing tables of minimum aberration pn-m designs, Tech-
[10] Y. Gong, Q. Li, D.G. Yang, The optimization of reflow soldering temperature nomertrics 26 (1984) 225–232.
profile based on simulation, in: IEEE 2006 7th International Conference on [36] Design-Expert Software Version 8.0 User’s Guide, 2009.
Electronics Packaging Technology, Shanghai, 2006. [37] S.P. Bradley, A.C. Hax, T.L. Magnanti, Applied Mathematical Programming,
[11] C.-Q. Li, Z.H. Wu, SPC system analysis and design of reflow soldering process, in: Addison-Wesley, 1997.
IEEE. Electronic Packaging Technology, ICEPT ‘06, 7th International Conference [38] NeuralPower user’s guide, CPC-X Software, 2005.
on Electronics Packaging Technology, Shanghai, 2006. [39] P.N. Houston, B.J. Louis, D.F. Baldwin, P. Kasmierowicz, Taking the pain out of
[12] T.-N. Tsai, T. Yang, A neuro-computing approach to the thermal profile con- Pb-free reflow, Lead-Free Magazine (2008) 3.
trol of the second-side reflow process in surface mount assembly, Journal of [40] KIC Thermal, A method for quantifying thermal profile performance, Retrieved
Manufacturing Technology Management 16 (2005) 343–359. 2008-12-10, (2008).
[13] J. Zhong, X. Hu, M. Gu, J. Zhang, Comparison of performance between differ- [41] U.D. Kumara, D. Nowickib, D. Ramírez-Márquez, J.E. Verma, On the optimal
ent selection strategies on simple genetic algorithms, in: Proceedings of the selection of process alternatives in a Six Sigma implementation, International
2005 International Conference on Computational Intelligence for Modeling, Journal of Production Economics 111 (2008) 456–467.
Control and Automation, and International conference Intelligent Agents, Web [42] NIST/SEMATECH e-Handbook of Statistical Methods, available at
Technologies and Internet Commerce (CIMCA-IAWTIC’05), 2005. http://www.itl.nist.gov/div898/handbook/, 2006.