Powder Technology 173 (2007) 153 162 www.elsevier.

com/locate/powtec

Parameters optimization of a nano-particle wet milling process using the Taguchi method, response surface method and genetic algorithm
Tung-Hsu Hou a,, Chi-Hung Su a , Wang-Lin Liu b,c
a

Department of Industrial Engineering and Management, National Yunlin University of Science and Technology, Taiwan, ROC b Department of Electrical Engineering, National Changhua University of Education, Changhua, Taiwan, ROC c Precision Machinery Research and Development Center, Taichung, Taiwan, ROC Received 3 June 2006; received in revised form 15 November 2006; accepted 23 November 2006 Available online 5 December 2006

Abstract Nano-particles have been successfully and widely applied in many industrial applications. The wet-type mechanical milling process is a popular method used to produce nano-particles. Therefore, it is very important to improve milling process capability and quality by setting the optimal milling parameters. In this research, the parameter design of the Taguchi method, response surface method (RSM) and genetic algorithm (GA) are integrated and applied to set the optimal parameters for a nano-particle milling process. The orthogonal array experiment is conducted to economically obtain the response measurements. Analysis of variance (ANOVA) and main effect plot are used to determine the significant parameters and set the optimal level for each parameter. The RSM is then used to build the relationship between the input parameters and output responses, and used as the fitness function to measure the fitness value of the GA approach. Finally, GA is applied to find the optimal parameters for a nano-particle milling process. The experimental results show that the integrated approach does indeed find the optimal parameters that result in very good output responses in the nano-particle wet milling process. 2006 Elsevier B.V. All rights reserved.
Keywords: Nano-particle; Wet-type milling process; Taguchi method; Response surface method (RSM); Genetic algorithm (GA)

1. Introduction Nano-particles are advanced materials with 1100 nm grain size. They have been widely applied in photo-catalyst, carbon nano-tube, nano-ceramics, fabric fiber, and compound material industries. The techniques for manufacturing the nano-particles can be classified into top-down and bottom-up methods [1]. The top-down method transforms the material with an initial size of a few micrometers into nano-particles with a size of only 40200 nm. Mechanical ball milling, sputtering, chemical etching and laser ablation are popular top-down methods. The bottom-up method is to generate the nano-particles by heaping up atoms, or assemble the nano-particles from nano-building blocks. It can produce nano-particles in any desired sizes. Examples of bottom-up methods include aerosol compaction, chemical synthesis, chemical vapor deposition and gasatomization methods. No matter what kind of methods are used to fabricate the nano-particles, the nano-particles will
Corresponding author. Tel.: +886 5 5342601x5115; fax: +886 5 5312073. E-mail address: houth@yuntech.edu.tw (T.-H. Hou). 0032-5910/$ - see front matter 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.powtec.2006.11.019

aggregate again because of the effects of Coulomb electrostatic force and Van der Waal force as soon as the grain size of the nano-particles is smaller than 100 nm. In this situation, the nano-particles don't remain at the nanometer size any more. Recently, the wet-type milling machine has been developed to produce the nano-particles and avoid the aggregation effect. It is a popular and suitable method for producing the nano-particles because of its simplicity and applicability to all classes of materials [24]. One of the most popular mechanical milling methods is stirred ball milling which consists of a rotating agitator with grinding ball media. The wet-type mechanical milling process uses the stirring and colliding motions in the stirred ball mill that generates the impact of collisions between milling balls and particles to break the materials into a nanometer size [6,5,7]. Many advanced mechanical milling machines for making nano-particles have been rapidly developed. The vertical chamber milling machine and the horizontal chamber milling machine are the examples. In addition, in order to increase the production capacity of the nano-materials, a lot of techniques or modules are integrated into the mechanical milling machine, such as the design of rotors, geometrical shape of the milling

154

T.-H. Hou et al. / Powder Technology 173 (2007) 153162

mill, the design of the flowing field within a milling mill, the separating system, the circulation and cooling system, the control system, and the user interface etc. The integration of these modules makes the mechanical milling process a complex process. Therefore it is critical to control the multiple variables to achieve the desired product quality. In this research, a vertical chamber milling machine developed by the Precision Machinery Research and Development Center (PMC) is applied to make the titanium oxide (TiO2) nano-particles. The vertical milling machine consists of a mill with 1000 CC. capacity, a vane agitator, a circulation system, a cooling system with 24,000 BTU/h, a filter separating system, a storage tank with 3000 CC. capacity, a power motor with 2400 rpm and 5500 W. Fig. 1 shows the structural diagram of the vertical chamber milling machine. The operation of the milling process is a wet-type and circulated mode. First, the processed material, the titanium dioxide (TiO2), has to be dissolved with a corresponding solvent, glycol, to make a colloidal solution in a suitable weight ratio. Second, this colloidal solution and a dispersant (phosphate) are poured into a stirring tank to mix up together, and sampled to measure the grain size before the milling process is started. Then the colloidal solution with dispersant, and the grinding ball media,

Zircon dioxide (ZrO2), are put into the mill. Forth, the power motor is started to drive the rotating vane agitator. In the mill, the impact of collisions between grinding ball media and particles is produced to break the TiO2 material to a nanometer size. Fifth, samples of colloidal solution are randomly sampled after operating the milling process some hours later to measure the quality responses, mean of grain size and variance of grain size, by using the Coulter Multisizer machine. Finally, the TiO2 nano-particles with nanometer size are produced by separating the solute (TiO2) and solvent (glycol) in a centrifuge machine. The required qualities of the wet-type milling process are that the mean of grain size and the variance of grain size for the nano-particles must be kept small. The parameters that may affect these qualities are the milling time, the flow velocity of circulating system, rotation velocity of agitator, solute-tosolvent weight ratio, filling ratio of grinding media, size of grinding media, grinding media material, mill geometry, mill size, temperature of the milling, type of dispersants and material types of milling milletc. However, the effects of these parameters on the output quality responses in the nano-particle milling process remain unknown. Moreover, it is important to find the optimal milling parameters in order to produce high quality nano-particles.

Fig. 1. The structural drawing of the vertical chamber milling machine developed by PMC.

T.-H. Hou et al. / Powder Technology 173 (2007) 153162

155

Recently, soft computing intelligence techniques, such as artificial neural networks (ANN) and genetic algorithms (GA), and statistical experiment design techniques, such as the Taguchi method and response surface method (RSM), have been widely applied in many engineering optimization problems. Sunada and Bi [8] used polynomial regression and an artificial neural network (ANN) to optimize the formulation and preparation of a rapidly disintegrating tablets process. Reh and Ye [9] applied the neural networks, Taguchi method and multiplier and Lagrangian methods to form the on-line prediction and optimization of a circulating fluidized bed process. They used the orthogonal experiments to gain experimental data and applied the backpropagation neural networks (BPN) to provide on-line predictions. Finally, the multiplier and Lagrangian methods are used to optimize the operation of the process using these BPN trained models. Fan [10] combined a specialized response surface method with state-of-the-art mathematical programming techniques to improve the quality of the chemical-mechanical planarization (CMP) in a semiconductor manufacturing process. He used the RSM as an external modeling technique and applied non-linear programming (NLP) approaches as an optimumseeking procedure to set the optimal parameters. This method could obtain a better polishing quality for the CMP process. Doniavi et al. [11] applied the integrated computer aided manufacturing definition (IDEF), experimental design and RSM to analyze, control and optimize the photolithography process in an electronic manufacturing environment. They used the IDEF technique to describe a manufacturing system with a modeling structure, and then used the RSM approach to suggest the operations with a significant statistical inference to find an optimal response for the given objectives. The optimal settings of the processing equipment resulted in an increase in the process yield. Kwak [12] proposed an application of the Taguchi method and RSM to minimize the geometric error and find the optimal grinding conditions in the surface grinding process. He evaluated the grinding parameters' effect on a geometric error by the Taguchi method and used the RSM to predict the geometric error. Zhou et al. [13] applied the Taguchi method and RSM to optimize the rapid prototyped stereo-lithography parts process parameters and improve the accuracy of the process. Kurtarn et al. [14] used the RSM and GA approaches to optimize the efficient warpage of thin shell plastic parts. The RSM is used to predict the response for warpage data and is then combined with an effective GA to find the optimum process parameter values. Moreover, He et al. [15] pointed out that the process optimization of the flow behavior in a grinding chamber can enhance energy efficiency and throughput in wet ultra-fine grinding operation. The Taguchi method combined with the RSM is a powerful optimum design tool in many engineering applications. It not only can save a lot of time but also can build models quickly and accurately in an optimization design. GA is one of the most powerful and broadly applicable optimization techniques in engineering design problems. Therefore, the objectives of this research are to integrate the parameter design of the Taguchi method, RSM and GA approach to find the critical parameters for a nano-particle milling process and to optimize the process parameters. In the following sections, the research methods

of Taguchi method, RSM and GA are briefly described in Section 2. The proposed integrated approach is then presented in Section 3. The implementation results of the proposed method are then illustrated in Section 4. Finally, concluding remarks are made in Section 5. 2. Research method 2.1. Taguchi method The Taguchi method, proposed by Genichi Taguchi, contains system design, parameter design, and tolerance design procedures to achieve a robust process and result for the best product quality. The purpose of system design procedure is to determine the suitable working levels of the design factors. The parameter design procedure determines the factor levels that can generate the best performance of the product or process under study. The tolerance design procedure is used to fine-tune the results of parameter design by tightening the tolerance levels of factors that have significant effects on the product or process. Regardless of the on-line or off-line engineering, the Taguchi method can efficiently improve the effectiveness of the product or process by using a loss function and achieve the robust product quality in terms of the parameter design. In general, the parameter design of the Taguchi method utilizes orthogonal arrays (OAs) to minimize the time and cost of experiments in analyzing all the factors and uses the signal-to-noise (S/N) ratio to analyze the experimental data and find the optimal parameter combination. Moreover, an analysis of variance (ANOVA) is employed to estimate the error variance and determine the significant parameters. Procedures for conducting a parameter design include the following steps: 1. Planning experiment (1) Determine the control factors, noise factors and quality responses of the product or process. (2) Determine the levels of each factor. (3) Select an appropriate orthogonal array (OA) table. The selection of the most appropriate OA depends on the number of factors and interactions, and the number of levels for the factors. For example, an L27(313) OA can lay out 27 trials, up to 13 factors in columns, and 3 factor levels. (4) Transform the data from the experiments into a proper S/N ratio. 2. Implementing experiment 3. Analyzing and examining result (1) Execute an ANOVA analysis to determine the significant parameters. (2) Conduct a main effect plot analysis to determine the optimal level of the control factors. (3) Execute a factor contribution rate analysis. (4) Confirm experiment and plan future application. 2.2. Response surface method The response surface method (RSM), developed by Box and Wilson in the early 1950s, is a collection of mathematical and

156

T.-H. Hou et al. / Powder Technology 173 (2007) 153162

statistical techniques that are used to model and analyze engineering applications. In these engineering applications, a response of interest is usually influenced by several variables and the objective of the engineering applications is to find the variables that can optimize the response. The RSM has been applied in a wide variety of industrial setting and parameter optimizations such as, chemical, semiconductor and electronic manufacturing, machining, and metal cutting processes. In general, the procedure of RSM consists of the following steps [16]: Step 1. Designing and conducting a series of experiments to get adequate and reliable measurements of the interesting response (e.g. orthogonal array experiment). Step 2. Developing mathematical models of the first and second order response surface with the best fittings. Step 3. Finding the optimal set of process parameters that produce a maximum or minimum value of the response. Step 4. Representing the direct and interactive effects of the process parameters through two and three dimensional plots. If all variables are assumed to be measurable, the response surface can be expressed as follows: Y f x1 ; x2 ; N xn 1

GA approach in solving an optimization problem can be summarized as follows: 1. Represent the problem variable as a chromosome of a fixed length, and choose the size of a chromosome population, the crossover probability, and the mutation probability. 2. Define a fitness function to measure the fitness of an individual chromosome in the problem domain. 3. Randomly generate an initial population of chromosomes. 4. Calculate the fitness of each individual chromosome. 5. Select a pair of chromosomes for mating from the current population. Parent chromosomes are selected with a probability related to their fitness. Highly fit chromosomes have a higher probability of being selected for mating. The roulette wheel method is usually applied in chromosome selection. 6. Create a pair of offspring chromosomes by applying the genetic operators: crossover and mutation. 7. Place the created offspring chromosomes in the new population. 8. Repeat step 5 until the size of the new chromosome population is equal to the size of the initial population. 9. Replace the initial (parent) chromosome population with the new (offspring) population. 10. Repeat steps 49 until the termination criterion is satisfied. 3. The proposed integrated approach In this research, an integrated method is proposed to investigate the effects of the process parameters for a nanoparticle milling process with multiple output responses and to find the optimal process parameters for the process. First of all, the parameter design of Taguchi method is used to determine the significant parameters and set the optimal level for each parameter based on the results of analysis of variance (ANOVA) and the main effect plot. The response surface method (RSM) is then applied to establish the linear and nonlinear multivariate relationships between the nano-particle milling process parameters and the quality responses. Finally, a GA approach is applied to find the optimal parameters by using the response function of the RSM model as the fitness function to measure the fitness value. The flow diagram of the proposed integrated approach is shown in Fig. 2. The detailed procedures of the proposed approach are stated in the following subsections. 3.1. Parameter design and response surface method (RSM) As stated in the Introduction section, a lot of parameters may affect the output quality responses in the nano-particle milling process. Therefore, the optimization of the mechanical milling process is a complex process. But the design experience of engineers show that some parameters have to be considered simultaneously when the milling machine is used to reduce a specific material into the nanometer size particles. For example, the parameters of mill size, mill material and mill geometry are always considered simultaneously in designing a mill. In order to reduce the complexity of the milling parameters, the mill size, mill material and mill geometry are fixed and not considered as the controllable

where n is the number of variables. The goal is to optimize the response variable Y. It is assumed that the independent variables are continuous and controllable by experiments with negligible errors. It is necessary to find a suitable approximation for the true functional relationship between independent variables and the response surface. 2.3. Genetic algorithm The genetic algorithm (GA), introduced by John Holland (1971), is a stochastic search technique based on the mechanism of natural selection and natural genetics to imitate living beings for solving difficult optimization problems with high complexity and an undesirable structure. The GA approach represents a powerful, general-purpose optimization paradigm in which the computational process mimics the theory of biological evolution [17,18]. It has been successfully used in job-shop scheduling, production planning, line balancing, lumber cutting optimization, and process optimization. Goldberg [19] proposed the most common and useful form of GA. Different from traditional point-to-point descending and ascending search techniques, a GA starts from one set of random solutions called a population. Each individual solution in the population is called a chromosome. At each generation, the GA performs genetic operations such as crossover and mutation on the selected chromosomes to yield offspring to produce the next generation. During each generation, these chromosomes evolve into better fitness by applying an evolution operation, called the selection. From generation to generation, eventually, the chromosomes in the population will converge. In this case, the best chromosome is found. Generally, the basic steps of a

T.-H. Hou et al. / Powder Technology 173 (2007) 153162

157

Fig. 2. The flow diagram of the proposed integrated approach.

process parameters in this research. The mill size is set at 1000 CC., the mill material is stainless steel material and the mill geometry is a vertical cylinder. Similarly, the zircon dioxide (ZrO2) is used as the grinding media material and the size of grinding media is set at 500 m. These two parameters are fixed and not considered as the controllable process parameters either. In addition, in order to keep the nano-particles in the nanometer at room temperature, the milling temperature is set at 2025 C and not considered as a control parameter, and the type of dispersants is set to the phosphate. In short, the five controllable process parameters considered in this research include the milling time, flow velocity of circulation system, rotation velocity of agitator shaft, solute-tosolvent weight ratio and filling ratio of grinding media. The choices of the levels on these five parameters are based on the property of the milling process and the limit of the milling machine. For the parameter of the milling time, the milling time should be limited at a reasonable range in order to achieve the milling efficiency and to decrease the degree of contamination. According to the engineers' experience, the reasonable range of milling time is between 2 and 8 hours. Therefore, the levels of the milling time are set at 2, 5 and 8 hours. For the parameter of the flow velocity of circulation system, the factor levels are set based on the mill size and storage tank size. In this research, the mill size

is 1000 CC and the storage tank size is 3000 CC. The maximum volume for the colloidal solution is 4000 CC. Therefore, the levels of the flow velocity of circulation system are set at 1, 2 and 3 (L/ min). For the parameter of rotation velocity of agitator shaft, the factor levels are set based on the limit of the power motor and the abrasion of the vane agitator. In this research, the limit of the rotation velocity of the power motor is 2400 rpm. Therefore, the levels of the rotation velocity of agitator shaft are set at 1200, 1800 and 2400 rpm. For the parameter of the solute-to-solvent weight ratio, the factor levels are set based on the processed material and its corresponding solvent (e.g. TiO2 and glycol). According to the engineers' design experience, if the consistency of the colloidal solution is too high, it will cause the milling equipment to shut down. Therefore, the levels of the solute-to-solvent weight ratio are set at 1, 4 and 7 (wt.%). For the parameter of the filling ratio of grinding media, the factor levels are set based on the mill size. Since the mill volume size is 1000 CC., the filling ratio of grinding media should be set at a reasonable range in order to achieve the milling efficiency. According the engineers' design experience, the levels of the filling ratio of grinding media are set at 5, 15 and 25 percentage of the mill volume. The five process parameters, each with three levels, and some interactions between process parameters are investigated in this research. The five process parameters and their factor levels are summarized in Table 1. The interest output responses of the nano-particle milling process are the mean grain size and the variance of grain size. Both of the output responses need to be kept to a minimum. In order to save on experimental costs and time, the orthogonal array (OA) experiment rather than a full factorial experiment design is applied to obtain the response measurements of the milling process. The selection of an OA to be used depends on the degree of freedom of the factors and interactions. In this research, five process parameters, each with three levels, and the interactions between milling time (A) and rotation velocity of agitator (C), milling time (A) and number of grinding balls (E) are investigated. The calculations of the degrees of freedom are shown as follows: The degrees of freedom of the main factors = (3 1) 5 = 10; The degrees of freedom of the interactions = (3 1) (3 1) 2 = 8. The summation of the degrees of freedom from the main factors and interactions is 18. This means we have to choose an L18 OA at least. However, all of the parameters have three levels, the OA must be a cube of 3 (e.g. L9, L27, L81). Therefore
Table 1 Factors and levels (%: the percentage of mill volume) Factors Levels Level 1 A B C D E Milling time Flow velocity of circulation system Rotation velocity of agitator shaft Solute-to-solvent weight ratio Filling ratio of grinding media (ZrO2) 2(H) 1 (L/min) 1200 (rpm) 1 wt.% 5% Level 2 5(H) 2 (L/min) 1800(rpm) 4 wt.% 15% Level 3 8(H) 3 (L/min) 2400(rpm) 7 wt.% 25%

158

T.-H. Hou et al. / Powder Technology 173 (2007) 153162 Table 2 The L27 (313) orthogonal array and output responses No Process parameters Output responses

an L27(313) OA is used to conduct the experiment in this research. The L27(313) can lay out 27 trials, up to 13 factors in columns, and 3 factor levels. The layout of the L27(313) OA is shown in Table 2. It also shows the two output responses for each experimental trial. The response surface method (RSM) combined with the parameter design is a powerful optimization design tool in many engineering applications. The RSM can be applied to obtain an approximation for a response function in terms of predictor variables. The linear and quadratic models are the most common response models used to build the relationship between the input parameters and output responses. They are generally stated as follows: For the linear (1st order) model: y b0 b1 X1 b2 X2 N bK XK e where K is the number of variables, is the error item. For the quadratic (2nd order) model: y b0
K X i1

A C A C A C E A E A E B e e D e e Mean Variance (grain (grain size) size) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 2 2 2 3 3 3 1 1 1 3 3 3 1 1 1 2 2 2 1 1 1 2 2 2 3 3 3 3 3 3 1 1 1 2 2 2 2 2 2 3 3 3 1 1 1 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 2 3 1 2 3 1 2 3 1 3 1 2 3 1 2 3 1 2 1 2 3 1 2 3 1 2 3 3 1 2 3 1 2 3 1 2 2 3 1 2 3 1 2 3 1 1 2 3 2 3 1 3 1 2 1 2 3 2 3 1 3 1 2 1 2 3 2 3 1 3 1 2 1 2 3 2 3 1 3 1 2 2 3 1 3 1 2 1 2 3 3 1 2 1 2 3 2 3 1 1 2 3 2 3 1 3 1 2 3 1 2 1 2 3 2 3 1 2 3 1 3 1 2 1 2 3 1 2 3 3 1 2 2 3 1 1 2 3 3 1 2 2 3 1 1 2 3 3 1 2 2 3 1 1 2 3 3 1 2 2 3 1 2 3 1 1 2 3 3 1 2 3 1 2 2 3 1 1 2 3 1 2 3 3 1 2 2 3 1 3 1 2 2 3 1 1 2 3 2 3 1 1 2 3 3 1 2 232.7 182.5 154.0 181.0 159.2 224.8 169.1 171.8 207.2 217.9 174.2 133.0 164.5 147.6 158.8 156.7 169.1 142.6 213.2 174.4 130.3 151.0 138.7 152.3 148.1 165.7 138.6 8.5 21.1 5.7 0.8 1.2 8.4 5.1 5.9 6.6 15.9 16.6 1.8 6.1 5.7 5.2 4.2 6.5 3.8 8.3 10.6 2.6 2.9 5.7 4.7 2.5 5.6 3.5

bi Xi

K X i1

bii Xi2 (3)

XX
ibj

bij Xi Xj e

3.2. The integrated GA and RSM approach In this research, an integration of GA and RSM is applied to find the optimal process parameters in a nano-particle milling process. The RSM is used to establish the linear and nonlinear relationships between the milling process parameters and the responses. The GA approach is then applied to find the optimal process parameters using the RSM response models as the fitness function to measure the fitness value for the process parameters. The detailed procedures of the integrated GA and RSM approach are stated as follows: Step 1. Encoding and generating the initial populations The process parameters are used to create the solution space for the GA approach. The searching ranges for the five parameters are shown in Table 1. Following normalization, each searching range is divided into 1000 equal intervals. The binary string representation for coding space (chromosome) is adopted for this GA approach and each input parameter is encoded into 10 binary digits. For example, if X1 = 0.010, X2 = 0.036, X3 = 0.064, X4 = 0.256 and X5 = 0.065, then the chromosome will be 0000001010,0000100100,0001000000, 0100000000,0001000001. Following the encoding process, we randomly generate Npop chromosomes to find the optimal process parameters (where Npop is the number of populations). Step 2. Calculate the fitness value The RSM response functions are used as the fitness function in the GA approach to calculate the fitness values. In this research, the required qualities of the nanoparticle milling process are that the mean grain size and

the variance of grain size of the nano-particle must be kept small. Therefore, these two output responses result in a two-objective optimization problem, i.e., minimizing both the mean grain size and the diameter variation. Because these two output responses have to be minimized, we can convert the multi-objective functions into a single objective using two strategies. One strategy is to merge these two output responses into one response by aggregating operator, and the other strategy is to combine them into one response by using the statistical operator, coefficient of variance. For example, if the RSM model of response 1 Y1 is and the RSM model of response 2 is, Y2 then the final fitness function used to calculate the fitness values by using the two strategies will be as follows: Fitness function of the first strategy Y1 Y2 4 Fitness function of the second strategy variance Y2 mean Y1 5

Step 3. Selection Select a pair of chromosome for mating, e.g. crossover and mutation, from the current population. Parent chromosomes are selected with a probability related to

T.-H. Hou et al. / Powder Technology 173 (2007) 153162 Table 3 The ANOVA table of mean grain size Source A B C D E AC AE Error Total Sum of square 1670.8 5613.2 2650.8 936.43 4780.6 65.21 143.97 57.11 15918 Degrees of freedom 2 2 2 2 2 4 4 8 26 Mean square 835.4 2806.6 1325.4 468.2 2390.3 16.3 35.99 7.14 F0

159

their fitness. Highly fit chromosomes have a higher probability of being selected for mating. The roulette wheel method is applied to the chromosome selection in this research. Step 4. Crossover The crossover operator is used to create a pair of offspring chromosomes. For each selected pair, a twocut-point crossover operation is applied to generate an offspring with the crossover probability Pc. Step 5. Mutation In this study, the one-gene mutation operation with a preset mutation probability Pm is applied to generate new chromosomes. By using the crossover and mutation operations, new Npop(offspring) populations are created. Replace the initial chromosome populations with the new Npop(offspring) populations and repeat Steps 2 3 4 5 until the termination criterion is satisfied. 4. Results and discussions 4.1. Results of the orthogonal array experiment and RSM Results of the orthogonal array experiment are shown in Table 2. The mean grain size and the variance of grain size are the

117.00 393.10 185.70 65.59 334.80 2.28 5.04

measurements from the Coulter Multisizer equipment. In each experiment, a sample of colloidal solution is randomly sampled and put into the Coulter Multisizer equipment to measure the mean grain size and the variance of grain size. For the sample of colloidal solution in the Coulter Multisizer equipment, the measurement of grain size is repeated five times and then the mean and variance are calculated from these 5 measurements. For example, the No. 21 output response of the orthogonal array experiment (shown in Table 2) is the measured result using the Coulter Multisizer equipment which is shown in Fig. 3. Each experimental trial results in one measurement. Due to the limitation of experimental cost, one trial was conducted for each

Fig. 3. The measured results of mean grain size and variance of grain size for the No. 21 orthogonal array experiment by using the Coulter Multisizer equipment.

160

T.-H. Hou et al. / Powder Technology 173 (2007) 153162

Fig. 4. Main effect plot of response 1.

experiment. Analysis of variance (ANOVA) is then applied to determine the significant parameters and the main effect plot is used to set the optimal level for each parameter in the nanoparticle milling process. The MINITAB 13.0 software is used to analyze the experimental data. Table 3 shows the ANOVA results with respect to the mean grain size. Because the F0.05(2,8) = 4.46 and F0.05(4,8) = 3.84, it is obvious to conclude that milling time (A), flow velocity of circulation system (B), rotation velocity of agitator (C), solute-to-solvent weight ratio (D), number of grinding balls (E) and interaction between milling time and number of grinding balls (A E) significantly affect the mean grain size with 95% confidence intervals. Based on the main effect plot, shown in Fig. 4, the optimal level of each parameter is set at A3B3C3D3E3 for the mean grain size response. Similarly, Table 4 shows the ANOVA results with respect to variance of grain size. It can be clearly seen that flow velocity of circulation system (B), rotation velocity of agitator (C ), solute-to-solvent weight ratio (D) and number of grinding balls (E) have significant effects on the variance of grain size response. Based on the main effect plot, shown in Fig. 5, the optimal level of each parameter is set at A3B3C2D3E3. Moreover, RSM is used to build the relationship between the process parameters and output responses. Both the 1st order model and the 2nd order model of RSM are conducted using the Design Expert version 6.0 software. The results are stated as follows: (1) Mean grain size (Response 1): For the 1st order model (Linear, R2 = 0.929, Adjusted R2 = 0.877): Y1 279:233:14X1 17:58X2 0:018X3 2:11X4 1:63X5 6 For the 2nd order model (Quadratic, R2 = 0.986, Adjusted R2 = 0.907): Y1 380:997:15X1 17:58X2 0:12X3 9:67X4 1:63X5 2 2 2 0:4X1 0:0000284X3 0:91X4 7

(2) Variance of grain size (Response 2): For the 1st order model (Linear, R2 = 0.454, Adjusted R2 = 0.324): Y2 14:260:0046X3 8

For the 2nd order model (Quadratic, R2 = 0.826, Adjusted R2 = 0.624): Y2 36:251:8X2 0:035X3 3:66X4 1:0X5 2 2 2 0:0000085X3 0:5X4 0:036X5

It is found that the 2nd order models yield a larger R2 value than the 1st order models in both responses. Therefore, we can infer that the relationships between the process parameters and output quality responses are nonlinear for both the mean grain size and the variance of grain size. The 2nd order model responses, Eqs. (7) and (9), are used as the fitness function to measure the fitness value in the GA approach. In addition, the 2nd order model responses, Eqs. (7) and (9), can be used to test the sensibility of the process parameter for the mean of grain size and the variance of grain size respectively. Let us take the Eq. (7) as an example, the coefficient of X2 (flow velocity of circulation system) is 17.58, it means if we increase one unit flow velocity of circulation system, the corresponding response Y1, the mean grain size of nano-

Table 4 The ANOVA table of variance of grain size Source A B C D E AC AE Error Total Sum of square 21.49 85.06 193.86 81.92 94.63 33.22 27.86 52.26 590.3 Degrees of freedom 2 2 2 2 2 4 4 8 26 Mean square 10.745 42.53 96.93 40.96 47.315 8.305 6.965 6.532 F0 1.64 6.51 14.84 6.27 7.24 1.27 1.07

T.-H. Hou et al. / Powder Technology 173 (2007) 153162

161

Fig. 5. Main effect plot of response 2.

particles, will reduce 17.58 nm. Similarly, the coefficients of X1, X3, X4 and X5 can be used to test the sensitivity for each process parameter for the mean of grain size using Eq. (7). Similarly, the Eq. (9) can be used to test the sensibility of the process parameters, X2, X3, X4 and X5, for the variance of grain size. 4.2. Optimal parameter setting using integrated GA and RSM approach The MATLAB 7.0 software is applied to develop the integrated GA and RSM approach and determine the optimal parameters in the nano-particle milling process. The 2nd order models of responses are used as a fitness function to measure the fitness value in the GA approach. In this research, the two response models are transformed into one single model. One transformation only adds the two models together and the other transformation uses the coefficient of variance. By using these two transformations, the new fitness functions are defined as follows: For the fitness function 1 which uses the add together transformation: F1 417:247:15X1 19:38X2 0:085X3 6:01X4 0:63X5 2 2 2 2 0:4X1 0:0000396X3 0:41X4 0:036X5 10 For the fitness function 2 which uses the coefficient of variance transformation F2 Y2 Eq:9 Y1 Eq: 7 11

Table 5 shows the five process parameter values and their corresponding output responses for the two fitness function transformations. The mean grain size and variances of grain size of the optimal parameters are clearly smaller than the values in Table 2. In addition, the obtained mean grain size is close to 100 nm. Therefore, we can conclude that the proposed approach does find the optimal process parameters that result in satisfactory output responses. Moreover, the obtained parameter values are continuous values. In the comparison of the optimal process parameters derived from the optimization procedure with the previously obtained optimal level form the main effect plots, the obtained A parameter value falls between level 2 and level 3. The B parameter value is closed to level 3. The C parameter is closed to level 2 for the fitness transformation strategy 1 and falls between level 2 and level 3 for the fitness transformation strategy 2. The D and E parameter values are closed to level 3. The results are similar to the result of the main effect plot. However, the proposed approach obtains more precise parameter values than the main effect plot. It is also clearly found from Table 5 that different transformation strategies of GA's fitness function result in different optimal solutions. Using the coefficient of variance fitness function results in a better optimal solution in mean grain size than the aggregating strategy, but it results in a worse solution in variance of grain size than the aggregating strategy. Although these two output responses have to be minimized, they are conflict. According to the engineering practices in producing nano-particles, the response of mean grain size is more important than the response of variance of grain size. Therefore, the optimal process parameters of the wet-type nano-particle milling process are set as follows: milling time (A) = 6.27, flow velocity of circulation system (B) = 2.96, rotation velocity of

These two new fitness functions are then used to measure the fitness value for each chromosome in the GA procedure. In GA's procedure, the population size is set at Npop = 30. Twocut-point crossover is employed with the crossover probability 0.98. The one-gene mutation operation probability is specified as Pm = 0.015 and the iteration number is 2000. Under these conditions, the best solutions are summarized in Table 5.

Table 5 The best solutions found by the integrated RSM and GA approach Parameter X1 (A) Strategy 1 Strategy 2 6.42 6.27 X2 (B) 2.97 2.96 X3 (C) 1793 1968 X4 (D) 6.96 5.80 X5 (E) 24.86 24.65 Response Y1 109.55 108.13 Y2 0.24 0.54

162

T.-H. Hou et al. / Powder Technology 173 (2007) 153162

agitator (C) = 1968, solute-to-solvent weight ratio (D) = 5.8, and number of grinding balls (E) = 24.65. 5. Conclusions Using a wet-type mechanical milling process to produce nano-particles is a popular method because of the applicability to all classes of materials. The required qualities of the milling process are that the grain size and the variance of grain size of the nano-particle must be kept small. There are several parameters that may affect these qualities. However, the effects of these parameters on the output quality responses in the nanoparticle milling process remain unknown. Moreover, it is also very important to improve the milling process capability and quality by setting optimal milling parameters in the nanoparticle milling process. In this research, the parameter design of the Taguchi method, response surface method (RSM) and genetic algorithm (GA) are integrated and applied to set the optimal parameters for a nanoparticle milling process. The orthogonal array experiment is conducted to economically obtain the response measurements and analysis of variance (ANOVA) and main effect plot are used to determine the significant parameters and set the optimal level for each parameter. The RSM is then used to build the relationship between the input parameters and output responses, and used as the fitness function to measure the fitness value in the GA approach. Finally, GA is applied to find the optimal parameters for the nano-particle milling process. In this research, there are two output responses and therefore, there are two response models. The two response models are converted into a single fitness function. The experimental results show that these two transformations can obtain similar optimal parameters in a nano-particle milling process. However, this research also finds that there are conflicts between the two output responses. The optimal process parameters that result in a smaller mean grain size may unfortunately result in a larger variance of grain size. Therefore, multi-objective optimization techniques may be applied to solve this problem in future research. Acknowledgements This research was partially supported by the Precision Machinery Research and Development Center (PMC) and the National Science Council of the Republic of China under grant number NSC 93-2218-E-224-002 and NSC 94-2218-E224-001. The authors also like to show their appreciation to the anonymous reviewers very much for their valuable

comments that are very helpful for the improvement of this paper. References
[1] R.V. Ramanujan, Nanostructured electronic and magnetic materials, Sadhana 28 (2003) 8196. [2] C.C. Koch, Synthesis of nanostructured materials by mechanical milling: problems and opportunities, Nanostructured Materials 9 (1997) 1322. [3] C. Suryanarayana, Mechanical alloying and milling, Progress in Materials Science 46 (2001) 1184. [4] C. Frances, On modelling of submicronic wet milling processes in bead mills, Powder Technology 143 (2004) 253263. [5] H. Mori, H. Mio, J. Kano, F. Saito, Ball mill simulation in wet grinding using a tumbling mill and its correlation to grinding rate, Powder Technology 143 (2004) 230239. [6] S. Mende, F. Stenger, W. Peukert, J. Schwedes, Mechanical production and stabilization of nanoparticles in stirred ball mills, Powder Technology 132 (2003) 6473. [7] D.L. Zhang, Processing of advanced materials using high-energy mechanical milling, Progress in Materials Science 49 (2004) 537560. [8] H. Sunada, Y. Bi, Preparation, evaluation and optimization of rapidly disintegrating tablets, Powder Technology 122 (2002) 188198. [9] L. Reh, H. Ye, Neural networks for on-line prediction and optimization of circulating fluidized bed process steps, Powder Technology 111 (2000) 123131. [10] S.K. Fan, Quality improvement of chemicalmechanical wafer planarization process in semiconductor manufacturing using a combined generalized linear modeling/ non-linear programming approach, International Journal of Production Research 38 (2000) 30113029. [11] A. DoniaVi, A.R. Mileham, L.B. Newnes, A systems approach to photolithography process optimization in an electronics manufacturing environment, International Journal of Production Research 38 (2000) 25152528. [12] J.S. Kwak, Application of Taguchi and response surface methodologies for geometric error in surface grinding process, International Journal of Machine Tools and Manufacture 45 (2005) 327334. [13] J.G. Zhou, D. Herscovici, C.C. Chen, Parametric process optimization to improve the accuracy of rapid prototyped stereolithography parts, International Journal of Machine Tools and Manufacture 40 (2000) 363379. [14] H. Kurtaran, T. Erzurumlu, Efficient warpage optimization of thin shell plastic parts using response surface methodology and genetic algorithm, International Journal of Advanced Manufacturing Technology 27 (2006) 468472. [15] M. He, Y. Wang, E. Forssberg, Parameter effects on wet ultrafine grinding of limestone through slurry, Powder Technology 161 (2006) 1021. [16] R.M. Myers, D.C. Montgomery, Response Surface Methodology, 2nd ed. Wiley, New York, 2002. [17] J.H. Holland, Genetic algorithms, Scientific American 267 (1) (1992) 6672. [18] R. Storn, K. Price, Differential evolutiona simple and efficient heuristic for global optimization over continuous spaces, Journal of Global Optimization 11 (1997) 341359. [19] D.E. Goldberg, Genetic algorithms in search, Optimization and Machine Learning, Addison-Wesley, 1989.