Introduction to Design of Experiments

by Michael Montero

University of California at Berkeley Mechanical Engineering Department Summer, 2001

Introduction to DOE

Part 1 Full Factorial Design and Analysis Part 2 Fractional Factorial Design and Analysis Part 3 3-Level and Mixed-Level Design and Analysis

M. G. Montero

UC-Berkeley, Mechanical Engineering

DOE Capability
Q: Why perform them? A: When deterministic reasoning on a system can not accurately describe all contributing factors that simultaneously effect the systems response for given a range of conditions. DOE Capability
Treatment or factor level comparison Detecting significant variables which effect the response the most Variable screening Detect most active main effects from larger set of variables Initial experiment followed by detailed DOE Response surface exploration Model building System optimization Finding minima, maxima to optimize response System robustness Identify robust factors to minimize variation in response
M. G. Montero

UC-Berkeley, Mechanical Engineering

Full Factorial Design

Factorial designs test all combinations of factors Columns are orthogonal to one another (estimates of effects can be computed) Example: Three variables (A, B, C) having 2,3, and 5 levels respectively. So: Total runs = 2 x 3 x 5 = 30 Example: Three variables (A, B, C) each having 2 levels. So: Total runs = 2 x 2 x 2 = 23 = 8
UC-Berkeley, Mechanical Engineering M. G. Montero

Epitaxial Growth Layer Experiment

Epitaxial Growth Layer DOE
Kackar, R.N. and Shoemaker, A.C., 1986

Semiconductor Equipment for Epitaxial Growing

Goal: Adjust process parameters to maintain as close as possible to nominal thickness (14.5 m) while minimizing variation. Responses: 1) Signal = Average thickness measurements 2) Variation = Dispersion of thickness observations

14.5 m

Statistical Operators Rules for Random Variables and Constants

Given: y is r.v. and a, b are constants

E(L(y,o )) = E (cy 2 ) E (2cyo ) + E (co 2 ) E(L(y,o )) = cE ( y 2 ) 2co E ( y ) + co 2

14 14.5 15 Epitaxial Layer (m)

1) E(ay + b) = aE(y) + b 2) Var(y) = Var(y b) 3) Var(ay b) = a2var(y) 4) E(y2) = Var(y) + [E(y)]2

E(L(y,o )) = cVar ( y ) + c[ E ( y )]2 2co E ( y ) + co 2 E(L(y,o )) = cVar ( y ) + c[ E ( y )]2 2co E ( y ) + co 2 E(L(y,o )) = cVar ( y ) + c( E ( y ) o ) 2
M. G. Montero

UC-Berkeley, Mechanical Engineering

= cy 2 2cyo + co 2 E(L(y,o )) = E (cy 2 2cyo + co 2 )

L(y,o ) = c(y o )2

L(y,o)

Taguchi Loss Function:

Full Factorial DOE for Epitaxial Growth Layer

Design: 24 = 16 runs

Response 1: Average Thickness

Response 2: Natural log of variance

Factors

Replicates

Randomized Tests
M. G. Montero

UC-Berkeley, Mechanical Engineering

Alternative Matrix Calculation of Effects

Effect Columns Responses

1) Multiply desired effect column by response column element by element 2) Sum terms 3) Divide sum by total number of positive or negative 1s
Example (Layer Thickness Response) 1) Column ABCD x yave = {+13.59, -14.59,-14.05, +14.24, -13.94, .} 2) Sum = 13.59 + (-14.59) + (-14.05) + 14.24 + (-13.94) + ... 3) Effect of ABCD = Sum/8 = 0.0359
M. G. Montero

UC-Berkeley, Mechanical Engineering

To Calculate a Variable Effect

Graphical Analysis of Factorial Plots - Main Effects

Layer Thickness Response

How to Plot (Example: Main Effect A)

1 ythickness = AVE + (Main Effect of A)x A 2

where xA is either +1 or -1
M. G. Montero

UC-Berkeley, Mechanical Engineering

Graphical Analysis of Factorial Plots - Interaction Effects

Layer Thickness Response

B+

A-

B-

A+
UC-Berkeley, Mechanical Engineering M. G. Montero

How to Plot (Example: Interaction Effect BC)

1 1 ythickness = AVE + (Main Effect of B)xB + (Main Effect of C)xC 2 2 1 + (Interaction Effect of BC)x BC 2

Confidence Interval (CI) Construction

CIs for effects will determine whether an effect is significant or not by determining whether 0 is included within the interval:
CI Construction

Effect t DOF,(1 / 2 ) seff

Effect t80,.975 .141

Effect 1.990 .141
Effect 0.2806

t = student-t statistic = confidence (typically .05) DOF = total DOF from observations Effect = Any effect value

Example for Thickness Response

CI for D: 0.8356 0.2806 (0.555,1.1162) 0 is not within interval, therefore, D is significant CI for A: -0.0541 0.2806 (-0.3347,0.2265) 0 is within interval, therefore, D is not significant
M. G. Montero

UC-Berkeley, Mechanical Engineering

Fundamental Principles for Factorial Effects

M. G. Montero

UC-Berkeley, Mechanical Engineering

Hierarchical Ordering Lower order effects are more likely to be important than higher order effects Probability of higher order effects being significant is low Analogous to Taylor expansion series Effects of the same order are equally likely to be important Typically in mechanical systems, 3-factor interactions or higher tend have a lower tendency of appearing significant within DOEs Effect Sparsity (Box and Meyer, 1986) Number of relatively important effects in a factorial experiment is small vital few and the trivial many Effect Heredity In order for an interaction to be significant, at least one of its parent factors should be significant (Wu and Hamada, 1992) Predictive models which contain interactions should always include main effects of the interacting variables even if main effects are not significant.

Introduction to DOE - Part 2

Part 1 Full Factorial Design and Analysis Part 2 Fractional Factorial Design and Analysis Part 3 3-Level and Mixed-Level Design and Analysis

M. G. Montero

UC-Berkeley, Mechanical Engineering

Confounding Controllable Factors with Uncontrollable Factors

Notion of Confounding
The inability to discern one effect from another on a systems response. Epitaxial Growth Layer DOE Effect of Rotation Method + Effect of Humidity Growth Layer Response Factor A is confounded with ambient factor Cannot determine which effect is truly influencing the growth layer

time

30 Humidity(%) 25 20 15 8 am Noon time

M. G. Montero

4 pm

UC-Berkeley, Mechanical Engineering

Fractional Factorial Designs

25
Run 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 28 29 30 31 32 A + + + + + + + + + + + + + + + + B + + + + + + + + + + + + + + + + C + + + + + + + + + + + + + + + + D + + + + + + + + + + + + + + + + E + + + + + + + + + + + + + + + + Run 1 2 3 4 5 6 7 8 A + + + +

5 2III 2
B + + + + C + + + + D + + + + E + + + +

Why Fractional Factorial Designs?

Test same number of factors but at lower number of runs Time and cost is reduced is most cases Used as variable screen out experiments

Give and Take Situation

Two ways of looking at fractional designs Reduce number runs Design resolution is lowered Add more factors Hierarchical order of effects (higher order less likely to be significant than lower order) A priori information or knowledge of physical system in question (clamping force) x (plastic moisture content) negligible (chemical pigment type) x (operator) negligible

M. G. Montero

UC-Berkeley, Mechanical Engineering

Designing 2k-p Fractional Factorial Designs

2k-p: Two-level fractional factorial design with k + total number of factors, p + number of added factors or 1/2p the fraction of runs Example:

Design : 2

4 1 IV

Basic Factors

Added Factors Added factor D is confounded or aliased with three factor interaction of basic factors A, B, and C. Column D is generated by the column multiplication of A, B, and C.
UC-Berkeley, Mechanical Engineering M. G. Montero

D + ABC, Design Generator D(D) + (ABC)D, Multiply both sides by D I + ABCD, Defining Relation Ave + ABCD Effect A + BCD Effect B + ACD Effect C + ABD Effect D + ABC Effect AB + CD Effect AC + BD Effect AD + BC Word Defining relation is used to identify confounding patterns within effect estimates. Multiply effect variable through defining relation.

Using the Defining Relation

I+ABCD AB(I) + AB(ABCD) AB + AABBCD AB + CD Effect AB + CD are confounded or aliased

Smallest word within defining relation determines resolution of design I + ABCD Word length is 4, therefore resolution is IV:
4 2IV1
UC-Berkeley, Mechanical Engineering

Resolution will quickly tell you to what degree of confounding exists within your design of experiment:
Resolution III: Main effects and 2-factor interactions are confounded Resolution IV: Main effects and 3-factor interactions are confounded and 2-factor interactions are confounded with each other Resolution V: Main effects and 4-factor interactions are confounded and 2-factor interactions are confounded with 3-factor interations Ave + ABCD Effect A + BCD Effect B + ACD Effect C + ABD Effect D + ABC Effect AB + CD Effect AC + BD Effect AD + BC
M. G. Montero

Employing the Hierarchical Principle in Analysis

Example: 6 2IV 2
Design Generators E + ABC I + ABCE F + BCD I + BCDF
Assume 3-factor intrxns or higher are negligible. Parent main effects of CE are active, therefore, CE is most likely significant.

Confounding Patterns and Effect Estimates

A + BCE + DEF + ABCDF + -0.300 B + ACE + CDF + ABDEF + 0.175 C + ABE + BDF + ACDEF + 8.675 D + AEF + BCF + ABCDE + 0.025 E + ABC + ADF + BCDEF + 4.350 F + ADE + BCD + ABCEF + -0.175 AB + CE + ACDF + BDEF + 2.575 AC + BE + ABDF + CDEF + 0.275 AD + EF + ABCF + BCDE + -0.125 AE + BC + DF + ABCDEF + 0.200 AF + DE + ABCD + BCEF + -0.025 BD + CF + ABEF + ACDE + -0.100 BF + CD + ABDE + ACEF + 0.550 ABD + ACF + BEF + CDE + -0.200 ABF + ACD + BDE + CEF + -0.350
UC-Berkeley, Mechanical Engineering M. G. Montero

Defining Relation I + ABCE + BCDF + ADEF

Introduction to DOE - Part 3

Part 1 Full Factorial Design and Analysis (2 levels) Part 2 Fractional Factorial Design and Analysis (2 levels) Part 3 Software Introduction and 3-Level or Higher Designs

M. G. Montero

UC-Berkeley, Mechanical Engineering

Available DOE Software

Commercial Software for Experimental Design
SAS JMP S-Plus Genstat Minitab State-Ease, Design-Expert Echip Statgraphics Systat Umetrics MODDE 6 Mixsoft Nutek Qualitek-4 StatSoft General Statistical Package Adept Scientific DOE_PC IV Process Builder STRATEGY S-Matrix CARD Qualitron Systems DoES RSD Associates Matrex

Example: Minitab v11.21

DOE Specific
UC-Berkeley, Mechanical Engineering M. G. Montero

Windows Based (Windows 9x, NT, and 2000) Spreadsheet-like interface and command line interface User-friendly menus 2k full and fractional factorial designs (regular and non-regular) Response surface building Analysis of Variance (ANOVA) Multiple linear regression Statistical Process Control (SPC), time-series analysis (autoregression) Reproducibility and Repeatability (R&R) And more...

Minitab: Create Factorial Design Step by Step

M. G. Montero

UC-Berkeley, Mechanical Engineering

Minitab: Factorial Designs Dialog Box

Summary of Possible 2-Level Designs Predefined Designs

2

Custom Designs Screening Design

4

Select # of Factors

5
UC-Berkeley, Mechanical Engineering M. G. Montero

Design Selection

Minitab: Summary of 2-Level Designs

M. G. Montero

UC-Berkeley, Mechanical Engineering

Further Topics: 3-Level Fractional Factorial Designs

3k-p designs rely on generators and defining relation not based on multiplicative column but modulus calculus:
Example: 34-1 (Generator: D = ABC) where xD = xA + xB + xC (mod 3) Where: x = coded value (0, 1, or 2) So: 3/3 = 1 remainder 0 1/3 = 0.3 remainder 1 2/3 = 0.6 remainder 2
UC-Berkeley, Mechanical Engineering M. G. Montero

Column Ds coded pattern is generated by the xA + xB + xC (mod 3) relation

Further Topics: 2m4n Mixed Designs

2m4n designs can be generated from fractional factorial 2k-p designs by method of column replacement:
Example: 27-4 (Generators Not Shown)

2441
UC-Berkeley, Mechanical Engineering M. G. Montero

Statistical Literature
Experimental Design and Optimization
Box, G. E. P., Hunter, W. G. and Hunter, J.S., Statistics for Experimenters: An Introduction to Design, Data Analysis, and Model Building, Wiley Series in Probability and Statistics, 1978. Devor, R. E., Chang, T. and Sutherland, J. W., Statistical Quality Design and Control: Contemporary Concepts and Methods, Macmillan, 1992. Ross, P. J., Taguchi Techniques for Quality Engineering, McGraw Hill, 2nd Edition, 1996. Wu, C. F. J. and Hamada, M., Experiments: Planning, Analysis, and Parameter Design Optimization, Wiley Series in Probability and Statistics, 2000. Myers, R. H. and Montgomery, D. C., Response Surface Methodology: Process and Product Optimization Using Designed Experiments, Wiley Series in Probability and Statistics, 1995

Statistics and Multiple Linear Regression

Walpole, R. E., Myers and R. H., Myers, S. L., Probability and Statistics for Engineers and Scientists, Prentice Hall, 6th edition, 1998. Sen, A. and Srivastava, M., Regression Analysis: Theory, Methods, and Applications, SpringerVerlag, 1990.

M. G. Montero

UC-Berkeley, Mechanical Engineering