AD-A233 316

AFWAL-TR-81-4097

REFERENCE DOCUMENT FOR THE ANALYSIS OF CREEP

AND STRESS-RUPTURE DATA IN MIL-HDBK-5

Richard Rice
Battelle-Columbus Laboratories
505 King Avenue
Columbus, Ohio 43201

September 1981

Interim Report for Period June 1980-June 1981

Approved for public release; distribution unlimited.

DTIC .

MATERIALS ILABORATORY
AIR FORCE WRIGHT AERONAUTICAL LABORATORIES
AIR FORCE SYSTEMS COMMAND
WRIGHT-PATTERSON AIR FORCE BASE, OHIO 45333
S pELECTE
L RE0 99_
APR 0..) 1991
D
914 08071
 NOTICE

When Government drawings, specifications, or other data are used for

any purpose other than in connection with a definitely related Government pro-
curement operation, the United States Government thereby incurs no responsi-
bility nor any obligation whatsoever; and the fact that the Government may
have formulated, furnished, or in any way supplied the said drawings, specifi- 4
cations, or other data, is not to be regarded by implication or otherwise as
in any manner licensing the holder or any other person or corporation, or con-
veying any rights or permission to manufacture, use, or sell any patented .L
invention that may in any way be related thereto.

This report has been reviewed by the Office of Public Affairs

(ASD/PA) and is releasable to the National Technical Information Service
(NTIS). At NTIS, it will be available to the general public, including
foreign nations.

This technical report has been reviewed and is approved for publication.

C. L. Harmsworth, Technical Manager

for Engineering and Design Data
Materials Integrity Branch
Systems Support Division
Air Force Wright Aeronautical Laboratory

FOR THE COMMANDER

T. D. Cooper, ief
Materials In rity Branch
Systems Support Division
Air Force Wright Aeronautical Laboratory

If your address has changed, if you wish to be removed from our mailing
list, or if the addressee is no longer employed by your organization, please
notify AFWAL/MLSA WPAFB, OH 45433 to help us maintain a current mailing list.

Copies of this report should not be returned unless return is

required by security considerations, contractual obligations, or notice on a
special document.
 1JNL"LASSIFTE1
SECURITY CLASSIFICATION OF THIS PAGE (When Date Entered,
REPORT DOCUMENTATION PAGE READ INSTRUCTIONS
BEFORE COMPLETING FORM
I. REPORT NUMBER 2. GOVT ACCESSION NO. 3. RECIPIENT'S CATALOG NUMBER
AFWAL-TR- 81- 409 7
4. TITLE (end Subtitle) S. TYPE OF REPORT & PERIOD COVERED

REFERENCE DOCUMENT FOR THE ANALYSIS OF CREEP Interim Report

AND STRESS RUPTURE DATA IN MIL-HDBK-5 June 1980- June 1981
6. PERFORMING ORG. REPORT NUMBER

7. AUTHOR(s) 6. CONTRACT OR GRANT NUMBER(s)

Richard C. Rice, Editor F33615-80-C-5037

9. PERFORMING ORGANIZATION NAME AND ADDRESS 10. PROGRAM ELEMENT, PROJECT, TASK
Battelle's Columbus Laboratories AREAS WORK UNIT NUMBERS
505 King Avenue FY1457-81-04013
Columbus, Ohio 43201
It. CONTROLLING OFFICE NAME AND ADDRESS 12. REPORT DATE

DCASMA, Dayton- Defense Logistics Agency, September 1981

Attention: DCRO-GDCA, c/o Defense Electronics 13. NUMBER OF PAGES
Supply Center, Dayton. Ohio 45444 F.9
14. MONITORING AGENCY NAME & ADDRESS(if dilferert from Controlling Office) IS. SECURITY CLASS. (of this report)

Materials Laboratory (AIVAL/1ILSA) Unclassified

Air Force Wright Aeronautical Laboratories (AFSC) E
Wright-Patterson Air Force Base, Ohio 45433 SCHEDULE

16. DISTRIBUTION STATEMENT (of this Report)

Approved for public release; distribution unlimited.

17. DISTRIBUTION STATEMENT (of the abstract entered in Block 20, it different from Report)

IS. SUPPLEMENTARY NOTES

19. KEY WORDS (Continue on reverse side if necesesry and identify by block number)

,t:reep .analysis of varianc

/: .- ,,stress rupture
design of experiments
regression analysis $-

20. ABSTRACT (Continue on reverese side If neceesry end identify by block number)

This report provides background information on creep and stress rupture

analysis which expands on the methods described in MIL-HDBK-5 guidelines.
The document includes discussions on (1) design of experiments for the purpose
of developing regression models, (2) important correlative information for
use in evaluating elevated temperature property data, and (3) a comprehensive
method of rupture data analysis.
j

DD 1473 EDITION OF I NOV 65 IS OBSOLETE UNCLASSIFIED

/ SECURITY CLASSIFICATION OF THIS PAGE (B'hen Date Entered)
 PREFACE

This interim report was compiled by Battelle's Columbus Laboratories,

505 King Avenue, Columbus, Ohio 43201, under contract F33615-80-C-5037 with
the Air Force Wright Aeronautical Laboratory, Wright-Patterson Air Force Base,
Ohio, as a part of the ongoing MIL-HDBK-5 coordination effort. Mr. C. L.
Harmsworth (MLSA) was the project monitor. This report was submitted by the
author, Mr. Richard Rice, in June 1981.
Major inputs to the document were provided by Lee Green of Pratt &
Whitney Aircraft (Section 2), Charles Maak, formerly of Sandia Corporation,
now retired (Section 3), and Lars Sjodahl, formerly of General Electric,
Evendale, now also retired (Section 4). Other important participants in the
Elevated Temperature Task Group who played a part in the development of the
creep and stress rupture guidelines in this report are:

Matthew Rebholz - Lockheed Missiles and Space Co., Sunnyvale

Gene Best and -- General Electric, Evendale

Jerry Cashman

Lou Fritz - Metcut Research, Cincinnati

Dennis Wolski -- AiResearch, Phoenix

Marty Marchbanks -- Westinghouse Hanford Company, Richland

Walter Hyler -- Battelle's Columbus Laboratories

James Laflen - Formerly Battelle, presently General Electric,

Evendale

Robert Stusrud -- Detroit Diesel, Allison Division, Indianapolis

i
 TABLE OF CONTENTS

Page

1.* INTRODUCTION............ ........................ 1

2. DESIGN OF EXPERIMENTS FOR THE PURPOSE OF

DEVELOPING REGRESSION MODELS... . .. .. ........... I

2.1 Planned Testing. ...... ,,*,....................a*.*. 2

2.2 Specification Daa......... .... see..... .... *. 5

3. CORRELATIVE INFORMATION FOR USE IN EVALUATING ELEVATED

TEMPERATURE PROPERTY DATA FOR METALLIC MATERIALS ................. 12

3.1 Need for Correlative Information................. 12

3.2 Detailed Uses for Data Generated and
Correlative Information....oo....... ................ 0 13
3.3 Nature of Correlative Information to be
Provided with Property Data ............................... 14
3.4 Information Regarding the Material Product ........ 000000.0.. 14
3.5 Information Regarding the Test Specimensooo..*oooooooeo..... 18
3.6 Information Regarding the Testing.....,.................... 20

4. A COMPREHENSIVE METHOD OF CREEP AND STRESS RUPTURE

DATA ANALYSIS WITH SIMPLIFIED MS.OEL. ..... .oo........... 21

4.1 Intntrodutction
0 000.o6 00 0 0.a .*00 0 0... oo 69 00 0 0 0a0 00 00 6 .
0o 21
4.2 Objectives of Data~nlss................ 21

4.4 Regression Anlss...... .. *. ........... 26

4o5 Evaluation of Fit. ~......... *....*.... ............... *. 28

4o7 Computer Programs. ...... ............. *,.......ooo...

...... 32
4o8 An Example Aayi... ............. ......... 33

APPENDIX A............................................................. 38

APPENDIX B... o..o...... so... ................. o................... 62

V
 LIST OF ILLUSTRATIONS

FIGURE 4.1 Observed and Predicted Stress Rupture

Lives for NRIM 304 Stainless Steel...............0. 35

FIGURE 9.3.6.2 Typical Creep-Rupture Curve....................... 40

FIGURE 9.3.6.7 Average Isothermal Rupture Curves for

FIGURE 9.3.6.8(a) Estimated Stress Rupture Curves for

FIGURE 9.3.6.8(b) Experimental Design Matrix for

Creep Rupture.......................... 57

FIGURE 9.3.6.8(c) Alloy 325 (MOD) Stress Rupture

Typical and Minimm Life................. 59

FIGURE 9.3.6.8(d) Alloy 325 (MOD) Stress Rupture

LIST OF TABLES

TABLE 2.1 Stress Rupture Times to Failure (Hours)*.*****o...... ee 6

TABLE 2.2 Computations for a Components of Variance

1 )0 ......
Analysis for Stress Rupture Lifetime (Log, 0 8

TABLE 2.3 Analysis of Variance Table for Stress

Rupture Lifetime (Logj 0 )........*....**a******......-o 9

TABLE 9.3.6.8 Results of Simulated Sampling of Creep-Rupture

vi
 STATISTICAL GLOSSARY

(1) Analysis of Variance. The analysis of the total variability of a set of

data (measured by their total sum of squares) into components which can
be attributed to different sources of variation. For example, scatter
measurements can be proportioned into (a) heat-to-heat and (b) repeat-
ability, or within heat scatter.

(2) Distribution Analysis determines (a) the underlying distribution, and

(b) the parameters of that distribution.

(3) Log-Log Plots are used to graph observations drawn from a distribution
whose underlying density function is log normal, Weibull, or extreme
value. Data from these distributions will graph a straight line. An
example would be stress rupture life in hours as a function of stress.

(4) Log-Normal Distribution is a density function that is not symmetrical,

but is positively skewed. If the logarithms of the values of a random
variable have a normal distribution, the random variable itself is said
to have the log-normal distribution. Creep and stress rupture values in
hours are just a few of the phenomena that may be log-normally
distributed.

(5) Normal Distribution is a probability law called the normal density

function and can be defined mathematically with parameters T (mean) and
a (scatter measurement). The graph of this function is a symmetric
bell-shaped curve. The normal distribution forms the cornerstone of a
very large portion of statistical theory.

(6) Outlier(s) are observations at either extreme (small or large) of a

sample which are so far removed from the main body of the data that the
appropriateness of including them in the sample is questionable. There
are statistical methods to determine the probability that the extreme
value observed is an outlier.

(7) Regression Model is the mathematical model chosen to represent the

universe that applies to the distribution from which the observations
were drawn. Regression models are developed by three methods:

(1) When the model is known in advance, regression analysis

derives the coefficients for the model.

(2) Step wise by adding terms representing the independent .vari-

ables and then testing to see if they have made a significant
improvement in the model.

(3) Backward elimination, where every term thought to be signifi-

cant is put into a model and then terms are eliminated by
removing the least significant terms after each regression
analysis.

vii
 GLOSSARY (concluded)

(8) Residual is the difference between the observed value and the corres-
ponding fitted or predicted value. Residuals are highly useful for
studying whether a given regression model is appropriate for the data at
hand. They can be used to locate areas of inconsistency in model
building. The sum of the residuals is zero.

(9) Scatter Measurements are estimates made on the dispersion of observa-

tions about a mean. This measurement is called the standard deviation
and symbolized with the Greek letter a. It is often referred to as the
repeatability error.

(10) Slope Measurements are observations made about isothermal regression

lines relating life as a function of stress. These observations are
made to estimate the slope of these lines.

(11) Standard Error of Estimate (SEE) is a measure of the scatter of obser-

vations about a regression line. It is the square root of the residual
variance. A mathematical derivation is available in any text on regres-
sion analysis.

(12) Tolerance Intervals, also referred to as tolerance limits, are the

intervals or limits within which 100 (1-a) % of future observations are
expected to fall. The width of these intervals is a function of the
size of the standard deviation and the degree of uncertainty resulting
from making estimates of the mean and the standard deviation from a
finite sample.

viii
 LIST OF SYMBOLS

bj = regression coefficients

BSSD - between-heat sum of squared deviations

c - intercept term in the regression - b0

ch = intercept term in the regression for a given heat of an alloy

= arithmetic average of the Ch's

f - degrees of freedom

H - number of heats of an alloy

h = heat index

k - the number of parameters in the regression, or the magnitude of

the noncentral statistic

m - thickness

MSB = between-heat mean square

MSW = within-heat mean square

N = total number of data values

ni sample size of ith heat

nE weighted average number of observations per heat

R2 coefficient of multiple determination

Si = sample sum of ith heat

s =- standard deviation of ith heat

SEE = standard error of estimate

SSI sum of squares of ith heat

SSD = sum of squared deviations

SSE sum of squares error

T temperature, degrees Fahrenheit

TA temperature of convergence of the iso-stress lines

ix
 LIST OF SYMBOLS (concluded)

t W time, hours

TSSD M total sum of squared deviations

VB W between-heat variance

Vh W variance of heat intercept

VT W total variance

Vw = within-heat variance

Wi M weight for the ith heat used to calculate regression intercepts

when separating heats

WSSD M within-heat sun of squared deviations

x M independent variable

x = average of the independent variable

xij M jth observation in the ith heat

y M dependent variable

M average of the dependent variable

y M estimate of the dependent variable

= VB/Vw, ratio of between to within-heat variance

- stress or standard deviation

aT = total standard deviation

0* - estimate of standard deviation for a multiple heat data

collection

*2 within-heat component of variance

2 between-heat component of variance

x
 1. INTRODUCTION

In 1976, a subcommittee from the MIL-HDBK-5 Coordination Committee

was established to investigate the state of the art in creep and stress rup-
ture data analysis procedures. The goal of this Elevated Temperature Task
Group (ETTG) was to develop a new and more comprehensive guideline on creep
and stress rupture data analysis for inclusion in Chapter 9 of MIL-HDBK-5.
After undertaking this task, it eventually became apparent that any
guideline resulting from this activity could not be totally comprehensive,
since the limited length of MIL-HDBK-5 guidelines precluded a comprehensive
review of the state of the art that would completely support recommended anal-
ysis procedures. In view of this predicament, a compromise approach was
taken. A relatively brief guideline which delineated the appropriate methods
for creep and stress rupture analysis for MIL-HDBK-5 was prepared. This
guideline, as approved at the 58th MIL-HDBK-5 Coordination Meeting is included
in Appendix A. The guideline covers all major considerations, but it does not
go into any great detail on the justification for such an approach, or on
particular difficulties that one might encounter in special cases which would
complicate the analysis.
The purpose of this reference document is to provide supplementary
background information on creep and stress rupture ankalysis which will be use-
ful in performing a data analysis according to the MIL-HDBK-5 guidelines. The
document is subdivided into three sections: (1) design of experiments for the
purpose of developing regression models, (2) required correlative information
for use in evaluating elevated temperature property data, and (3) a comprehen-
sive method of rupture data analysis with simplified models.

2. DESIGN OF EXPERIMENTS FOR THE PURPOSE OF

DEVELOPING REGRESSION MODELS

There are hundreds, perhaps thousands, of curves published and used

in industry today for the purpose of graphically displaying the mechanical
properties of metals and their alloys. The vast majority of the curves that
manufacturers use in designing everything from metal fasteners to aircraft

1
turbines are developed from data not specifically generated to produce mathe-
matical models. The data, if excessive, raise costs; and if not sufficient,
decrease reliability. Often creep and stress rupture data are both excessive
and inefficient because the observations, although more than required, do not
adequately cover the temperature-life matrix. In practice, mathematical
models are derived and curves are drawn from those models. Product reliabil-
ity depends on the accuracy of these curves. To develop precise and accurate
models, a system of test planning for regression analysis is required. It is
the purpose herein to present one simplified method of experimental design
that will enable engineers to produce a more precise model at less cosL. The
production of the models is not covered in this section since it is discussed
in detail in Section 4.

2.1 Planned Testing

Planned testing using experimental design techniques for selection of

experiments is a logical approach to understanding the properties of an alloy,
yet it is almost never done. The designer's success is directly related to
the efficiency of the design curves and mathematical models developed from
test data.
One method for planning testing is to develop a test layout in matrix
form with the test temperatures listed in rows and lifetimes of interest tabu-
lated in columns. Then, through testing, the blocks iormed in the row-by-
column matrix can be filled out. This approach ensures coverage of all the
areas of interest.
Although there are other methods, this method is probably the sim-
plest, oldest, and most generally used one; the most efficient and sophisti-
cated method is the "Central Composite Design". Other methods are discussed
in Reference 2-1.
Using the matrix format shown in Figure 2.1, this method involves the
steps listed as follows:
1. Select and bracket the range of temperatures and time variables
desired. Insert the temperatures selected in the extreme right-
hand column.

2
 2. From estimated log-log or Larson-Miller typical (average) life
plots select and record the stress expected to produce the column
life at the row temperature.
3. If no stress-temperature or time related interactions are
expected, some of the experiments can be omitted, so long as no
less than 20 experiments remain.
4. Omitted blocks must be selected randomly as follows:
a. Omit blocks of experiments in sets equal to the number of
temperatures In the model.
b. In each set, omit one observation from each temperature level
in the rows. (See the example given in Section 2.3.)
c. Do not omit more than one value in any life (hours) columns.
5. For small designs do not omit any of the corner experiments (in
the matrix).
6. If interactions are suspected or found, the entire matrix must be
completed (no omitted blocks are allowed).
7. Perform the experiments in random order, mixing temperatures,
machines, operators, heats, etc., to ensure that unaccounted for
(nuisance) variables are randomized.
8. Small differences in the experimental life obtained and the
estimated life are to be expected. Reset the stress levels and
readjust the matrix if an experimental life is off by more than
two columns from the estimate. All experimental results may be
used in the final regression.
Before the test matrix, as shown in Figure 2.1, can be formed, the
interval sizes must be selected, first for test temperatures, and then for
desired lifetimes.
(a) Temperature - A range of temperatures is usually required. For
example, if the test range is from 1000 F through 1500 F, the
basic question is: should tests be performed at six levels
(1000 F, 1100 F, 1200 F, 1300 F, 1400 F, 1500 F) or at three
levels (1000 F, 1300 F, 1500 F)? The decision can be quite
complicated and based on such considerations as:

3
 (1) The expected spacing of the isothermal lines
(2) The likelihood of parallel or divergent isothermal lines
(3) Anticipated precipitation of secondary phases during the
life ranges of interest.

HOURS
3 6 K)15 UB356 100116013201560 1000 F

TI
T2
T3
T4
T5
w T6
T7
TS

FIGURE 2.1. MATRIX FORMAT

If reasonable estimates of anticipated isothermal lines can be

constructed, this selection can be greatly simplified with very
little risk. Starting with the lowest temperature, the next
temperature line should be chosen such that at least one level
if testing stress, on the log-log stress-life plot, will be
common to both temperatures. This process should be repeated
for each temperature line, ensuring like stress values for
adjacent temperature levels.

4
 (b) Life - A log life cycle should normally be divided into four
equal intervals. For example, between 100 hours and 1000 hours,
the divisions would be approximately 180, 320, and 560 hours on
the log scale.
These divisions are far enough apart to insure a well defined
curve and a minimum overlap of data. To convert from tempera-
ture and life desired to temperature and test stress requires
some prior knowledge of the time-temperature-life relation-
ship. If there is no prior knowledge, a series of "probe" tests
must be made to locate the isothermal lines on a log-log plot.
An example of an experimental design for the purpose of develop-
ing regression models is given in Section 9.3.6.8 of Appendix A.

2.2 Specification Data

Virtually all alloys are controlled and purchased to a material spec-

ification which provides a process control variable generally called the "spec
point". Therefore, there will often be large quantities of data available
from quality control data records at the "spec" condition. The data will con-
tain many heats and will provide an excellent indication in regression equa-
tions of scatter. Therefore, in regression modeling, "spec" data are often
the major source of the scatter or variability measurements. The slope
measurements must come from the experimental design matrix.
"Spec" data can also be used to: (1) determine through analysis-of-
variance techniques the fractions of the scatter due to heat-to-heat varia-
tions, etc., and (2) determine through distribution analysis if the data are
normal, log normal, etc., and to determine, if the distribution is not normal,
what is required to "normalize" it.
When no "spec" data are available, the scatter measurements are de-
termined from the residuals of the regression model. However, they are mixed
(confounded) with some small curve fitting error. Curve fitting usually in-
creases scatter in stress rupture regressions by 5-15 percent. However, curve
fitting errors can be much larger in some cases.

5
 I °

2.3 Heat-to-Heat Variation

A batch of an alloy is generally referred to as a heat. Batch varia-

tions in chemistry, heat treating, etc., can cause considerable variations in
the mechanical properties of the alloy. This difference is referred to as the
heat-to-heat component of variance as opposed to the within-heat component of
variance. Heat-to-heat standard deviation is often 30-70 percent of the with-
in-heat standard deviation although in some cases it may be much larger. The
root sum 3quare of the two components of variance produces a measure of scat-
ter about the regression, that when added to the curve fitting error, gives
the regression parameter called SEE (Standard Error of Estimate). It is this
parameter which is used to fix the design minimums about the regression esti-
mates (typical, or mean values). The SEE is rarely determined as defined
above, rather it is a product of regression analysis.
Two methods are generally used to obtain the major components of var-
iance, between (heat-to-heat) and within-heat. One method is described in
Section 4.6 on Multiple Heat Data. A second method, based on Analysis-of-
Variance (ANOVA) techniques, is much simpler and is described by an illustra-
tion. This method requires repeat observations, from several heats, at a com-
mon stress-temperature level (such as at the specification point).
Table 2.1 shows 19 times to stress rupture, in hours, for specifica-
tion data associated with four heats, BJJK, BJJJ, BKLJ, and BLLD. Since a
log-normal distribution of rupture lives will be assumed, base 10 logarithmic
transformations must be made of these lifetimes before making subsequent
analyses.

TABLE 2.1. STRESS RUPTURE TIMES TO FAILURE (HOURS)

Heat Label
BJJK aiJJ SKLI BLLD

35.0 51.3 29.0 41.4

33.1 37.5 36.1 16.5
33.4 48.6 47.5 33.6
42.7 74.2 32.6
70.5 27.4
26.4
34.9

6
 Table 2.2 shows standard computations (see Reference 2-2) that par-
tition the total sum of squared deviations TSSD into two components: the
between-heat component, BSSD, and the within-heat component, WSSD. These com-
ponents are defined in algebraic terms, together with their associated compu-
tational formulas, as follows:

2
h 1( 2 h S /n h 12/h

BSSD I
h 2 Ih SI I I ni
i- i( .I (Jl=

h ni 2 h
WSSD I I (Xij - i - WSSDJ

im imll i )/i-
h ni 2 hhh
and TSSD I j-_ S, IS In

From the algebraic expressions, it is seen that the TSSD consists of the sum
of the squared differences between each observation xij (the jth observation
in the ith heat) and the mean taken over all observations, T.
In contrast, the WSSD is seen to consist of the sum of the squared
differences between each observation, within a given heat, xij, and the mean
for that heat, xi" These within-heat sums of squared deviations, WSSD i are
then summed over the h heats to obtain the final WSSD. The between-heat term,
BSSD, is obtained by summing the squared differences between the mean of each
heat xi and the overall mean x. These differences are weighted in accord
with the number of observations in each heat ni. An algebraic expansion of
these expressions shows that

TSSD - BSSD + WSSD

The right-hand side of the defining expressions gives the conventional compu-
tational formulas, where the symbol Si denotes the sum of the observations in
the ith heat, and SSi denotes the sum of the squares of the observations in

7
 0
0N 0%
a- ~ 4 4-

r- Les

1-4s ~ 0%-
"4 a, I- 1. *0
z W 0E 0

Z- C en 0 I.: IA

z0 Ei r -O

0
enr ,
s

s 'C 00 C0

eq I..~ VIN ..

w 0 1

C.N .

S C I .~ p8
the ith heat. The computational formulas are used to reduce round-off errors
in hand calculations.
Based on the totals shown in Table 2.2, it is seen that

BSSD - 47.152 - (29.880)2/19 - 0.162

WSSD - 0.244

and

TSSD - 47.397 - (29.880)2/19 - 0.407.

Table 2.3 shows an analysis of variance table that summarizes these

results. The degrees of freedom f for the between-heat SSD, the within-heat
h h
SSD, and total SSD are given by h-1, ni -h and I ni -1, respectively;
i-i i-l
and with h - 4 these expressions yield 3, 15, and 18 as shown in Table 2.3.
The corresponding mean squares are given by SSD/f and are also shown in the
table.

TABLE 2.3. ANALYSIS OF VARIANCE TABLE FOR

STRESS RUPTURE LIFETIME (LOGIo)

Sum of Degrees
Squared of Mean
Deviations Freedom Square
Source of Variance SSD f SSD/f

Differences Between Overall Average Lifetime

and Average Lifetime for Each Heat 0.162 3 0.0540

Differences Between Average Lifetime for Each

Heat and the Individual Lifetimes Within
Each Heat 0.244 15 0.0163

Total 0.406 18 -

9
 The components of variance are obtained from the mean squares as fol-
lows (see Reference 2.3). Let MSB and MSW denote the mean squares for the be-
2
tween-heat variation and the within-heat variation. Next, let W2 and 0
denote the components of variance for the between-heat and within-heat
variations. The expected value of MSW is given by a 2 , so that an estimate of
02 is equal t- 0.0163 for this example. The expected value of MSB is given by
2
2
2 n , where i is an "average" number of observations per heat and is
computed using

n 1[n, - ( ni /ni)]/(h-1)

which for this example becomes

n - (19-(99/19))/3 - 4.60

An estimate of w 2 is then obtained from the relation: 0.0163 + 4.60w 2 -

0.0540 and is found to be given by w 2 _ 0.0082. Taking square roots then

shows that the standard deviations of the between-heat and within-heat compo-
nents of variance are estimated by w - 0.091 and a - 0.128, approximately.
Because a single measurement of lifetime is likely to be affected by
heat-to-heat variation, and by the variations within a heat, an estimate of
the standard deviation for a single subsequent measurement, a , can be
obtained using the expression:

a* VC2 + a 2 - V0.0082 + 0.0163 = 0.157.

The ratio of the heat-to-heat component of variance to the within-

heat component of variance is given by w 2 /a2 - 0.0082/0.0163 - 0.50. Thus,
for this example, the heat-to-heat variation is approximately equal to 50
percent of the within-heat variation. As a rule of thumb, the following
requirements for the number of heats have been established for creep and
stress rupture data analysis in MIL-HDBK-5:

10
 (1) When the heat-to-heat component of variance is less than 25 per-
cent of the within-heat variance, use at least two heats equally
for the sample sources.
(2) When the heat-to-heat component of variance is between 25-65
percent of the within-heat variance, use at least three heats
equally.
(3) When the heat-to-heat component of variance is greater than 65
percent of the within-heat variance, use at least five heats
equally.
For the sample set of data examined above in Tables 2.2 and 2.3, data on at
least three heats of material would be recommended by MIL-HDBK-5 for analysis
purposes.
When regression models are developed from data that were not taken
from an experimental design, the heats are rarely chosen randomly. Therefore,
unless there are large quantities of data in all areas of the regression
matrix, this imbalance of heat sample sizes must be accounted for. The manner
in which this is done will not be covered here; one method is described in
Reference 2.4 and is also briefly reviewed in Appendix B.

2.4 Summary:

To design the experiments necessary to produce reliable stress rup-

ture or creep curves and to develop realistic mathematical regression equa-
tions from those experiments, the following factors must be established:
1) Temperature: Choose the specific isothermal levels from the tem-
perature range of interest.
2) Number of Observations: Determine how many observations are to
be made, and at what stress levels, for each isothermal tempera-
ture level.
3) Number of heats: Determine how many heats should be selected and
randomize the heats equally throughout the test matrix.

11
 3. CORRELATIVE INFORMATION FOR USE IN EVALUATING ELEVATED TEMPERATURE

PROPERTY DATA FOR METALLIC MATERIALS

3.1 Need for Correlative Information

The properties of metallic materials operating at elevated tempera-

tures are influenced strongly by processing variables. Every phase of pro-
cessing, from solidification of the alloy from the molten state to the final
heat treatment, affects the magnitude of individual property values and the
statistical variation of those values representing both individual pieces and
tonnage quantities of those pieces. Paralleling this fact is the complexity
of the interplay at elevated temperatures among purely metallurgical phenom-
ena; the mechanical aspects of stress, short-time strain, long-time strain;
and fracture. The properties of metallic materials at room temperature vary
much less as a function of processing variables.
Data for elevated temperature properties of alloys frequently are not
generated for the purpose of providing a basis for allowables to be used in
design. Variables frequently are not selected and controlled to provide opti-
mum data, but are those encountered in evaluating materials for other pur-
poses. Thus, it is necessary to evaluate not only the property data, but
correlative information as well, to assess the suitability and reliability of
those data for use in establishing design allowables for general use.
Frequently when data are generated for other than pure "data genera-
tion" purposes there is a significant amount of correlative information avail-
able, but it is not recorded or is recorded incompletely. In some instances,
of course, information is simply lacking due to the narrow objectives of the
specific testing concerned. The originator of the test can be assisted by
having available a list of the various identifying information which are (a)
necessary and (b) desirable for potentially extending the usefulness of the
data generated beyond its original intent. The intent of the sections which
follow is to establish a list of correlative identifying information for use
by anyone generating elevated temperature property information for metallic
materials.

12
3.2 Detailed Uses for Data Generated and Correlative Information

Elevated temperature property data are used for many individual

purposes, some of which are broader and of greater overall significance than
others. Nevertheless, all elevated temperature data must be supported by a
certain amount of correlative information. The amount of required correlative
information is a function of how the data are used.
Examples of different data uses are:
a. Accumulation, analysis, and presentation of data for MIL-HDBK-5.
b. Evaluation of new materials.
c. Evaluation of field-retired parts.
d. Evaluation and investigation of failures of parts.
MIL-HDBK-5 data analyses normally concern materials which have been
in production in relatively large quantities over a period of some years,
which have been processed by different methods suiting end requirements, which
are relatively well established, and for which there exists a substantial body
of information. Due to the far-reaching significance and use of property
values in MIL-HDBK-5, considerable care must be used to assure that the data
being evaluated are indeed representative of production material, and a great
part of that assurance comes from correlative information relating to those
data.
For new materials, a relatively modest amount of experience will
have been accumulated since smaller quantities will have been produced and
processing methods will probably still be in a state of development. In this
case, documentation of the processing is vital not only for immediate pur-
poses, but to permit future consideration of the data (perhaps for inclusion
in MIL-HDBK-5) and to determine whether those data are representative of
material currently being produced.
For field retired parts and failed parts, less information is
normally available. Failure analysis at times can be quite straightforward
[defect not detected by nondestructive testing (NDT)], but at other times
(such as abnormally low short-cycle fatigue properties) it can involve
investigation to obtain complete processing information.

13
3.3. Nature of Correlative Information to be Provided with Property Data

Collection of correlative information: The first step is to collect

all information immediately at hand without additional effort or cost,
regardless of the kind of information. Reports or documentation from the
producer, NDT inspection reports, analytical data from chemical laboratories,
and specification data are examples of information which may already be at
hand.
Categories of information to be supplied: The second step consists
of determining the identity of information falling into the following
categories:
(1) Always to be supplied.
(2) Supplied if obtaining it is feasible and not excessively
costly or time consuming.
(3) Supply only if already on hand.

3.4 Information Regarding the Material Product

General identifying information regarding the material product should

include:
Identity of producer of
(2) a. Intermediate material used in producing final product.
(1) b. Final product tested.
Information relating to all products
(1) a. Identity of metal or alloy
(1) b. Procurement specification to which it was produced
(3) c. Purchase order number to which the final product was
procured
d. Usual pedigree information:
(1) Chemical analysis determined at the producing mill
(2) Check analysis
(1) Heat number, If applicable
(3) Ingot number, if applicable
(2) Mill test report (frequently contains supporting
supplementary-data)

14
(2) Tensile property and fracture toughness data generated at
mill or elsewhere to assure that product conformed to
specification requirements
(3) Hardness
(3) Crack propagation data (da/dN) if available
e. Quality aspects
(3) Cleanliness rating data (AMS 2300, 2301)
(3) Macroetch rating
(2) Ultrasonic class to which product conforms if required
(AA, A, B)
(2) Radiographic rating, as applicable
(2) Eddy current rating, as applicable
f. Heat treatment
(1) Conducted by whom (producing mill, forge shop, user, etc.)
(1) Final heat treatment conducted on as-produced stock,
rough-machined stock, or finish-machined part
(2) In air, vacuum, inert gas, cracked gas, etc.
(2) Statement of time, temperature, quenching, tempering,
aging, etc., if not adequately covered elsewhere
(1) Sequence and nature of mechanical working if such is
associated with achieving desired heat treat condition
Information relating to specific products
(1) Results of any microstructural studies made on the
specific product tested, or on specimens from it
Aspects of melting and casting process used in producing
ingot from which final product is made (usually applies
only to iron-and titanium-base alloys)
(1) a. Melting process
(2) b. Number of remelt cycles, as applicable, especially if number
is abnormal
(3) c. Special aspects of melting and casting, if applicable (all
high purity melting stock; inert gas used over ESR slag
layer, etc.)
(3) d. Ingot size, especially if abnormal

15
(3) e. Conditioning of ingot surface, especially if abnormal
Forgings
(2) a. Hammer, hydraulic press, HERF machine, ring
(1) b. Forging practice: Closed die, pot forging, hand (Smith)
forging, ring rolling, loose mandrel ring forging
(1) c. Forging stock used: Forging billet, forging ingot, powder
metallurgy preform
(3) d. Size of forging stock used
e. Out-of-the-ordinary aspects, as applicable:
(1) Low forging temperatures in HERF product (initial and
final)
(2) Creep forging in superalloy dies
(2) Ausforming time-temperature sequence
(3) Time-temperature-percent reduction of powder metallurgy
preforms
Extrusions
(3) a. Size of press used
(1) b. Extrusion stock used: Billet, ingot, powder metallurgy
preform
(3) c. Size of extrusion stock used
(2) d. Out-of-the-ordinary aspects, e.g., low or high reduction
ratios
Impact extrusion (impact forging, cold forging)
(3) a. Kind and size of press used (mechanical, hydraulic)
(2) b. Type of extrusion: forward, backward, forward and backward
(3) c. Shape and design of preform
(1) d. Stock used in fabricating preform
(3) e. Temperature of preform
Castings
(1) a. Process: green sand, baked sand, shell mold, plaster,
investment, etc.
(3) b. Type of cores used
(3) c. When available, location of gates and risers (as they can
affect properties in their vicinity)

16
(1) d. Whether casting was repair welded before final heat
treatment (very common), and if so where and how
(1) b. If specimen is from a casting, does it have as-cast surfaces
or was it machined all over
(1) c. If specimen represents a cast product, was the specimen cast
separately
(1) d. Does the specimen retain some surfaces from the original
product, e.g., as forged and as-heat treated surface of a
die forging
(3) e. Sequence followed in rough machining, finish machining, and
heat treatment
(1) Sequence and nature of mechanical working if such is
associated with achieving desired heat treat condition
Information relating to specific products
(1) Results of any microstructural studies made on the specific
product tested, or on specimens from it
Aspects of melting and casting process used in producing
ingot from which final product is made (usually applies
only to iron-and titanium-base alloys)
(1) a. Melting process
(2) b. Number of remelt cycles, as applicable, especially if number
is abnormal
(3) c. Special aspects of melting and casting, if applicable (all
high purity melting stock; inert gas used over ESR slag
layer, etc.)
(3) d. Ingot size. especially if abnormal
(3) e. Conditioning of ingot surface, especially if abnormal
(1) f. Heat treatment practice followed, if such was conducted on
the machining blank or on the finished specimen, thickness
of blank heat treated
(3) g. Heat treat oxide left on or removed; if removed, how was it
accomplished
(1) h. Identity of any surface-finishing processes used, and
surfaces on specimen so affected: Dissolving

17
 surface-damaged layer in beryllium; peening (specification
used and practice followed)
(2) i. NDT of finished specimen, including radiographic quality of
welds
Specimen proper
(1) a. Basic design and dimensional tolerances
(1) b. Location of specimen in product, including orientation with
respect to grain flow and with respect to weld direction
(1) c. Notch orientation with respect to L, LT, and ST directions
(2) d. Precracking practice
(1) e. Sequence of precracking and heat treatment.
Powder metallurgy end product
(2) a. Powder production method
(3) b. Powder size and, if applicable, shape characteristics
(3) c. Pressing method (axial ram-type, isostatic, hot or cold,
pressure, time)
(3) d. Sintering temperature, time, and atmosphere
(3) e. Special aspects, e.g., use of activators with powder
(3) f. Pressing and sintering practice: pressures, atmospheres,
density as pressed, sintering time-temperature-atmosphere,
density after sintering
Weldment
(1) a. Process: TIG, MIG, EB, coated-electrode, etc.
(2) b. Weld face preparation
(1) c. Filler metal used: kind, size, procurement spec.
(2) d. Practice: number of passes, preheat, interpass temperature,
postheat, amperage, voltage, type current, limitation on
joules per inch, etc.
(1) e. Heat treat condition: as welded, reheat treated, etc.

3.5 Information Regarding the Test Specimens

General identifying information regarding the test specimens should

include:

18
 Identity of stock from which specimens were machined
(1) a. Virgin product, e.g., plate as produced by the mill
(2) b. Product subject to prior testing, e.g., tests on a
field-retired part, or on a statically tested forging
Fabrication
(1) a. If applicable, identity of any nontraditional (non-
mechanical) machining process used, such as EDH
(electrical discharge machining) or ECG (electrochemical
grinding). Some of these processes, such as EDM, may
alter test surfaces significantly
(1) b. If specimen is from a casting, does it have as-cast surfaces
or was it machined all over
(1) c. If specimen represents a cast product, was the specimen cast
separately
(1) d. Does the specimen retain some surfaces from the original
product, e.g., as forged and as-heat treated surface of a
die forging
(3) e. Sequence followed in rough machining, finish machining, and
heat treatment
(1) f. Heat treatment practice followed, if such was conducted on
the machining blank or on the finished specimen, thickness
of blank heat treated
(3) g. Heat treat oxide left on or removed; if removed, how was it
accomplished
(1) h. Identity of any surface-finishing processes used, and
surfaces on specimen so affected: Dissolving surface-
damaged layer in beryllium; peening (specification used
and practice followed)
(2) i. NDT of finished specimen, including radiographic quality of
welds
Specimen proper
(1) a. Basic design and dimensional tolerances
(1) b. Location of specimen in product, including orientation with
respect to grain flow and with respect to weld direction

19
 (1) c. Notch orientation with respect to L, LT, and ST directions
(2) d. Precracking practice
(1) e. Sequence of precracking and heat treatment.

3.6 Information Regarding the Testing

(1) Basic type of test

(2) If available, specification for test procedures followed,
e.g., ASTH E139-70 for creep-rupture testing
(1) Details of testing if not covered completely by a
specification
(2) Identity and capacity of equipment used for loading specimen
(3) Identity of strain-measuring instrumentation
(3) Identity of recording instrumentation
(1) Environment of specimen during testing: air, percent
relative humidity, dew point if low humidity, salt fog,
inert gas, torr vacuum, etc.
(2) If fracture faces evaluated, method used.
(1) Testing conducted by:
a. Name of company, laboratory, or other applicable
organization
b. Name of individual personally conducting the test
(1) Correlation and analysis of data conducted by: (specify)
(3) Authorization and funding by: (specify)

20
 4. A COMPREHENSIVE METHOD OF CREEP AND

STRESS RUPTURE DATA ANALYSIS WITH SIMPLIFIED MODELS

4.1 Introduction

Until the dream of a universal creep and stress rupture equation is

achieved, (that includes among its variations all the parametric and other re-
lationships found useful in the past), it is necessary to use less general
equations in curve fitting. The method described herein uses standard param-
etric equations as starting points and adjusts the best of these to correct
for inadequate fit to the data. Depending on the type of lack-of-fit
observed, the adjustments are made by the analyst, using metallurgical infor-
mation where applicable.
The proposed method requires the analyst to consider the material as
an inherent part of the analysis and to apply his mathematical skills where
needed to closely approximate in a rational way the behavior interpreted from
the data or learned from other sources. On the other hand, the method has
been designed to provide the analyst with mathematical tools for obtaining
good results from typical engineering data, permitting more efficient design.
Access to a computer is required and a degree of automation is
strongly recommended.

4.2 Objectives of Data Analysis

The first objective of a creep or stress rupture analysis is to find

the underlying relationship between stress, temperature, and life, modified if
necessary by inclusion of the effects of such auxiliary variables as specimen
geometry, grain size, coating, etc. In a creep or stress rupture analysis,
the logical dependent variable is the logarithm (or log) of life; this vari-
able is approximately normally distributed with uniform variance and has the

21
highest variability.* Least squares regression minimizes the sum of the
squares of the differences between observed and predicted logarithms of
life. The analysis must minimize fitting error to ensure that predicted lives
lie in the central part of the envelope of the observed data over the range of
test conditions.
Once the equation is obtained, it may be used to predict the stress
to obtain a given life at a given temperature, or to predict the temperature
to obtain a given life at a given stress. Either prediction can include
appropriate statistical limits.
A second objective is to present the results in a form having high
utility to the users-especially those in design engineering. In addition to
graphs of the relationship, the results may be presented in the form of equa-
tions allowing predictions of probable creep or stress rupture life at any
condition of design interest.
Any limitations on the range of applicability of published equations
can be indicated on associated graphs by terminating the curves at the limit
of applicability. Equations made available by publication or storage in com-
puter programs require a definition of circumscribing limits, for example,
stress and temperature limits on reliable extrapolation.
A third objective is to use the results of the analysis to improve
the data mix to be obtained in the same or subsequent experiments. Principal
interest lies in identifying test conditions which will optimally determine
curve shape. Test conditions most frequently lacking in a creep or stress
rupture analysis are those resulting in long lives. Other test conditions can
frequently be found where cost-effective contributions to curve shape deter-
mination can be made.

*The use of the logarithm of life as the dependent variable is supported by

considering data at one stress and temperature, such as quality control data
at a specification point. Life is the only variable. Statistical results
from such data can readily be used in conjunction with regression analysis
using the logarithm of life as the dependent variable for data at other test
conditions.

22
 4.3 Guiding Principles

In general, an attempt should be made to use all available data in

the analysis, including those from different heats, having different grain
size, section size, etc. However, all the data should not normally be made to
fit a single curve. Effectively, a separate time-stress-temperature curve for
each significantly represented heat and each level of each auxiliary variable
should be obtained through the use of additional regression variables (terms
in the equation). Regression variables for such physical characteristics as
grain size and section size should be added to the equation when they are
known quantitatively and when preliminary plots show they have significant
effects on creep or stress rupture life. If the effect is not the same at all
stresses and temperatures, cross-product terms of these regression variables
with stress or temperature can also be added. The procedure that should be
used for handling ultiple heats is described in Section 4.6. Certain vari-
ables such as heat treatment or large changes in chemistry may affect curve
shape so drastically that it will not be realistic to analyze data with dif-
ferent values of these variables in a single analysis, even when they are in
the range allowed by the controlling specification.
Initial screening analyses should be made using standard parametric
equations. These equations recommended in the MIL-HDBK-5 guidelines are known
by their originator's name: the Larson-Miller, Dorn, Hanson-Succop, and
Manson-Haferd equations. An unnamed equation with linear terms in stress,
reciprocal stress, and reciprocal temperature, and cross-product terms in
stress with temperature and reciprocal temperature has also been found to be
useful in some cases. The parametric equations which are recommended are
cubic functions in the logarithm of stress or, in the case of Larson-Miller, a
cubic function in stress.
When one of these equations fits the data adequately by the criteria
described in Section 4.5, the analysis can be considered complete. For some
alloys, there are regions of stress and temperature where none of the para-
metric equations will be adequate. In these instances, the most promising
equation should be modified to obtain good fit.

23
 A few words are in order about the recommendation to use a cubic log
stress function in the screening equations. A cubic equation implies a sig-
moidal shape on a logarithmic plot, and is usually concave downward at high
stress and concave upward at low stress (the coefficient of the cubic term is
negative). The high-stress curvature is needed in order for the equation to
approach the ultimate tensile strength at reasonably short life and is
generally accepted as valid. The curvature at the low stress end is ques-
tioned; some highly qualified observers favor an asymptotically linear curve
in this region or even one with concave-downward curvature. With typical air-
craft engine alloys the cubic representation is generally successful.
The inflection point of cubic analyses, where the curve is nearly
straight, occurs near the low-stress limit of typical sets of data. Such data
will not provide sufficient information to allow selection between the cubic
and other curve shapes essentially linear in the low-stress region. This lack
of definition is especially severe if long-time data are excluded from the
analysis in order to test extrapolation capability. Relatively few data sets
clearly show the existence of an inflection point; most show only a tendency
toward linearity in this region. The use of the cubic may provide unconserva-
tive predictions with extrapolations greater than a decade toward longer creep
or stress rupture lives if the linear or concave-downward assumption is really
correct and data are simply not available to demonstrate this trend. The
cubic is normally acceptable for moderate extrapolation, however.
One method that can be used to improve the fit is to use a double
cubic, or cubic spline fit, in log stress to allow different curvatures at
high and low stresses. The selection of the location of the knot (or value of
log stress where the two cubic equations join) is arbitrary, but not generally
cricical. Frequently the log stress value from the mid-range of the data is
effective. When a cubic spline fit is used, both cubic coefficients for the
equation of the line should be negative.
Analysis of data covering a limited range of stress will occasionally
determine a positive coefficient of the cubic term in the parametric equa-
tions. This effect can be caused by unusual experimental error in a few data
points strategically located to affect curve shape, and is not considered to
represent a true relationship between stress and life. The following steps

24
should force the coefficient of the cubic term to be negative. First, elimin-
ate the second order term in log stress. Second, by inspection of a log-log
plot of the data, select a reasonable stress for the inflection of the
curves. Divide all stresses by the value of the inflection stress before
taking logarithms. This will force the inflection to the selected stress and
allow the data to determine the curve shape under this restraint.
If insufficient high-stress data exist to determine curve shape,
ultimate tensile strength data may be entered as rupture data at a short time,
such as 0.01 hour, to control extrapolation in short times. Since the curve
is extremely flat in the region of ultimate strength, the time assigned to
these entries is not critical.
Runouts, or tests not run to completion for any reason, can be used
in two ways. The first is to treat the data as censored and use a computer
program which calculates the maximum likelihood relationship between the vari-
ables. This method is rather cumbersome, however, and tends to underestimate
the standard error. Therefore it should generally be avoided. The method
found more suitable, though less rigorous, is to include in a second regres-
sion only those runout data that lie above the curve calculated for completed
tests (valid failures). It is reasonable to asswne that runouts below the
curve for valid failures contribute little to the determination of curve
shape, while the inclusion as failures of those above the curve will tend to
raise the curve where they occur. While not proven, this approach is consid-
ered to give conservative results on the average.
The use of higher order polynomials should be avoided because of
their typically very poor extrapolative characteristics. When interaction
terms between auxiliary variables are needed, these should generally be formed
by multiplying the auxiliary variable by one of the stress or temperature
terms already in the equation, thereby keeping the curve shape as simple as
possible.
The usual temperature term is either linear, or the reciprocal of
absolute temperature. If it is necessary to improve the fit with temperature,
other powers than I or -1 can be used to give a simple curve shape with lower
error. If the optimum exponent appears to be close to zero, the logarithm of

25
temperature is suggested since very small exponents cause roundoff errors in
computer calculations.

4.4 Regression Analysis

Regression analysis can be performed using centered data to avoid

roundoff error. To do this, the equation fit by the computer should be of the
form

k
y -y i ij-X )
b(X

for k dependent variables, where

y and X are the averages of these variables,

j is the index of a particular life datum and
yj is the corresponding predicted value of log life.

For example, if y - log time, T - temperature, and X = log stress,

the cubic Larson-Miller equation for a single heat can be written

2 3
S 2 3
) +
y -y a b (T-L-)+b (T- ) +b (X -!E b (- 2E) (4.1)
1 T T 2 T T 3 T -T)+ 4 -- ) (4.1

or

y -y - bIX 1 + b2X 2 + b3 X3 + b4 X4

where X - log stress and the subscripted X's have an average value of zero.
The bis are the regression coefficients. The sum of squares of the deviations
(or sum of squares error) of observed y from predicted y is

SSE - (y -2- 1
bIX 1 - b2X2 - b3X3 - b4 X4 ) (4.2)

The solution for the values of bi can be obtained by setting the de-
rivative of the last equation with respect to each bi equal to zero, and

26
obtaining a set of equations called the normal equations. For the example
given in Equation 4.1 these are:

2
SIX (y-y)J - b Z X 1 + b2 Z X1X 2 + b3 X1X 3 + b4 XX

Zx2 yY b1 Z2X1+b2X 2 2+b 3 ZX2X3 + E X2 X4 (4.3)

S[X 3(Y-)] b I X3X + b2 E X3X2 + b3 E X 2 + b 4 L X3X 4

Z IX4 (y-y )] b1I X4 X1 + b2 X 2 + b3 ' X4 X3 + b 4L X42

The summations on the right side of these equations form the X matrix
which can be inverted to provide a solution for the bi's; the solution mini-
mizes the SSE, independent of the form of the distribution of error in y.
A more accurate solution can be obtained if, before inversion, the X
matrix is first converted to the correlation matrix. In the correlation
matrix all entries on the principal diagonal are 1, and the off-diagonal
entries are between -1 and +1. the Inverse of the correlation matrix can be
converted to the inverse of the X matrix without significant loss of accuracy.
With this approach and occasional rewriting of variables to reduce
correlation among the "independent" variables, it is normally not necessary to
use double precision or complete orthogonalization of the independent
variables to obtain good results.
It is generally not possible to identify suspect data until at least
a preliminary curve is obtained, either graphically or by regression. Any
data unusually distant from this curve, called outliers, are suspect. If re-
view of the original test record indicates any serious question as to their
validity, the data should be removed from further analysis.
Data may also appear to be outliers because of the selection of the
wrong model for the analysis. However, if unusual deviations occur for
several customary curve fits, the data may be deleted even if no experimental
cause is found. This is a judgment decision. Although there is a certain
probability of valid outliers, inclusion of outliers can cause the prediction

27
for the material to be in error. The reasons for any deletions should be made
a part of the record of the analysis.

4.5 Evaluation of Fit

Frequently, the size of the standard error of the regression is ade-

quate to determine the suitability of a regression model. However, the equa-
tion with the smallest standard error may fit the data poorly in one important
region and compensates for this by unusually good fit elsewhere. A detailed
study of the residuals (observed minus predicted log life) will be helpful to
evaluate whether an adequate fit has been obtained.
A simple plot against the predicted lives is one method of evaluating
residuals. Any trends upward or downward in the residuals, particularly at
long lives, indicate an equation with poor extrapolation potential. This is
the most frequent reason for choosing an equation with a slightly higher stan-
dard error over that with the lowest.
A table of the average deviation in each of several ranges of predic-
ted life is another tool for evaluating residuals. Five cells are suggested
at each temperature, each cell covering a fifth of the total range in predic-
ted log life. The observed pattern of plus and minus deviations can suggest
changes to the equation which will increase randomness. A computer program
can easily be modified to print such a table as part of the analysis output;
this can improve efficiency by avoiding plot routines for every trial.
An extension of a plot of the predictions, particularly to longer
lives than those measured, will allow engineering judgment of the suitability
of the equation for extrapolation. Typical problems that can be uncovered by
this technique are crossing of isotherms and radical changes in curve slope
outside the data range.
A good standard error is 0.10 in common logarithms of life; typical
values usually range from 0.15 to 0.30. Occasionally standard errors as low
as 0.05 can be obtained, particularly with small amounts of data and when com-
plex equations are used. When this happens, it is likely that some of the
error of the measurements is being fit by the equation. Extrapolation is
usually unreliable in such cases.

28
 The R2 static, or coefficient of multiple determination is also
useful in evaluating the quality of a regression fit. Lipson and Sheth (2-2)
provide a clear definition of this statistic and its uses.

4.6 Multiple Heat Data

Draper and Smith (4-1) recommend the use of dummy variables to

account for differences between subsets (e.g., heats) of data. This method
requires adding a variable to the equation for each heat except one, increases
the size of the matrix to be inverted, and thereby increases the chance of
serious computer round-off errors.
Another approach is to use tensile or short-time stress rupture data
as measures of heat variability by including functions of one or both as inde-
pendent variables. This approach is particularly useful when predictions are
to be made for a particular heat for which such data are available. Heat sep-
aration is then based entirely on the tensile or short-time stress rupture
data used in the analysis; optimun heat separation in the sense of minimum
within-heat variance is not likely to be obtained. Tolerance intervals for
unknown heats are also likely to be in error. (Tolerance interval predictions
are frequently needed for production runs using many heats.)
The method of heat separation recommended here, though equivalent to
the dummy variable method, avoids the difficulties inherent in the two methods
above. The data for each variable should be centered about the mean value for
each heat. Equation 4.1 is replaced by

k
Yhjyh i (lbiCXij-Xih) (4.4)

where h is the index of the heat to which it belongs, and Yhj is the predicted
value of log life for a specific heat. The heat separation results are iden-
tical to those obtained with the use of dummy variables, but are more pre-
cisely and efficiently determined. Including tensile data as very short-time
stress rupture data has only a minor effect on determiningthe differences be-
tween heats since all data are used for this purpose.

29
 When the method of heat separation is used, the sum of squares mini-
mized by regression is based on the differences between observed values and
the predicted values of log life for each heat. Variability between heats is
accounted for, and an average constant for the material is calculated, by
additional operations (described below) on the values of

k
Chh - (b 'i h )
i-i

the heat constants of the equation.

The dummy variables method and the heat separation method both assume
parallel curves for each heat, that is, a constant ratio of lives between any
two heats, at all test conditions. This is approximately true in many sets of
creep or stress rupture data, although some lack of parallelism can be
observed.
A statistical test of the significance of lack of parallelism can be
applied. The residuals from each heat, which always sum to zero, may be fit
to a linear equation against any independent variable, or against the predic-
ted logarithms of life. The coefficients of such equations are generally not
significantly different from zero, indicating that the data do not disprove
the assumption of parallelism. In other words, apparent lack of parallelism,
at least in part, is due to random scatter in the data.
Even if the true relationships for the various heats are not repre-
sented by exactly parallel curves, the average curve determined by heat separ-
ation is still a valid prediction for average material if the heats analyzed
are representative of heats in the total population of heats. The assumption
of parallelism will increase the within-heat sum of squares where the curves
are not parallel. The tolerance interval will be more conservative and should
not lead to over-optimistic predictions for design. Multiple-heat analyses
using heat separation techniques usually give within-heat standard errors in
the range of 0.10 to 0.30, the same as analyses for single heats.
Heat separation allows each heat tested at more than one condition to
contribute to curve shape. Heats with only one test condition, such as spe-
cification point data, contribute only to the determination of the average
constant and the between-heat variance for the material. When all data are

30
analyzed together without regard to heat, specification point data can sig-
nificantly distort curve shape.
The analysis by heat separation techniques gives an equation with a
separate constant term for each heat and a set of common coefficients. The
relationship between the constant terms represents the relationship between
the logarithm of life among the heats at any test condition. When the heats
included in the analysis were not preselected for any particular characteris-
tic, the constants may be considered to be a random sample from the population
of all heats, and are assumed to have a normal distribution. When sufficient
heats are present, a good measure of between-heat variance may be obtained.
The next step is to determine the constant for the material. It
obviously should be based on some average of the constants for the individual
heats. One way to do this is a simple arithmetic average of the heat con-
stants. This is correct only when the between-heat variability is much larger
than the within-heat variability. The second approach is to weight each heat
constant by the number of data points in that heat. This procedure is correct
only when the between-heat variability is much smaller than the within-heat
variability. The heat separation method includes these two approaches as
extreme cases and also correctly treats the more usual case where the between-
and within-heat variabilities are of the same order of magnitude. Logically
consistent between-heat variance is calculated at the same time as the average
equation (material) constant.
Mandel and Paule (2-1) studied the situation where several labora-
tories measure a given property, but do not each make the same number of mea-
surements. A mathematically simple iterative procedure was reported for
estimating the "best" single value to represent the property, and of the be-
tween-laboratory variance. Creep or stress rupture data can be made to fit
Handel and Paule's model if "heats" (or other subsets of data) are substituted
for "laboratories," "heat constants" for the "average property value from each
laboratory," and the "square of the standard error from the regression" for
the "pooled within-laboratory variance." A short paraphrase of this analysis
is presented in Appendix B. Applied to creep or stress rupture data, the
method is simple, iterates quickly, and gives results which stand up well to
engineering inspection.

31
 A batch program in a version of FORTRAN is recommended when data are
retrieved from a data storage system or large amounts of data are supplied on
punched cards. Several standard analysis packages are available. These
include Statistical Analysis System (SAS) and Biomedical Computer Programs
(BMDP).
It is recommended that a standard version of a batch program perform
the analysis of the screening equations described previously. The batch
program should be capable of operating on these screening equations with or
without additional terms in auxiliary variables, or even with completely non-
standard equations. The minimum output should include the standard error of
estimate, the coefficients with their standard errors, and a table of pre-
dicted average lives for each heat and the average material curve. Some
analysts may also wish to define a lower level tolerance bound on the average
curve. Two plots of the data (coded by heat) should be generated, one with
curves for all equations on isothermal plots and another with all isotherms on
a single plot for the equation with the smallest standard error. Isothermal
deviation plots should also be produced for this equation. The table of aver-
age deviations for five cells of predicted lives for each temperature can be
printed for each equation. These recommendations reflect experience in
obtaining the minimum necessary information in a labor-efficient manner with-
out overwhelming the analyst with both paper and information.

4.8 An Example Analysis

A collection of 304 stainless steel data, reported by the National

Research Institute for Metals in Japan, was obtained. Six heats of hot ex-
truded material were represented in the collection of data. One of these
heats, AAA, showed on preliminary plots and individual heat regression
analyses to vary systematically from the other five heats and was not used in
the analysis. Except for its unusually low boron content, no metallurgical
reason was found for this difference.
One datum from heat AAB (1127.5 hours, 700 C, 8 kg/mm 2 ) was also dis-
carded since it appeared as an outlier in both the individual heat analysis

33
and in comparison with the data at 700 C from the other heats. The remaining
78 points from five heats were used in the analysis.
Separate analyses of each heat using a Larson-Miller equation with a
cubic function of the logarithm of stress had a pooled standard error of
0.092, an adequate measure of experimental error. Only one of these
individual heat analyses resulted in an equation with acceptable extrapolative
capability.
A plot of the data shows that the logarithm of time is approximately
linearly related to the logarithm of stress, which may account for the unac-
ceptable extrapolation of the cubic fits. The plot also shows a systematic
variation in the relationship between heats. Although this indicates that the
individual heat curves are not parallel, the. assumption of parallelism was re-
tained since it was expected to produce a reasonable curve for average
material.
The first step taken to arrive at the final equation was to use a
Larson-Miller equation that was linear in the logarithm of stress. The tem-
perature exponent was adjusted from -1 to -0.1. Finally an exponential term
of the form exp [A • (T-1202) 2]/a , was added with T in degrees F and a in
MPa. The within-heat standard error of the final equation is 0.127. The
square root of the total variance is 0.175.
The final within-heat standard error is higher than the pooled stan-
dard error of the individual heat analyses (0.092) largely because of non-
parallelism of the five heats tested. Two heats, AE and AF, fell above the
average curve except at low stress and high temperature. The other three
heats behaved generally in the opposite manner.
Figure 4.1 shows the data plotted as solid symbols for heats AAE and
AAF and open symbols for the other heats. The curve for average material is
also shown, extrapolated to 10 5
hrs. Its equation is

lo
logt-b 1b
° + b/T0 log
+ b2 loaT /T0 1

+ b3 *exp [A. (T-1202)2] /o

where T is degrees F and a is Hpa.

34
250

-30
200* *
U -25

120220

50 C5

00 C))2

904-

CAl

0 AAF

0 AA 5
310 102 4o5
Rupture Life, hours
FIGURE 4.1. OBSERVED AND PREDICTED STRESS RUPTURE LIVES FOR
NRIM 304 STAINLESS STEEL

35
 A better analysis could have been performed on the two groups of
heats separately, with considerably less lack of parallelism in each case, but
this was not done because two or three heats are insufficient to accurately
determine between-heat variance, and because there was no way to describe the
subpopulation of heats to which each analysis would belong.
The exponential term in the equation represents an effect centered at
1202 F (650 C). This temperature was determined from the data upon itera-
tion. It may be coincidence that 1202 F is also the center of the temperature
range in which carbide precipitation and intergranular corrosion occur in this
material, although the negative exponential term may be reflecting such a
phenomena.
The statistical significance of the variation between heats under the
assumption of parallelism is very high. The ratio of between-heat and within-
heat mean squares is 11.6 with 4 and 70 degrees of freedom, and is significant
with greater than 99.9 percent confidence. These data result in a somewhat
unsatisfactory analysis in the absence of information about the differences
between the two groups of heats (and the sixth heat which was not included in
the analysis). The method of heat separation, however, can be expected to
approach the true rupture relationship better than the simpler approach of
analyzing all data as one population irrespective of heat, especially with
only some of the heats being tested at some conditions of stress and temper-
ature. The lack of parallelism did cause the standard error to increase, and
consequently the prediction limits to expand, so as to compensate at least
partially for this defect in the model.

4.9 Summary

The method of creep and stress rupture data analysis described in

this section involves regression analysis of mathematical models, starting
with standard parametric models and modifying them as needed to improve the
fit and the ability to extrapolate reasonably. A method of handling mltiple
heat data is included in which the variability between specimens from a single
heat and the variability between heats are treated separately. The method
presented allows the determination of a single curve for the average
properties of a material from data on representative heats of that material.

36
 REFERENCES

2-1 Myers, R., Response Surface Methodology, Library of Congress Catalog No.
71-125611, 1976.

2-2 Hald, A., Statistical Theory With Engineering Applications, John Wiley &
Sons, Inc., 1952 (See Chapter 9, Table 16.4, Table 16.5, p. 426).

2-3 Brownlee, K. A., Statistical Theory and Methodology in Science and

Engineering, John Wiley & Sons, Inc., 1960 (See Chapter 10, Article 10.6,
Table 10.5, pp. 262-264).

2-4 Mandel, J., and Paule, R. C., "Interlaboratory Evaluation of a Material

with Unequal Numbers of Replicates", Analytical Chemistry, Vol. 42, No.
11, pp. 1194-1197, Sept. 1970; correction in Vol. 43, No. 10, Aug. 1971.

4-1 Draper, N. R. and Smith, H., Applied Regression Analysis, John Wiley &
Sons, Inc., 1966.

37
 APPENDIX A

MIL-HANDBOOK 5 GUIDELINE ON
CREEP AND CREEP RUPTURE DATA

ANALYS IS

Approved as shown here at the

58th Coordination Meeting in
Portland, Maine

October, 1979
9.3.6 Creep and Creep-Rupture Data

9.3.6.1 Introduction - Creep is defined as the time-dependent de-

formation of a material under an applied load. It is usually regarded as an
elevated temperature phenomenon, although some materials creep at room tem-
perature. If permitted to continue indefinitely, creep terminates in rupture.
(First stage or logarithmic creep exhibited by many materials at lower temper-
atures is not the subject of this section.) Creep in service usually occurs
under varying conditions of temperature and complex (multiaxial) stress
states, leading to an infinite number of stress-temperature-time combinations.
Creep data for use in general design are usually obtained under conditions of
constant uniform temperature and uniaxial-stress; this type of data is the
subject of this section.

9.3.6.2 Terminology - The definitions presented below will be help-

ful in preparing creep-rupture data for inclusion in MIL-HDBK-5.
Creep - The time-dependent deformation of a solid resulting from
force.
Note 1 - Creep tests are usually made at constant load and at
constant temperature. For tests on metals, the initial loading
strain, however defined, is not included.
Note 2 - This change in strain is sometimes referred to as creep
strain.
Primary Creep - Creep occurring at a diminishing rate, sometimes
called initial stage of creep.
Secondary Creep - Creep occurring at a constant rate, sometimes
called second stage creep.
Tertiary Creep - Creep occurring at an accelerating rate, sometimes
called third stage creep.
Creep Test - A creep test has the objective of measuring deformation
and deformation rates at stresses usually well below those which
would result in fracture during the time of testing.
Creep-Rupture Test - A creep-rupture test is one in which progressive
specimen deformation and the time for rupture are both measured. In

38
general, deformation is much larger than that developed during a
creep test.
Stress Rupture Test - A stress-rupture test is one in which time for
rupture is measured, no deformation measurement being made during the
test.
Total Strain - The total strain at any given time, including initial
loading strain (which may include plastic strain in addition to elas-
tic strain) and creep strain, but not including thermal expansion.
Loading Strain - Loading strain is the change in strain during the
time Interval from the start of loading to the instant of fill-load
application, sometimes called initial strain.
Plastic Strain During Loading - Plastic strain during loading is the
portion of the strain during loading determined as the offset from
the linear portion to the end of a stress-strain curve made during
load application.
Creep-Strain - The time-dependent part of the strain resulting from
stress, excluding initial loading strain and thermal expansion.
Total Plastic Strain - Total plastic strain at a specified time is
equal to the sum of plastic strain during loading plus creep.
Creep Stress - The constant load divided by the original cross-
sectional area of the specimen.
Elapsed Time - The time interval from application of the creep stress
to a specified observation.
Creep Rupture Strength - The stress that will cause fracture in a
creep test at a given time, in a specified constant environment.
Note: This is sometimes referred to as the stress-rupture strength.
Creep Strength - The stress that causes a given creep in a creep test
at a given time in a specified constant environment.
Rate of Creep - The slope of the creep-time curve at a given time
determined from a Cartesian plot.
Creep-Rupture Curve - The results of material tests under constant
load and temperature; usually plotted as strain versus time to
rupture. A typical plot of creep-rupture data is shown in Figure
9.3.6.2. The strain indicated in this curve Includes both the

39
initial deformation due to loading and the plastic strain due to
creep.

Initial Second Third

stage stage stage

' -Rupture

point
C.
Cr

"'"-Tronsition point
Slope: minimum creep rate
Creep intercept
''"Deformation due to loading

Time, hours

FIGURE 9.3.6.2. TYPICAL CREEP-RUPTURE CURVE

Master Creep Equation - An equation expressing combinations of

stress, temperature, time and creep, or a set of equations expressing
combinations of stress, temperature and time for given levels of
creep.
Master Rupture Equation - An equation expressing combinations of
stress, temperature, and time, that caused complete separation
(fracture or rupture) of the specimen.
Isothermal Lines - Lines of uniform temperature on a creep or stress
rupture curve.
Isostrain Lines - Lines representing constant levels of creep.

40
 9.3.6.3 Data Generation - The following paragraphs provide guide-
lines on testing methods and designing an experimental matrix for developing
creep and creep-rupture data.
Test Methods - Test methods must conform to ASTM E-139. However, it
is recognized that this standard allows considerable latitude in
procedures such that both the level and scatter of results can be
significantly affected.
In case of significant difference in results from different testing
sources, the following should be evaluated:
Material Condition (see Section 9.3.6.4)
Specimen Dimensions and Configuration (geometry effect)
Specimen Surface Preparation (residual stresses)
Specimen Alignment (concentricity, fixturing, load train, and loading
method)
Temperature Control (number, type, and location of sensors, reference
junction temperature control, monitoring and recording)
Extensometers (type, fixturing, and recording)
Strain Recording (records inelastic strain on loading and creates a
record that can be evaluated for test stability)
Documentation (testing procedures)
General Laboratory Conditions, Personnel Qualifications, Calibration
Intervals.
The submittor of a proposal should be prepared to provide documenta-
tion sufficient to permit a comparative evaluation of data. Inability to do
so may cause the rejection of some of the associated data or the entire
proposal.
Design of Experiments - A design of experiments approach to creep
data development is highly recommended because it provides the maximum amount

of useful data for the least expenditure of time and testing funds. If such
an approach is not used, it is quite likely that several times as many test
data will not serve as well in developing the desired mathematical models of
creep behavior as data developed through design of experiments. This section
is devoted to a description of the design of experiments approach which can be

41
used to develop regression models to mathematically portray creep rupture life
and creep as a function of temperature and stress.
One method for planning testing is to develop a test layout in matrix
form with temperatures in rows and expected creep lives in columns. Then
through testing, simply fill out" the blocks within the matrix. There should
be a minimum of eight observations per isothermal line or twenty observations
per Larson-Miller or other regression model. This ensures coverage of all of
the conditions of interest. Further explanation of this method, by way of an
example, is provided in Section 9.3.6.8.
Choosing the Number of Temperatures and Life Intervals - Before the
test matrix can be formed, the interval sizes must be considered, first for
temperature and then life.
(a) Temperature - A range of temperatures is usually required. For
example, if the experiments must range from 1000 F through
1500 F, a choice must be made whether to perform tests at six
levels (1000 F, 1100 F, 1200 F, 1300 F, 1400 F, 1500 F) or maybe
at three levels (1000 F, 1300 F, 1500 F). The decision for this
can be quite complicated and based on such phenomena as:
(1) The relative closeness of the isothermal lines
(2) Parallel or divergent isothermal lines
(3) The precipitation of secondary phases within the life ranges of
interest.
However, this selection can be greatly simplified with very little
user risk. Start with the lowest temperature and then choose the
next temperature line such that at least one level of testing stress,
on the log stress - log life plot, will be common to both tempera-
tures. Then, proceed to the next temperature line, etc., ensuring
like stress values on adjacent temperature levels.
(b) Life - Divide a log-life cycle into four equidistant segments.
For example, between 100 hours and 1000 hours, the division
would be approximately 180 hours, 320 hours, and 560 hours on
the log-life scale. These divisions are far enough apart to
insure a well-defined curve and a minimum overlap of data. To
convert from temperature and life desired to temperature and

42
 test stress requires that there be some prior knowledge of this
relationship. If there is not prior knowledge, a series of
"probe" tests must be made to locate the isothermal lines on a

log-log plot.
Choosing the Number of Heats - Batch variations in chemistry, heat
treating, etc., can cause considerable variations in the mechanical properties
of an alloy. This difference is referred to as the heat-to-heat component as
opposed to the within-heat components of variance. * Heat-to-heat standard
deviation is usually 50-70 percent of the within-heat standard deviation. The
root sum square of the two components of variance produce a measure of scatter
about the regression that when added to the curve fitting error gives the re-
gression parameter called the SEE (Standard Error of Estimate). The SEE is a
product of the regression analysis; it is rarely determined as defined above.
It is this parameter which fixes the design minimums about the regression
estimates of the typical or mean values.
To make a mathematically sound decision on the minimum number of
heats that should be used in a given analysis, it is necessary that an esti-
mate of heat-to-heat and within-heat variance be known. This can usually be
estimated from like alloys, or calculated from development data. Simulation
has shown the following minimum number of heats to be satisfactory:
(1) When the heat-to-heat component of variance is less than 25
percent of the within-heat variance, use two heats equally for
the sample sources.
(2) When the heat-to-heat component of variance is between 25-65
percent of the within-heat variance, use three heats equally.
(3) When the heat-to-heat component of variance is greater than 65
percent of the within-heat variance, use five heats equally.
Heats should be distributed randomly and essentially equally throughout the
test matrix to ensure an unbiased heat distribution.

*The within heat variance is the pooled variability of data from all heats,
where the variability for each heat is calculated about its own average
regression line. The heat-to-heat variance is calculated from the vari-
ability of each heat's average regression line about the overall average
regression line of all heats. All heat average curves are assumed to be
parallel in log life.

43
 When regression models are developed from data that were not taken
from an experimental model, the heats are rarely chosen randomly. Therefore,
unless there are large samples of data in all areas of the regression matrix,
this imbalance of heat sample sizes must be accounted for as described in
Section 9.3.6.5. The order of testing must also be randomized so that
anytime-oriented, operator-oriented, or machine-oriented effects are randomly
distributed within the test matrix as described in Reference 9.3.6.3.

9.3.6.4 Data Collection and Interpretation - After a desired group

of creep and/or creep-rupture data have been experimentally developed or
isolated in pre-production files, it is necessary to carefully collect and
interpret these data in accordance with the following guidelines:
Data Collection - For iso-strain creep, the collected data will
include stress, temperature, modulus and plastic strain on initial loading,
and strain-time pairs sufficient to define a curve. While the strain-time
pairs will be only those for the iso-strains of interest after inelastic
strain on loading has been included in the reported strain, it may be that the
reported data may not correspond to the iso-strain levels. Consequently, the
iso-strain-time pairs may be read from a smooth curve drawn through the values
recorded during the test.
For rupture, the collected data will include stress, temperature,
time-to-rupture, and percent elongation, and reduction of area. The percent
elongation and reduction of area can then be used to define the rupture
ductility curves or equations.
Data Interpretation - The state of the art for interpreting these
types of creep and rupture data requires that a certain amount of judgement be
allowed. The general approach will be to optimize one of several empirical
equations that best follows the trend of the data, using life (or time) as the
dependent variable. The independent variables will include stress and temper-
ature for rupture and iso-strain creep curve and will also include strain for
iso-strain creep curves.
Rupture ductility can be an exception to the above because of complex
behavior and data scatter. At least a cautionary note should be given in the
introductory material on the time and temperatures included in the rupture

44
data. Some materials exhibit such low elongation in certain time-temperature
regions that normally reasonable values of design creep strain cannot be
achieved without risk of fracture.
The interpretation of creep and rupture data should also include the
variables that are reflected in the background data reporting requirements
(discussed in the next subsection). Depending on the information content of
the data and the type of variable, it may be desirable to develop a series of
equations or to include additional physical variables in the regression
analysis. The proposal should demonstrate that these additional variables
have been evaluated and appropriately treated in the analysis.
The individual interpreting the data should also take note of the
following special types of data, and consider the following recommendations on
their use:
Specification Data - Virtually all alloys used for high-temperature
applications are controlled and purchased by a process control vari-
able generally called the "spec point." Therefore, there will often
be large quantities of data available from quality control data re-
cords at the specification condition. The data will contain many
heats and serve as an excellent measurement source in regression
equations of what is referred to as "scatter." Therefore, in regres-
sion modeling, specification data are often the major source of the
scatter measurements. The slope measurements must come from the ex-
perimental design matrix.
Specification data can also be used to (1) determine through
analysis-of-variance techniques the fractions of the scatter due to
heat-to-heat variations, etc., (2) determine through distribution
analysis if the data are normal, log normal, etc., and (3) find out,
if it is not normal, what transformation is required.
Outliers - These can be excluded only if the tests are demonstrably
invalid or if the effect on the equation and the statistical param-
eters is unreasonable. Since the exclusion of outliers normally in-
volves a certain degree of judgement, it should only be done by a
knowledgeable and experienced individual.

45
 Discontinued Tests - These can be included if longer lived, or ex-
cluded if shorter lived, than the average life of the data subset
(lot, section thickness, etc.) to which they belong. (Also, see
censored distributions in Section 9.3.6.5 on Data Analysis
Procedures.)
Stepped-Tests - If the load on the specimen had been increased or
decreased after the initial loading, this test result shall be
excluded.
Truncating Data - Certain equations, notably the parametrics, are
often unable to properly represent a mix of shorter and longer time
data; these equations can severely over-predict creep and rupture
lives less than ten to thirty hours. Similarly a preponderance of
short time data can cause longer lives to be over-predicted. Elimin-
ating such data requires truncating the data (or subset): this is
done by removing all data above (or below) a fixed stress level, even
though normally acceptable data are excluded.
Background Data Reporting - The significance and reliability of creep
data generated at elevated temperatures for heat-resistant alloys are, to a
major extent, a function of detailed factors which relate to the material, its
processing, and its testing. Hence it is necessary to evaluate not only the
property data, but also correlative information concerning these factors.
It is not possible to specify the individual items of correlation in-
formation, or the minimum thereof, which must be provided with elevated tem-
perature property data to make those data properly meaningful. Individual
alloy systems, individual product forms, and individual testing practices can
all be quite unique with regard to associated information which should be pro-
vided with the data. A certain minimum amount of information is required for
all data, including:
(1) Identity of alloy
(2) Chemical composition of the specific material tested
(3) Form of product (sheet, forging, etc.)
(4) Heat treatment condition
(5) Producer(s)

46
 (6) Specification to which the product was produced
(7) Date when part was made.
Lack of such information is a sufficient basis for rejection of a particular
data set.
In addition, it is vital that the individual submitting data consider
those factors which contribute to the uniqueness of the alloy, the processing,
and/or the testing, and give thought to information which is pertinent to that
uniqueness. Thus, grain size can be a significant variable not only between
cast turbine blades but within a single blade; thermomechanical working pro-
cesses may result in significar' 'y different properties (not only higher, but
lower as well); and test specimen design can affect resultant data. It is
mandatory that knowledgeable personnel be involved when data are submitted for
evaluation and potential use. Any correlative data that can be provided will
aid the analyst in identifying valid reasons for rejection of data which may
not fit the trends of the other data (outliers). Such apparent outliers may be
indicated through the analysis of between-heat variance as described in
Section 9.3.6.5.
These examples illustrate the need for adequate information:
(1) Creep-rupture specimens are being machined from cast high-
strength nickel-base alloy turbine blades. At center span
location, specimens are 0.070 to 0.090 inch diameter, while at
the trailing edge specimens are flat and 0.020 inch thick. The
flat specimens are typically about one Larson-Miller parameter
weaker than the round specimens, which is attributable both to
the thickness effects of the thin specimens and to the finer
grain size at the trailing edge. In addition, the trailing edge
specimens exhibit more scatter. Hence, the availability of
associated information is vital when considering data from
specimens machined from cast turbine blades.
(2) Comparison of the creep-rupture properties of Waspaloy and
Superwaspaloy shows that the latter is much weaker at temper-
atures approaching the upper bounds of utility of the alloy.
The significantly lower properties at higher temperatures are
attributed to a finer grain size of the Superwaspaloy and also

47
 to a recovery process that may well be occurring at these tem-
peratures. This alloy is subjected to extensive thermome-
chanical working, and some of the strengthening gained by the
associated warm working is lost at the higher testing tempera-
tures. This effect clearly indicates that processing history
significantly affects the levels of mechanical properties and
hence must be adequately documented when property data are
submitted.

9.3.6.5 Data Analysis Procedures - After an acceptable data

collection has been obtained and interpreted, it is possible to proceed in
analyzing those data and developing mathematical models of creep and creep-
rupture behavior. The objective of the procedures described in the following
paragraphs is to calculate creep and rupture life as a function of test con-
ditions and other significant variables. This calculation is done to provide
an average curve and a measure of the expected variability about the average.
The approach that is discussed involves regression analysis to optimize the
fit of an equation to the data set. Linear regression analysis is described
in Section 9.6.3. The following information provides guidelines in the
application of regression analysis to creep and rupture data and recommends
approaches to specific problems that are frequently encountered.
General - It is assumed that life or time is the dependent variable
for the rupture or iso-strain creep equation analysis, respectively, and the
logarithmic transformation of the dependent variable is normally distributed.
The data set will nearly always contain a variety of stresses and
temperatures. If the data set is the product of a very well-balanced test
design (see Section 9.3.6.3), good results may be obtained by independently
fitting each temperature. Since this type of data set is often not available
and the approach sacrifices the opportunity for interpolation, the discussion
will assume that at least temperature and stress are used as independent
variables.
In order to achieve good results, it may be necessary to consider
other variables. Some variables are continuous physical variables that are
incorporated into the regression variables, e.g., section size. Other

48
variables may occur as discrete subsets that require modifying the regression
analysis (this is discussed under Subsets of Data). In such cases, it may be
necessary to group the data per subset for data reporting if the regression
analysis cannot easily accommodate the observed subsets.
Selection of Equations - For iso-strain and rupture time, as a func-
tion of stress and temperature, a number of relationships have been proposed.
Some useful ones are:

(1) log t - c + bl/T + b2 X/T + b3 X2 /T + b4 X 3 /T

(2) log t - c + bl/T + b2 X + b3 X2 + b4 X3
2 3
(3) log t - c + biT + b2 X + b3 X + b4X
(4) log t - c + (T - Ta) (bl + b2 X + b3 X2 + b4X 3 ).
These are the Larson-Miller, Dorn, Manson-Succop, and Manson-Haferd,
respectively, where
c is the regression constant
bi are the coefficients (bI through b4 )
t is time
T is absolute temperature (Ta is the temperature of convergence of
the iso-stress lives)
X is log S (stress).
While all of the forms may be used to model a data set with varying degrees of
goodness of fit, experience and practice indicates the Larson-Miller relation-
ship adequately models most materials and is usually the preferred equation
form.
If none of these standard forms satisfactorily follow the data
trends, various combinations of stress and temperature may be tried. For
example, terms can be selected from P matrix obtained using the cross products
of T-1 , TO, T 1 with S- 1, S0 and S1 . methods for generalizing and
applying these equations are discussed by M. K. Booker in a paper, "Regression
Analysis of Creep-Rupture Data - A Practical Approach", published in Reference
9.3.6.5.
The exact fo in of the functions should reflect the data and reason-
able boundary conditions. Quadratic, quartic, etc., can be expected to give
poor boundary conditions, e.g., zero life at zero stress, and should be
avoided. Extrapolation by users of the equation is inevitable (though it is

49
not recommended) so other general equations must be checked for unusual behav-
ior beyond the data - this can be done in many cases by differentiating to ob-
tain maxima and minima. In general, short times should give strengths approx-
imately corresponding to tensile yield and ultimate strength; zero stress
should predict infinite life.
Metallurgical instabilities and transition regions may present diffi-
culties in some analyses. Methods for handling such problems have been dis-
cussed by L. H. Sjodahl in a paper, "A Comprehensive Method of Rupture Data
Analysis With Simplified Models", published in Reference 9.3.6.5.
Optimum Fit - Guidelines for an optimum fit are:
(1) A minimum number of terms. With two independent variables
(a, T), six regression variables are reasonable, with each
additional physical variable allowing two additional regression
variables.
(2) Reasonable curve characteristics for material behavior,
including extrapolation.
(3) Minimum standard error and maximum correlation coefficient (as
long as I and 2 are not violated). Standard errors are
typically between 0.1 and 0.2.
(4) Uniform deviations (see a later paragraph on Weights for a brief
discussion of nonuniform deviations and their analytical
treatment).
Subsets of Data - A nonnormal or multi-modal population, or an
excessive standard error may indicate the presence of subsets. However, an
apparently typical data set may contain subsets that should receive special
consideration.
One type can be treated by adding physical variables to the regres-
sion analysis. For example, different thicknesses of sheet material may give
different average lives. Including sheet thickness in the regression should
not only improve the fit but also avoid the risk of misrepresenting the behav-
ior of the material. Section thickness, distance from surface, and grain
size, are other examples of subsets that can be treated as regression varia-
bles. Section thickness and distance from surface refer to the location of

50
the specimen in terms of the geometry of the original material, e.g., finish
work thickness, final heat treat thickness, etc. .
A second type is not typically subject to use as a regression vari-
able. Examples of these are orientation (L, LT, and ST), or different heats
(chemistry). A decision must be made whether to treat these as unique subsets
to be analyzed separately (if properties are different) or as randomly dis-
tributed subsets. Orientation will usually be analyzed separately while heats
will usually be randomly distributed subsets.
The theory of the treatment of randomly distributed subsets has been
developed in Reference 9.3.6.3 while the application to lots of material
(actually "heats" based on chemistry) is considered in a paper by L. H.
Sjodahl in Reference 9.3.6.5. Treating subsets as random affects the calcula-
tion of both the average curve and the standard error. While the effect on
the standard error may become insignificant as the number of subsets exceeds
ten (depending on the relative contribution to the total standard error), the
effect on the trend of the calculated average remains. Lots whose average
lives are uniformly displaced (parallel) in logarithm of life, or are not sig-
nificantly nonparallel, are discussed by L. H. Sjodahl in Reference 9.3.6.5.
There is no known published reference for treating nonparallel lots. The data
permitting, individual lots can be fitted, the within-lot variances pooled,
and the average and variance of lot averages calculated for selected stress-
temperature combinations. After calculating the total variance and the desired
lower level tolerance limit* (X - ks) at each stress level curves can be drawn
and, if desired, equations be fit to the X's and (X - ks)'s. It should be
noted that the equation for (X - ks) is not likely to properly reflect uncer-
tainty in coefficients that would be obtained by normal fitting procedures.
Alternatively, all the data for nonparallel lots can be pooled and the var-
iance weighted, providing sufficient lots are represented and the average
curve is reasonably similar to that of the first approach.
Weights - Rupture and iso-strain creep curves will not normally re-
quire weights to obtain uniform variables. Analyses including strain as a

*Tolerance limits used here are one-sided and are normally developed for
tolerance levels of 90 or 99 percent at a confidence level of 95 percent.

51
variable frequently will. Variables other than strain, temperature, and
stress will require evaluation for uniform variance. A paper by L. H.
Sjodahl, Reference 9.3.6.5, provides further discussion of weighting.

9.3.6.6 Preparation of Creep-Rupture Data Proposals - Creep-

rupture proposals developed for review and possible inclusion in MIL-HDBK-5
should contain the following information and meet the associated criteria.
Data Reporting - The background information shall meet the require-
ments of Section 9.3.6.4. The test results shall be listed in a manner such
that all data are identifiable in terms of the material and test background
information, as well as the test conditions used in generating the data (see
Section 9.3.6.7 for an example).
Analysis Reporting - The analysis report will display the following
(see Section 9.3.6.5 for details):
(a) Trials - Equations tried and reason for rejecting.
(b) Data rejected - Reason.
(c) Best-fit details - Listing of data, calculated values, and
deviations. All data are to be clearly traceable in terms of
data reporting requirements.
(d) Standard error or total variance and correlation coefficient.
(e) Subset variance - If random subsets ate used, report both the
pooled within-subset variance and the between-subset variances
as well as the total variances.
(f) Constants - Report the average regression constant and the
regression constants for any subsets.
(g) Coefficients - Report the numerical value of the coefficient of
each regression variable and its standard error.
(h) Equation - Exhibit the equation used; with the coefficients,
bi, traceable to the numerical listing in above item (g).
(i) Deviation - Exhibit plots of deviations in life vs. calculated
life for each temperature and, in so far as possible, identify
according to subsets. It is also possible to provide a summary
table of deviations. As an example of iso-strain creep or
rupture, divide the life range of the data in five equal

52
 logarithmic increments and, for each temperature, give the
algebraic sum of the deviation within that increment. If random
subsets are used, the deviations summed are to be those from
within the respective subsets.
(J) Data and Curve Comparison - Display the data against the calcu-
lated curves, giving both the average and, for proposal consid-
eration, at least 99 percent lower level tolerance limits lines
at 95 percent confidence level. Encode the data with symbols as
in the deviation plots. Scale the coordinates such that the
curves have an apparent slope of about -1.0. Use scales appro-
priate for the most significant form of the regression variable,
usually log stress versus log life, with life (the dependent
variable) on the abscissa and stress on the ordinate.
(k) Curve Extrapolation Tests - Exhibit the 99 percent probability
of exceedance curve from one hour to 106 hours and the average
curve for the corresponding stress levels. Representative
curves may be used that Include the extreme values of the inde-
pendent variables as represented in the data.
The above recommendations apply to the incorporation of new creep
and/or stress rupture curves in the Handbook. The incorporation of creep
monographs in the Handbook has been discontinued. The creep monographs in the
Handbook will be replaced as the data are reanalyzed and new analytically
defined creep and stress rupture curves are developed.

9.3:6.7 Data Presentation - The presentation for MIL-HDBK-5 will

include one or more pages of correlative information, equations, and curves as
needed. The requirements on each will vary with the problem and should be
reasonably obvious from the data, background information, and analytical
results. The following data presentation is representative.

53
 HS 188 RUPTURE

Alloy Designation:

Specification(s) AMS 5608

Product Form Sheet
Heat Treatmear 2150 F anneal, rapid cool
Number of Vendors One
Number of Lots Ten

Specimen Description:

Type Sheet
Gage Length 1.125 inches
Gage Width 0.250 inch
Gage Thickness 0.01 to 0.10 inch

Test Conditions:

Number of Test Three

Laboratories

Number of Tests 540

The lines shown on Figure 9.3.6.7 are limited to the available test
times and temperatures, because extrapolation in time or temperature has def-
inite risks. If extrapolation is necessary, the equation defining these lines
should not be used below 1300 F or above 60 ksi. (These extrapolation limits
will, of course, be different for each alloy.) The standard error of estimate
(SEE) is given on Figure 9.3.6.7, along with the associated degrees of free-
dom. This parameter can be used to compute desired minimum curves, e.g., 95
percent, 99 percent, etc.

54
 50

40 0007.
0.3 8.

I ] il
0.4 9.
0.5010.
150F00601.

100

c,~170 thei
define
eqatonofth aerge

X~~ isE i is th
log90)
where 178
inlcio5onti0s
10s og 0

t ~~ ~ ~ ~
isthetorpurrnsor ie

Aofiins Th i verage
nteodrgvn
onstant,()-7.2
n h
co effiinentse eqaibi)th veae
-2.710+0
4.3257E+y-0c4 +
3177E+0 + b.*j3E4 -I.3014+

b4 3Tb(2/-00055 9
Mean Values of Regression Variables
4.6704E-04 2.8074E-08 8.4749E-06 1.3323E+01 3.0104E-01
Within-Heat Variance 0.02713
Ratio of Between-To- 0.16787
Within-Heat Variance
Intermediate Variance Ratio 1.01029

9.3.6.8 An Example of the Use of Experimental Design for the Purpose

of Developing Regression Models - By a slight chemistry change and modifica-
tion of the heat, the former Alloy 325 is now believed to have an increased
stress rupture life of 20-30 percent. It is desired to fully characterize
these properties over the 1600 F to 1900 F range. Average creep life is to be
from 10 hours to 1000 hours.
Nineteen stress rupture tests from two heats of the new alloy aver-
aged 37.4 hours at 30 ksi/1800 F, s - 0.150 (log base 10). Figure 9.3.6.8(a)
is a log-log mean life plot of the predicted stress rupture properties of the
modified Alloy 325 based on a predicted value. A 1750 F line has been added
to the original plot. From this log-log plot, it can be seen that only three
temperatures need to be tested because there are stress levels in common with
the 1600 F line, and the same is true for the 1750 F and 1900 F lines.
Next, the three temperature lines are bracketed with the 10 hour-1000
hour life range. [See Figure 9.3.6.8(b)]. The stress levels are then chosen
to give the desired life. There are 25 tests required doing it this way. All
25 could be run, or three tests could be randomly eliminated from the center
cells of the matrix (see the circled cells). That would leave 22 tests, which
is near the minimum of 20. These tests could then be run, added to the 19
specification data points @ 30 ksi/1800 F. This would now constitute the data
set. Table 9.3.6.8 shows the results of a simulated sampling.
A Larson-Miller analysis of the data produced the curves seen in
Figures 9.3.6.8(c) and (d). The data plotted with the temperature lines of
Figure 9.3.6.8(d) confirm a good fit over the range of data. The approach
described in this example can be used for any creep or rupture experimental
design.

56
 100 _ _- - - - - -

*~ F)

10

Time, hours

FIGURE 9.3.6.8(a). ESTIMATED STRESS RUPTURE CURVES FOR ALLOY 325 (MOD)

HOURS
3 6 10 15 32 56 100 180 1320 5601000 5600 F

TI 63 59 54 52 48 45 42 139 36 16001
T2 42 39 36 32 29 27 25 22 20 1750
T3 25 22 20 7 15 12 10 1900
T4
CL T5
w T6
T7
T8

FIGURE 9.3.6.8(b). EXPERIMENTAL DESIGN MATRIX FOR CREEP RUPTURE

57
TABLE 9.3.6.8. RESULTS OF SIMULATED SAMPLING OF CREEP-RUPTURE DATA

ksi 1600 F ksi 1750 F ksi 1900 F

63 19.0 hra. 42 8.8 hrs. 25 27.6 hrs.

59 11.1 hra. 39 35.5 hrs. 22 23.9 hrs.
54 36.3 hrs. 36 52.3 hre. 17 65.4 hrs.
52 170.7 hrs. 32 71.8 hrs. 15 140.3 hrs.
45 148.0 hrs. 29 121.9 hrs. 12 257.5 hrs.
42 376.0 hra. 27 355.9 hrs. 10 623.5 hrs.
39 806.0 hra. 22 389.0 hra. *
36 878.0 hra. 20 2912.4 hrs. *

*No interest.

SPECIFICATION DATA
@ 30 KSI/1800 F

Hours

41.4 33.1 70.5 36.1

16.5 27.4 37.5 34.9
35.0 33.4 48.6 74.2
33.6 51.3 29.0 47.5
32.6 42.7 26.4

n - 19, X - 37.4, S(log 10) - 0.150

58
 10

-0 LA

10

40 42 44 46 48 50 52
Larson-Miller Parameter, C=19

FIGURE 9.3.6.8(c). ALLOY 325 (MOD) STRESS RUPTURE TYPICAL LIFE

100 _ _

X oerage
specimendata at-3 ki80---
*0

10FIGUR 1000 3000

Time, hours
FIUE9.3.6.8(d). ALLOY 325 (MOD) STRESS RUPTURE TYPICAL LIFE

59
 REFERENCES

9.2.2.3(a) Weissberg, A., and Beatty, G. H., "Tables of Tolerance-Limit

Factors for Normal Distributors", Battelle Memorial Institute
(December 1969).

9.2.2.3(b) Owen, D. B., "Tables of Factors for One-Sided Tolerance Limits for
a Normal Distribution", Sandia Corporation (April 1958).

9.2.2.3(c) Owen, D. B., Handbook of Statistical Tables, Addison-Wesley

(1962).

9.2.2.3(d) Guenther, W. C., Concepts of Statistical Inference, McGraw-Hill

(1965).

9.3.2.3 Templin, R. L., et al, "Typical Tensile and Compressive

Stress-strain Curves for Aluminum Alloys 14S-T, Alclad 14S-T,
24S-T, and Alclad 24S-T Products", Alcoa Research Laboratories
Technical Paper No. 6 (1942).

9.3.2.4(a) Ramberg, Walter, and Osgood, William R., "Description of Stress

Strain Curves by Three Parameters", NACA TN 902 (July 1943).

9.3.2.4(b) Hill, H. N., "Determination of Stress-Strain Relations from

Offset Yield Strength Values", NACA TN 927 (February 1944).

9.3.3 Burt, C. W., et al, "Mechanical Properties of Aerospace Structural

Alloys Under Biaxial-Stress Conditions", Battelle Memorial
Institute, Technical Report AFML-TR-66-229 (August 1966).

9.3.4(a) "Manual on Fatigue Testing", American Society for Testing and

Materials (1949).

9.3.4(b) "A Guide for Fatigue Testing and the Statistical Analysis of
Fatigue Data", American Society for Testing and Materials ASTM STP
91-A (1963).

9.3.4(c) Aircraft Fatigue Handbook, ARTC Report W-76 (January 1957).

9.3.6.3 Mandel, J., and Paule, R., "Interlaboratory Evaluation of a

Material with Unequal Numbers of Replicates", Analytical
Chemistry, Vol. 42, No. 11, pp 1194-1197, Sept. 1970; correction
in Vol. 43, No. 10, August 1971.

9.3.6.5 "Characterization of Materials for Service at Elevated

Temperatures", Report No. MPC-7, Presented at 1978 ASME/CSME
Montreal Pressure Vessel and Piping Conference, Montreal, Quebec,
Canada, June 25-29, 1978.

9.4.1.3(a) "Military Standard, Fasteners, Test Methods, Test 4, Joint Shear

Strength", MIL-STD-1312, Department of the Navy, Naval Air Systems
Command, Custodians.

60
 9.4.1.3(b) "Military Standard, Fasteners, Test Methods, Test 13, Double Shear
Test", MIL-STD-1312, Department of the Navy, Naval Air Systems
Command, Custodians.

9.4.1.3(c) "Military Standard, Fasteners, Test Methods, Test 20, Single Shear
Test", MIL-STD-1312, Department of the Navy, Naval Air Systems
Command, Custodians.
4

9.4.1.3(d) "Fastener and Joint Testing Procedure, Room Temperature", ARTC

Report No. ARTC-33, Aerospace Industries Association of America,
Inc.

9.4.3 Hood, D., et al, "An Investigation of the Generation and

Utilization of Engineering Data on Weldments", The Boeing Company,
Technical Report AFML-TR-68-268 (October 1968).

9.5.1.4.1 "Standard Method of Test for Plane-Strain Fracture Toughness of

Metallic Materials", ASTM E399-72, ASTM Standards, Part 31 (1972).

9.5.1.5.1 Feddersen, C. E., "Evaluation and Prediction of the Residual

Strength of Center-Cracked Tension Panels", Damage Tolerance in
Aircraft Structures, ASTM STP 486, American Society for Testing
and Materials (1971).

9.5.2.2(a) Sprowls, D. 0., and Brown, R. H., "What Every Engineer Should Know
About Stress Corrosion of Aluminum", Metals Progress, Vol. 81, No.
4 (April 1962).

9.5.2.2(b) Handbook on Corrosion Testing and Evaluation, Edited by W.H.

Ailor, John Wiley and Sons, Inc., page 252 (1971).

9.6 Natrella, M. G., Experimental Statistics, National Bureau of

Standards Handbook 91 (August 1, 1963).

9.6.1.1 Lilliefors, Hubert W., "On the Kolmogorov-Smirnov Test for

Normality With Mean and Variance Unknown", American Statistical
Association Journal, pp 399-402 (June, 1967).

9.6.4.1 Owen, D. B., "Factors for One-Sided Tolerance Limits and for
Variables and Sampling Plans", Sandia Corporation Monograph
SCR-607 (March, 1963).

61
 APPENDIX B

A METHOD OF CALCULATING BETWEEN-HEAT VARIANCE

AND AVERAGE EQUATION CONSTANT

Mandel and Paule's analysis equates two expressions for the within-
heat variance, Vw . The first, from the regression, is hj
Z2 /hj-Yh)2/
(N-H-k)

where h and j are indices respectively of the heats (1 < h < H) and of points
in a heat (1 < J <nh), N is the total number of points (Enh), Yh is the
predicted value of the dependent variable for heat h, and k is the number of
independent variables.
The second expression for Vw is derived from the spacing between the
parallel heat curves. This can be measured at any point along the curves; for
convenience the constants, ch, of the individual heat equations can be used.
This is equivalent to adjusting all data to the test condition represented by
zero values of all variables.
The variance of each ch is given by Vh - VB + Vw/nh, where Vb is the
between-heat variance, since the within-heat variability contributes to the
observed scatter in the ch. By introducing the term A - VB/VW (the ratio of
between- to within-heat variance),

Vh - Vw (A+l/nh) - E (ch-c)2/(H-1) on the average,

h

giving the second expression for Vw,

Vw E[(Ch- c)2 /(Hl)/(A+l/nh)]

These two expressions are made equal iteratively by adjusting the

value of X. Of course, being the ratio of variances, A cannot be negative. X
is thereforeset equal to zero when the calculated optimum value is negative
(between-heat variance is negligible). VB is now AVw . Mandel and Paule fur-
ther show that the weights used in calculating c (the average equation con-
stant) are wh - 1/(A+1/nh), and the variance of 1Eis equal to Vw /Ewh
h

62