Proceedngs of the PCI.
fHWA Internalional Symposium on High Performance Concrete New Orleans, Louisiana, Oclober 2022, 1997
CONCRETE
MIXTURE OPTIMIZATION USING STATISTICAL DESIGN METHODS
MIXTURE
Marcia J. Simon Research Highway Engineer Federal Highway Administration McLean, Virginia Eric S. Lagergren Mathematical Statistician National Institute of Standards and Technology Gaithersburg, Maryland Kenneth A. Snyder Physicist National Institute of Standards and Technology Gaithersburg, Maryland
ABSTRACT
The optimization of mixture proportions for highperformance concretes, which contain many constituents and are often subject to several performance constraints, can be a difficult and timeconsuming task. Statistical experiment design and analysis methods have been developed specifically for the purpose of optimizing mixtures, such as concrete, in which the final product properties depend on the relative proportions of the components rather than their absolute amounts. Although mixture methods have been used in industry to develop products such as gasoline, metal alloys, detergents and foods, they have seen little application in the concrete industry. This paper describes an experiment in which a statistical mixture experiment was used to optimize a sixcomponent concrete mixture subject to several performance constraints. The experiment was performed in order to assess the usefulness of this technique for high performance concrete mixture proportioning in general.
INTRODUCTION
In the simplest case, portland cement concrete is a mixture of water, portland cement, fine aggregate, and coarse aggregate. Additional components, such as chemical admixtures (air entraining agents, superplasticizers) and mineral admixtures (coal fly ash, silica fume, blast furnace slag), may be added to the basic mixture to enhance certain properties of the fresh or hardened concrete. Highperformance concrete mixtures, which may be required to meet several performance criteria (e.g., compressive strength, elastic modulus, rapid chloride permeability) simultaneously, typically contain at least six
components.
230
coarse aggregate. a statistically designed mixture experiment was used to identify the best factor settings for optimizing properties of high performance concrete. they may not provide the best combination of materials at minimum cost.
In this study... trial and error or "one factor at a time" approaches sufficient for a conventional mix will be inefficient and costly. but it does not provide a procedure for finding the proportions which provide the best settings to meet a number ofperfonnance criteria simultaneously. Because mixture experiments have not been readily used in the concrete industry..06 MPa (3200 psi). However. The factorial (independent variables) approach permits the use of classsical factorial and response surface designs':" . but has the undesirable feature that the experimental region changes depending on how the q mixture components are reduced to qI independent factors. cement. For concrete. "The optimization of the composition of a high performance concrete (HPC) is at present more of an art than a science . The materials (components) used included water. may not require a significant number of trial batches to find an appropriate mix. the sum of the volume fractions is constrained to sum to one. Because the volume fractions must sum to unity. target 42day "rapid chloride" (ASTM C1202) test (RCn measurement less than 700 coulombs.
231
. and minimum cost (dollars per m'). and fine aggregate. the American Concrete Institute (ACI) guideline document for mix proportioning' provides a method for proportioning one mix."
Even for conventional concrete mixes. but the analysis of such experiments is more complicated. More importantly.
One viable experiment design option for concrete mixtures!" is the factorial design. There are advantages and disadvantages to both the mixture and factorial approaches. the utility of this approach for optimizing concrete properties was investigated. as in the ACI mix design approach'. for a concrete mixture containing six or more components which must satisfy several performance constraints.02 MPa (7400 psi). The selection of mix proportions for a conventional concrete mix. in which the q mixture components are reduced to qI independent factors by taking the ratio of two components. lday target compressive strength of 22. with strength as the primary criterion. micro silica (silica fume). An experiment was designed to find the optimum proportions for a concrete mix meeting the following conditions: 50 to 100 mm (2 to 4 in) slump for the fresh concrete. the component variables in a mixture experiment are not independent. For example. the experimental region of interest is defined more naturally in the mixture experiment approach. In a mixture experiment. 28day target compressive strength of 51. highrange water reducing admixture (HRWRA). the total amount (mass or volume) of the mixture is fixed and the factors or component settings are proportions of the total amount.In a recent paper. Rougeron and Aitcin' stated that. The recent ACI guideline document for proportioning highstrength concrete containing fly ash' also does not provide a means for optimizing mixtures.
This design Figure 2.1. Layout of experinrent design includes all vertices. midpoints of edges. For example.0)
lQ(V. The axis for each component X.1)
Figure 1. where each X. The points shown in Figure 2 represent mixtures included in the experiment. For three components. The volume fractions of these components sum to one. = 0). the linear polynomial for a response y is
(2)
232
. extends from the vertex it labels (x. the vertex labelled x I is the pure water "mixture" with x I = 1. The vertex represents the pure component. X 2 = 0. Typically. since much of this region does not represent feasible concrete mixtures.113. but other functional forms can be used as 'well. consider concrete as a mixture of three components: water (XI)' cement (x2). This example is included for illustrative purposes only. represents the volume fraction of a component. and x 3 = 0.0.0).113)and is called the centroid.
(1)
and the region defined by this constraint is the regular triangle (or simplex) shown in Figure 1.II. Experinrental regionfor !hreeco~onent~ure
A good experiment design for studying properties over the entire region of a threeXf component mixture would be the simplexcentroid design shown in Figure 2. Assume the coarsetofine aggregate ratio is held fixed. and aggregate (x3). and for three co~onent simplex~entrold the overall centroid. The coordinate where the three axes intersect is (113.EXPERIMENT Background on Mixture Experiments
DESIGN
As a simple (hypothetical) example of a mixture experiment.
Xf (l. = 1) to the midpoint of the opposite side of the triangle (x. polynomial functions are used for modeling.O.CI)
lC2(0. or (l. All properties of interest ~ure would be measured for each mix in the design and modeled as a function of the components.
represents the combined effects of all variables not included in the model.).60
~XJ
~X2~
s 0. The minimum and maximum levels of each component were chosen based on typical volume fractions for non airentrained concrete" with the constraint that the volume
Figure 3.70
In this case the simplex designs are generally no longer appropriate and other designs" are used.10 0. 0.where the b.15 0. Xl = cement. a meaningful subregion of the full simplex must be defined by constraining the component proportions. An example of a possible subregion for the three component example is shown in Figure 3. XJ = aggregate): .g. * are constants and e. the random error term. Example of constrained experimental region
233
. These designs typically include the extreme vertices of the constrained region and a subset of the remaining centroids (e. It is defined by tlle following volume fraction constraints (XI = water.25 0. This model is typically reparameterized in the form (3)
using bo*= bo *(xJ + Xl + xJ and is called the Scheffe Similarly the quadratic polynomial
9
linear mixture polynomial. Experiment Design for the SixComponent Study
)(1
Selection of proportions and constraints .20 ~X3~ 0.. centers of edges.
is reparameterized as y
= b1x1 + br2
+
br3 + b1r1X2 + b1r1X3 + b
2rr3 + e
(5)
Since feasible concrete mixes do not exist over the entire region shown in Figure 1. etc.The proportions for the sixcomponent mixture experiment were initially selected in terms of volume fraction and converted to weights for batching. faces.
Although air is incorporated into concrete during mixing. Since there are 21 coefficients in the model. These issues are addressed below.185 .The selection of an appropriate experiment design depends on several criteria. but these are not important for the small trial batches and can be adjusted later after a final mix is selected. the paste fraction of the concrete (water. the design must have at least 21 runs (21 distinct mixes) to estimate these coefficients.25
Maximum volume fraction . a single mix was replicated during each
234
. In addition to the individual constraints on each component. The "best" experiment design depends on the choice of an underlying model which will adequately explain the data. the following quadratic Scheffe polynomial was chosen as a reasonable model for each property as a function of the six components: (6)
This model is an extension of Equation 5 for the six component case. and HRWRA) was required to range from 25 to 35 percent by volume.2924
Xz
Microsilica HRWRA
Coarse Aggregate Fine Aggregate
xJ ~
Xs
X. Table 1. ability to provide an estimate of repeatability.16 . and five mixes were replicated to provide an estimate of repeatability allowing us to test the statistical significance of the fitted coefficients. microsilica. Experiment Design Details .0074 .027 . and ability to check the adequacy of the fitted model.
The six components and the final ranges of their volume fractions for this experiment are shown in Table 1. Finally. cement. In addition to the 21 required runs. For this experiment.fractions sum to unity. The volume fractions were converted to corresponding weights using the specific gravities and percent solids (where applicable) obtained from laboratory testing or from the material supplier. such as ability to estimate the underlying model.40 .Mixture Components and Volume Fraction Ranges
Component Water Cement ID
XI
Minimum volume fraction . seven additional runs (distinct mixes) were included to check the adequacy of the fitted model.13 .4424 .013 . Ignoring the air content as a mix component affects yield calculations.0046 . it is not an initial component and therefore was not considered to be a component of the mixture.15 .
the cylinder ends were ground parallel to meet the ASTM C39 requirements using an endgrinding machine designed for this purpose. Compressive strength tests (ASTM C39) were conducted on the cylinders at the ages of one day and 28 days. were prepared over a fourweek period. In all. Before testing.
235
. and ten 100 mm by 200 mm (4 in by 8 in) cylinders. Thirtynine batches of concrete. cylinders for concretes with slumps less than 50 mm (2 in) were vibrated on a vibrating table.0 W) capacity was used to mix the concrete. two fresh air content (ASTM C231) tests. from which 36 sets oftest results were analyzed. Table 2 summarizes the mixes used in the experiment. A total of39 batches were prepared. Three specimens (50 mm (2 in) thick slices taken from the middles of the concrete cylinders) were tested at an age of42 days. The first three mixes were repeated at the end of the program because an incorrect amount of water was used in batching them. A modifieddistance design" was chosen to ensure that the design selected could estimate the quadratic mixture model while spreading points as far away as possible from one another.
SPECIMEN FABRICATION AND TESTING
The materials used in this study included a Type IIll Portland cement. a total of 36 mixes were planned. In most cases.5 W) in volume. #57 crushed limestone coarse aggregate.week of the experiment to check statistical control of the fabrication and measurement process. tap water. Each batch included sufficient concrete for two slump tests.17 m3 (6. Commercially available computer software for experiment design was used for design and analysis of the experiment. otherwise. and a naphthalenesulfonate based superplasticizer (ASTM C494 Type FIG). natural sand. The test results from the incorrectly batched mixes were not included in the subsequent analysis. three cylinders were tested for each age. The program selected thirtysix points from a list of candidate points that is known to include the best points for fitting a quadratic polynomial. after which they were stripped and placed in limewaterfilled curing tanks for moist curing at 23 ± 2°C (73 ± 3 OF). The cylinders were covered with plastic and left in the molds for 22 hours. each approximately. A fourth test was performed in some cases if one result was significantly lower or higher than the others. the cylinders were rodded. two unit weight tests. The run order was randomized to reduce the effects of extraneous variables not explicitly included in the experiment. The three remaining cylinders from each batch were used for "rapid chloride" testing according to ASTM C1202. In order to obtain adequate consolidation. A rotatingdrum mixer with a 0. microsilica (in slurry form).04 m' (1. The cylinders were fabricated in accordance with ASTM C192.
3 126.3 141.3
(I)
Coarse aggregate {dry) (kg) 867.1 506.9 312.6
0.66 5.3 541.9 21.3 131.18.Table 2 .52 5.3 141..43 0.6 81O.34 0.l 810.1 829.9 354.9 21.33 0.59 4.59 5.3 122.43 0.39 28 27 31 25 29 5.66 4.81 oz.9 337.35 11 24 19 9 17 10 14 21 33 36 1·.1 810.37 20 122.9 21.33 0.8 530.9 337.8 537.3 506.52 4.3 563.52 5.22 6.7 853.3 122.41 0.3 141.9 21.1 810.2046 lb.9 45.9 45.4 33.
Design Run Order Water
of mix proportions (per cubic meter of concrete)
Cement Silica fume (kg) 45.1 312.4 21.6 526.3 122.38 8 13J34 4 16 30 3· 26.8 323.2 506.33 0.52 5.66 4.52 3.1 824.41 0.12. 1 kg = 2.66 3.8 131.1 312. lday strength.35 0.0 361.4 21.9 875.0 312.1 81O.3 532.52 3. along with the estimated cost per cubic yard of concrete..4 45.0 335.52 3.l 810.6 21. The cost of each batch was calculated from the mix
236
.66 5.8 526.8 312.9 810.4 122.4 122.66 3.1 361. r) indicates replicated mix c) indicates control mix • indicates mix which was repeated due to incorrect batchina
RESULTS AND ANALYSIS The average values for slump.38 0.36 0.4 829.9 21.35 0.8 122.1 810.9 857.9 21.33 0.4 21. 32.0 312.4 45.3 141.9 361.3 592.37
~otes: I 1= 33.9 549.59 3.35 0.6 122.4 122.35 0.9 21.32 0.1 513.35 0.52 3.4 21.5 126.3 506.6 (kg) 312.52 5.8
HRWRA
ID
(kg) 5(r) 11(r) 13 15 16 20{r) 22 28 37 38(r) 48 63 65 66 70 71(r) 78 87 89 91 98 101 103 110 116 123 127(c) 163 7.35 0.9 312.38 0.37 0.6 845.3 506.2 548.37 0.8 506.9 506..3 506.9 521.1 361.0 810.Sum"".0 341.52 3.3 561.3 130.66 5.52 3.9 45.9 312.3 810.1 312. rapid chloride test measurement (coulombs) for each batch are shown in Table 3.59 5.66 3.6 841.0 131.9 337.4 122.8 122.4 45.1 834.1 895. 28day strength.37 0.52 5.9 337.8 852.1 361.38 0.3 529.4 21.66 5.9 45.9 21.1 810.9 312.9 21.9 45.3 122.8 131.4 141.35 0.33 0.9 21.3 130.3 506.38 0.9 810.9 27.43 0.2 506.8 122.3 135.9 361.1 506.12
506.1 312.52 3.0 810.2 571.23 15 2·.5
Fine aggregate (dry)
(kg)
w/(c+st)
3.66 3.7 836.9 312.4 45.9 337.3 122.0 532.7 832.1 836.
1
"
28day s1r (MPa) 48.32 118.6 24. 1 m'
=
237
.0 51.67 123.1 57.2 56.99 91.0 51.67 94.32 102.6 58.99 97.5 1.2 58.0 54.2 50.4 53.42 126.7 23.2 25.6 51.6 19.8 53.5 18.2 48.308 yd1
42dayRCT (coulombs) 1278 862 1162 387 776 1027 744 492 842 903 583 684 292 604 847 720 554 792 348 968 700 316 390 302 682 505 245 310 636 356 820 553 340 640 239 332
Cost ($/ m') 95.3 60.6 28.2 54.8 22.0394 m.5 27.20 124.2 22.4 23.68 120.2 21.67 128.85 91.8 22.5 27.8 26.2 28.2 53.2 27.1 54.22 91.56 96.85 92.1 54.24 118.67 91.32 99.3 27.80 96.7 23.8 26.9 54.0 21.4 21.20 96.42 92.85 94.3 61.4 21.0 16.0 59.44 120.32 92.41 96..8 22.41 103.85 123.7 50.2 51.4 16.8 53.89 96.1 54.9 54.1 48.22 110.6 65.6 16.7 23.34 99.6 53.Table 3 .65 126.9 18.5 22.85
67 57 102 13 35 200 22 127 99 118
64
57 29 32 92 38 95 51 25 114 67 76 29 124 171 51 35 32 121 114 127 108 99 102 51
24.6 51.0 27.24 123.6 145 psi.5 48.53 96.2 55.8
=
39 25 .3 22.1 60.5 53.5 56.Test Results and Costs
Design
Run 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38
Slump
(mm)
ID
22 71 II 5 16 91 101 78 127 20 103 13 28 98 127 89 163 110 5 II 87 66 38 63 48 70 37 65 127 116 20 71 123 127 15 38 Note: 1 mm
lday s1r (MPa) 21.3 21.9 24.65 96.2 55.20 118.18 102. 1 MPa
=
23.
062E+05 51. This is assessed using analysis of variance (ANOVA).19 147. any mixture would give the same response). a linear model may provide a better fit to the data.50 9.e. we conclude that quadratic terms should not be included in the model.84 2965. fit by least squares.05.. The analyses for the other properties was performed in a similar manner. The ANOVA results for 28day strength are shown in Table 46.01 9.5665
IS
36
coefficients of the linear terms are equal. and interpreting the model graphically using contour and trace plots. The resulting linear model for 28day strength (y I). we conclude that linear terms should be included in the model. The statistical analysis is described in detail for 28day strength.36xs + 87.81x3 + 1972x4 + 38.
238
. Since the Prob > Fvalue is 0.proportions using approximate costs for each component material obtained from a local readymix concrete producer. Since the Prob > Fvalue of 0.062E+05 257.0011. is YI = 45. We conclude that the coefficients differ for low values (say less than 0.19x6
(7)
with residual standard deviation s = 3. Each of the four responses was analyzed by fitting a model. this means that mixing does not affect response (i.37
F Value
Prob > F
5.22x1 + 89. In the absence of quadratic terms.92
0. The row with source linear tests whether the
Table 4 ANOVA
Source Mean Linear Quadratic Residual Total Sum of Squares
l.52 135.
Model Identification and Validation
The first step in the analysis is to identify a plausible model.07 MPa (445.7. Even though the design selected permits estimation of a quadratic model.5667 exceeds 0.0011 0.4 psi).068E+05 5 1~
Table/or 28day strength
Mean
DOF
Square
1.15x2

3. validating the model (by examining the residuals for trends and outliers).46 0.62 1. The row with source quadratic tests whether any quadratic coefficients differ from zero.05) of the Prob > F (also called the pvalue).
39 MPa (492. in the model. This figure consists of six overlaid plots. random residuals.01
DtMsIion from Referenoe Blend.00
0. should exhibit similar properties. a plot of residuals versus run sequence as well as a plot of the control mix data revealed a linear trend in the data for each response. voIlIT78 fraction
0. here the centroid.
'i'
'I'
Traa1 Pkt (Cox)
Graphical Interpretation
Once a valid model is obtained. For a given component the fitted value of the response is plotted as the component is varied from its low to high setting in the constrained region. residual plots will often suggest how to modify the model to remove the structure. A response trace plot is shown in Figure 4. A value of s near the repeatability value (replicate standard deviation) is an indication of an adequately fitting model. an adequate model should capture all information in the data leaving structureless. one for each component.~
0. while the other components are held in the same relative ratio as a specified reference mixture. However.02
0. The repeatability value is 3.The residual standard deviation s is defined as
np
(8)
where the number of observations n = 36 and the number of parameters in the fitted model p = 6. Trace plot/or 28day strength
239
. Essentially. The residuals are the deviations of the observed data from the fitted values. The fitted model is then validated by examining residual plots. In this study. estimates the error term e.04
Figure 4. The residual j.j'. The plot shows the "effect" of changing each component on 28day strength. which estimate these errors.03
0. The residuals. If structure remains in the residuals. The e/s are assumed to be random and normally distributed with mean 0 and constant standard deviation. which is close to s. it can be interpreted graphically using response trace plots and contour plots.
58
50
. because the run sequence was randomized.01
0.3 psi). this trend had little impact on the fitted models.
with the other components fixed at their centroid values. and microsilica.As expected. cement. with a predicted value of strength of 59. and HRWRA.40. several must be examined. and fine aggregate = 0. In each case. confirming the result from the response trace plot. HRWRA = 0. Figure 6 shows a contour plot of 28day strength in water. The best settings (expressed as volume fractions) are water = 0. high cement.53 MPa (8634 psi). cement and microsilica (HRWRA = .410. coarse aggregate. microsilica = 0. and HRWRA. and fine aggregrate. cement. and high fine aggregate.0074. Contour plot/or 2Bday strength (MPa) in water. These plots show that strength increases for low water. Therefore. while increasing the amount of cement increased strength. HRWRA will be set at its high value. in subsequent
28daystr
mIc:roIIIca
Figure 5. CA =. FA =. low microsilica. CA =. Figure 5 is a contour plot of 28day strength for water.013. FA =.15. This apparent reduction may not be significant when compared to the underlying experimental error.16. and Figure 7 shows a contour plot of 28day strength in water.410. coarse aggregate. This may be due to the improved dispersion of the cement and silica fume caused by higher amounts ofHR WRA. Contour plot/or 2BdIIy strength (MPa) in water. HRWRA had the largest effect with higher amounts of HRWRA yielding higher strength. Contour plots are used to identify conditions which give maximum (or minimum) response.259)
contour plots.259)
Figure 6. cement = 0. low coarse aggregate.0074. and fine aggregate at the best settings of water. HRWRA is fixed at its high value and the other components are fixed at the centroid settings. increasing the amount of water decreased strength. cement. Because contour plots can only show three components at a time (the others components are set at fixed conditions). cement and HRWRA (microsilica =.01B. Surprisingly. The plot indicates that strength increases rapidly by increasing HR WRA.27. coarse aggregate = 0. an increase in silica fume appears to reduce strength. The best overall settings can be found using the contour plot shown in Figure 8 for microsilica.
240
.
1376. since residual plots showed that the standard deviation of 42day ReT was proportional to the mean.
(9)
Y3
= I.):
< •.16. for which the fitted model includes four quadratic terms which were found to be significant. cement=.328E+06x~4+6.118E+05xr6+6.1. HRWRA = . the following models were fit to slump (Y2)' lday strength (Y3).481E+06x. and coane aggregate (water =. 732E +06x4 +8632xs 15245x6 + 6. and 42day ReT results (y. The natural logarithm of 42day ReT was used for modeling.lS.
241
. microsilica = .
752E +05x1 +2. Contour plotfor 28day strength (MPa) ill microsilica.573E +05x1 I 0723X3 I.fme aggregate.0074)
Figure B.01B. Contour plotfor 2Bday strength (MPa) in water.HRWRA =.8701E01
waw
Figure 7..0074)
Models for Other Responses Using the same procedure described above for 28day strength.107E+05xlx68.x6
(10)
(11)
Linear models were adequate for all responses except the lday strength. C4 and FA (cement = .
287. The desirability function ~ takes on values between 0 and 1. Therefore. Desu:a~ilUJ: unctions for f (3200 psi) and I above 22.404. Since cost is to be minimized.94 per cubic meter.. resulting in desirability functions with values of 1 above the minimum or below the maximum and zero otherwise. and 12. HRWRA = 0. cement = 0. respectively.. nonlinear function instead of linear for cost). as I. for lday strength the desirability value is 0 below 22. with a linear decrease in desirability to a value of'zero at the lower and upper specifications (see Figure 9).48 MPa (5000 psi) the desirability becomes 0. Desirabilities for 28day strength and 42day ReT are defined similarly. at a cost of $92.160. is water = 0.7 200 91Ja 128. It is also possible to develop more complex desirability functions (e. expressed in volume fractions. but the most desirable value is the midpoint of this range.. the mix which maximizes D.. Minimum and maximum specifications are used for strength and ReT. Numerical optimization using desirability functions' can be used to find the optimum mix. Iday strength = 22. over the feasible region of mixtures. The response values for this mix are slump = 75 mm (3 in).Selection of Optimum Mix
The optimum concrete mix is defined here as that mix which minimizes cost while meeting the specifications. coarse aggregate = 0.06 MPa (3200 optimization psi).
fitted models:
(12)
Based on the experimental results. First. however this strength was chosen to be well beyond the maximum value in the observed data.g.J indicated in Figure 9.
In the numerical optimization scheme.130.e8 219 1162 .013. or 75 mm (3 in). the optimum mix maximizes the geometric mean D of the individual desirability functions d. For example. the desirability function for cost decreases linearly over the range of costs observed in the data (see Figure 9). the maximum desiqlbility is given to the target value of75 mm (3 in).62 MPa (7922 psi)..00493.06 MPa Figure 9. For slump a range of 50 to 100 mm (2 to 4 in) was specified.06 MPa (3200 psi).
242
. a desirability function must be defined for 1'5 each property. using the . and fine aggregate = 0. microsilica = 0. At 34. and 42daY ReT value = 653 coulombs.2cIor IICT may be defined in several ways. 28daY strength = 54.
00533. CONCLUSIONS AND RECOMMENDATIONS
In high performance concretes consisting of many components.160. and 42day ReT = 617 ± 81 coulombs. the number of experimental runs can be halved.0131.97 MPa (3200 ± 140 psi).135. The best mix for this new set of constraints (expressed as volume fractions) is water = 0. 28day target strength> 53. 21 mixes are required to fit a quadratic model. The predicted values and 95% uncertainties for the remaining responses at the current best mix are slump = 75 ± 15 rnrn (3. Fitted models are obtained from the experimental data and are used to identify optimal mixes over the region. and 42day ReT < 620 coulombs.72 MFa (8047 ± 394 psi).285 at a cost of$72. For the materials and conditions of this experiment. which depends on the location of the mix in the feasible region. The uncertainties in the properties of the current mix can be used to modify the constraints and identify a revised optimal mix for these new constraints. 28day strength = 55. lday target strength> 23. the analysis would be complete. 23..09 ± 0. i. where several properties are of interest.06 MPa (3200 psi).6 in).If the fitted functions for each property were known without error.54. A minimum of 31 runs is recommended. HRWRA = 0. However.03) contains the true lday target strength for this mix. if a linear model is found to be adequate for all responses of interest. Therefore.06 ± 0.6 in). cement = 0. quadratic models are assumed to provide an adequate representation of each property over the region of interest. The uncertainty provided is for a 95% confidence interval. it is critical to use a systematic approach for identifying optimal mixes given a set of constraints. although additional runs should be included for checking the adequacy of the fitted model and estimating repeatability. there is uncertainty in the fitted functions since they are estimated from a sample of data. 28day strength = 54. The predicted values and 95% uncertainties for this mix are slump = 75 ± 15 rnrn (3 ± 0. microsilica = 0.77 MPa (3349 ± 112 psi). However. Statististical experiment design and mixture experiments provide such an approach.78 MPa (7800 psi). They permit a thorough examination of a feasible region of interest in which to identify optimal mixes.62 ± 2. The revised mix must then be checked to see that the specifications are met.6 in < slump < 3. lday strength = 23. The modified constraints on the responses which take into account the uncertainties are 66 rnrn < slump < 86 rnrn (2. a linear model was adequate for all but one response (Iday strength). For a sixcomponent mixture.99 MPa (7922 ± 434 psi). The lower or upper bound values (as appropriate) for all responses meet the specifications.4 in).03 MPa (3340 psi). coarse aggregate = 0. it is quite possible that the true lday target strength would fall below 22. at the current mix the predicted lday target strength is 22. So if this mix is used. Typically.0 ± 0. and 28day ReT = 653 ± 81 coulombs. and fine aggregate = 0. Since materials and conditions will vary by location. each specification must be modified to account for the uncertainty in the fitted function. For example. we are 95% confident that the interval (21.
243
.401. the quadratic model should be considered initially.48 ± 2.e.09.
P. 1990. Boston: Irwin. "Standard Practice for Selecting Proportions for Normal. Cornell. "Optimization of the Composition ofa HighPerformance Concrete. H. McLean.. New York: Wiley. trial and error methods could easily miss the optimal conditions...
11. Illinois: Portland Cement Association. 447454.
6. Detroit: American Concrete Institute. 3rd edition. Hunter. D." Technometrics. V.. Scheffe.K. Wasserman.
2. & Kutner.A. P. May.2. No.P. Volume 1." Cement. V.
244
.Extra care is required to run a designed experiment. "Guide for Selecting Proportions for HighStrength Concrete with Portland Cement and Fly Ash. Panarese. Volume 1. ACI Committee 211. and Aitcin. and Mass Concrete. 1958.. Applied Linear Statistical Models. pp.H.
9. 1995.
REFERENCES
1. Statistics for EXperimenters. Response Surface Methodology: Process and Product Optimization Using Designed Experiments. and Anderson. 2nd ed. pp. S. and W.. the results are well worth the additional effort..H.H. New York: Wiley. "Experiments with Mixtures. J. "A Novel Approach to Developing HighStrength Concrete. pp. Thirteenth Edition. & Hunter.. 115124. J. Experiments with Mixtures: Designs." InACI Manual of Concrete Practice.
8. 1994. 20. resulting in higher costs to producers over the longterm.L. B.. pp. V.A. Kosmatka. M." InACI Manual of Concrete Practice. & Montgomery.. R.
5. December. however. Models.J. 1978. W. ACI Committee 211.'
r
7.
10. Box. 1988.. "Extreme Vertices Design of Mixture Experiments. Skokie. Detroit: American Concrete Institute. 2529. Luciano. 1991.C.R DelGado. V. 8. G. Myers.C. 16. Concrete and Aggregates. J. R. With many components and several properties of interest.
4.
Neter." Concrete International. New York: Wiley. and the Analysis Mixture Data. C..
3.S. Design and Control of Concrete Mixtures. Heavyweight. 1995. 1995. Nmai.C. and J." Journal of the Royal Statistical Society. J. Rougeron.
0/
. 1990.E. 344360.