Vous êtes sur la page 1sur 4

Chapter 9 The clich image of experimentation is a crazed person in a lab coat pouring fluorescent liquids into bubbling beakers.

Ironically, the standard approaches for DOE do not work very well on experiments that involve mixtures. For example, what would happen if you performed a factorial design on combination of lemonade and apple juice? Table 9-1 shows the experimental layout with two levels of each factor, either one or two cups. Notice that standard orders 1 and 4 call for mixtures with the same ratio of lemonade to apple juice. The total amount varies, but will have no effect on responses such as taste, color, or viscosity. Therefore, it makes no sense to do the complete design. When responses depend only on proportions and not the amount of ingredients, factorial design do not work very well.

Shotgun approach to mixture design Chemists are famous for creating mysterious concoctions seemingly by magic. A typical example is Hoppe is Nitro Powder Solvent Number 9, invented by Frank August Hoppe. While fighting in the Spanish-American War, Captain Hoppe found it extremely difficult to ream corrosion from his gun barrel. The problem was aggravated by antagonistic effects from mixing old black powder with new smokeless powder. After several years of experimenting in his shed, Hoppe came up with a mixture of nine chemicals that worked very effectively. A century or so later, his cleaning solution is still sold. The composition remains a trade secret. Another approach to this problem is to take the variable of amount out of the experiment and work on a percentage basis. Table 9-2 shows the layout for a second attempt at the juice experiment, with each component at two levels: 0 or 100 percent. This design obviously do is not work either, because it asks for impossible totals. It illustrates second aspects of mixtures: the total is constrained in that the ingredients must add up to 100 percent.

Two-Component Mixture Design: Good as Gold Several thousand years ago, a jewelry maker discovered that the addition of copper to gold reduces the melt point of the resulting mixture. This led to a breakthrough in goldsmithing, because small decorations could be soldered with a copper-rich amalgam to a main element of pure gold. The copper blended in with no noticeable loss in luster of the finished piece.

Table 9-3 lays out a simple mixture experiment aimed at quantifying this metallurgical phenomenon. This is a fully replicated design on the pure metals plus the binary (50/50) blend. To do a proper replication, each blend must be reformulated, not just retested for melt point. The latter approach would cause error to be underestimated, because the only variation would be that due to testing, not the entire process. Also, the run order must be randomized. Do not do the same formulation twice in a row, because you will more than likely get results that reflect less than the normal process variation. Notice the depression in melt point in the blend. This is desirable and therefore an example of synergistic behavior. It can be modeled with the following equation, called a mixture model: Melt point = 1034.0A + 1073B 553.2AB Where components A and B are expressed in proportional scale (0 at 1). This second-order model, developed by Henri Scheffe specifically for mixtures, is easy to interpret. The coefficients for the main effects are the responses for the purest blends for A and B. the negative coefficient on the

interaction term (AB) indicates that a combination of the two components produces a response that is less that what you would expect from linear blending. This unexpected interaction becomes more obvious in the response surface graph shown in figure9-1. For purpose of illustration, we`ve shown a simplified view of the actual behavior of copper an gold mixtures. The predicted value for equal amounts of gold and copper is: Melt point = 1043.0 (0.5) + 1073.5 (0.5) 553.2(0.5*0.5) = (1043.0 + 1073.5)/2 553.2/4 = 1058.25 138.3 = 919.95 The equation predicts a deflection of 138.3 degrees C from the expected melt point of 1058.25. statistical analysis of the overall model and the interaction itself, shown in table 9-4, reveals a significant fit. Mixture problems require special treatment for computing probabilities. Specifically, the main effects (A and B in this case) cannot be independently estimated. For example, if you specify 40 percent of gold, then the copper must be 60 percent to make up the differences to the total of 100 percent. Therefore, the main effects for the mixture are collected in one test labelled linear in the ANOVA. Only one degree of freedom is available because the levels of gold and copper are completely interdependent, when the proportion of one component changes, the proportion of the other one does also. No matter how many components you include in a mixture design, the last one will always be fixed after setting the levels on all the others. The ANOVA in this case lists pure error rather than residual, because the estimate for error comes from replicated blends. Each pair of results contributes one degree of freedom for pure error, so there are three degrees of freedom in all.

Worth it is weight in gold?

An ancient king suspected that his goldsmith mixed some silver into a supposedly pure gold crown. He asked the famous mathematician Archimedes to investigate Archimedes performed the following experiment: 1 create a bar of pure gold with same weight as the crown 2 put the gold in a bath tub. Measure the volume of water spilled. 3 Do the same with the crown 4 compare the volumes. Archimedes knew that silver would be less dense than gold. Therefore upon finding that the volume of the crown exceeded the volume of an equal weight of gold, he knew that the crown contained silver. According to legend, Archimedes then ran naked from his bath into the street shouting Eureka!, which in Greek means I have found it. The principles of mixture design can be put to work in this case. Gold and silver have d

Vous aimerez peut-être aussi