Using statistically designed experiments
to improve investment casting quality
Although it’s been practiced for centuries, investment
casting is still regarded by many as part science, part
art, and part black magic. This is due in large part to
the unpredictable behavior of precious metals in the molten
state. Much of the data available about lost-wax casting
is based on trial-and-error experimentation, not scientific
procedure.
In today’s competitive business environment, unless you
get lucky with trial-and-error experimentation and hit
an answer, you could find yourself out of business quickly.
Instead, it would be ideal to find a scientific yet relatively
simple way to problem solve your casting issues without
very expensive hardware, software, or rocket science.
Fortunately, methods that fit the bill have been around
for a long time, and they’ve been used in many other industries
with great success. One of these methods is statistically
designed experiments, also called Design of Experiments
(DOE). This fairly simple tool can be used to quickly zero
in on important parameters for casting, while weeding out
the trivial ones. It prevents time wasted working with
variables that you may think are important, but in reality
don’t affect your results. More important, this method
allows you to see how different factors may interact with
each other to influence results, which is very difficult
to determine through observation alone.
Background on DOE
The classical view of experimentation is to change one factor
at a time, run the experiment, observe the results, and move
on to the next factor. This approach has several drawbacks:
It takes much longer and uses up more resources; the optimum
combination of all variables may never be revealed; and the
interaction between factors may never be revealed.
Unlike classical experimentation, statistically designed
experiments can provide the following benefits:
• Many factors can be examined simultaneously.
• Some input factors that cannot be controlled, which
are called noise factors, can influence the output; however,
other input factors can then be controlled to reduce the
effect of the “noise.”
• In-depth statistical knowledge is not necessary to perform
these tests and reap their benefits.
• Relatively few experiments can look at a large number
of factors and separate the trivial from the important.
• In most cases, quality and reliability can be improved
without increasing costs.
Although experimental design was developed in England
and the U.S. many years ago, it was aimed at statisticians,
not engineers. The latter were not exposed to it until
a Japanese engineer named Gen’ichi Taguchi took the method
and adapted it for use by Japan’s engineering community.
In the 1980s, as Japan was fast becoming a world economic
leader, Americans noticed the Japanese using a powerful
new experimental tool: the Taguchi method. Today this method
is used in many industries to shorten product development
lifecycles while at the same time producing higher quality,
lower cost products. In the jewelry manufacturing industry,
it can be applied to casting in order to determine the
proper parameters needed to produce good, consistent castings.
It also enables manufacturers to assess what factors are
important in ultimately determining how to produce high
quality castings.
Setting Up an Experiment
In the sample experiment presented in this article, my objective
was to examine what factors would influence the filling of
a fine piece when casting platinum on my centrifugal caster—a
Galloni automatic induction, double-broken-arm machine. The
fine piece that I cast was a 10 square by 10 square grid made
of 0.05 mm diameter wax wire (figure 1). It was gated along
one edge with a main gate in the middle.
After establishing an objective, the first step of any
DOE is to determine what factors you want to examine. To
assess influences on fill, I felt the obvious choices were
oven temperature and metal temperature because they are
two of the most important factors in casting. Since I can
control acceleration and final velocity on my casting machine,
I chose them as the other two factors, leaving me with
four factors to examine.
Typically, to keep the experiment from getting out of
hand, I look at two levels for each of those factors, a
high level and a low level. It is useful to pick levels
that are within the realm of your machine and process capability.
For a four-factor, two-level experiment, you will need
to run 16 experiments. This is typically referred to as
a 2 by 4 factorial.
Once you’ve determined what factors you want to examine,
you must set the matrix (figure 2). The pluses represent
the higher level and the minuses represent the lower level.
As you can see, there is a pattern to the pluses and minuses
to ensure that every combination of factors is covered
at every level.
This matrix will provide the main factors. If you want
to explore the interaction between two factors, you can
expand this matrix (figure 3).
There are two additional columns in figure 3. Column five
examines the interaction between velocity and acceleration,
while column six examines the interaction between metal
temperature and oven temperature. A result that shows a
strong interaction indicates that you can’t consider one
factor without considering the influence of the other.
If there is no interaction, the factors are independent
of each other.
To set up the additional columns, use basic math to multiply
the pluses and minuses in the columns of interest. A plus
multiplied by a plus equals a plus, as in run number four
for velocity x acceleration. A minus multiplied by a minus
equals a plus, as in run number one for velocity x acceleration.
A plus multiplied by a minus equals a minus, as in run
number two for velocity x acceleration. Once again, you
can see a pattern developing.
Plotting the Numbers
When the high and low levels are plotted in the
matrix, you need to fill in the actual numbers. Since my
Galloni casting machine provides a velocity choice of 200
to 500 rpm, I make these levels my plus and minus.
It’s important to note here that the setup and results
of your experiment will be specific to the equipment you
are using. For example, some centrifugal casting machines
don’t measure velocity in rpm like mine does. In this scenario,
you could count the winds of the spring, using three winds
as the low level and six winds as the high level, for example.
Or, consider my treatment of the acceleration factor.
My casting machine allows an acceleration of 1 as the
fastest and 4 as the slowest. Although these are simply
levels of acceleration, not specific measurements, I can
still plug them into the matrix and use them for my calculations.
(Note: In this circumstance, you must be careful when working
with software, as it will be looking for the low level
number to be less than the high level number. You can plug
in a -4 to get around that.)
For a metal temperature, I wanted to look at two extremes.
I initially chose 1,800°C (3,272°F) and 1,975°C (3,587°F);
however, early trials showed that at 1,800°C (3,272°F)
I couldn’t get the metal to leave the crucible, so I changed
the low temperature to 1,825°C (3,317°F).
Lastly, I set the oven temperatures to 1,550°F (843°C)
and 1,450°F (788°C). When casting, I usually set the oven
temperature as my last factor, so I can run all the low
temperature trials first, followed by the high temperature
trials. Although this is not the preferred method—random
testing provides more accurate results—it saves a lot of
time if you are using only one burnout oven. If you have
the luxury of using two ovens, this is not an issue.
When you plug all the numbers into the matrix, it will
resemble the one shown in figure 4.
Additional Factors
In addition to the factors in question, you can examine potential
“noise” factors. For example, does the position of the object
on the button influence fill? Logic would dictate that as the
mold travels in a counterclockwise motion, the metal as it
leaves the crucible will want to stay behind, traveling in
the opposite direction (figure 5).
You could hypothesize that pieces at the back end of the
swing would see more force and fill better than pieces
at the front end of the swing. You could also hypothesize
that pieces at the top and bottom of the button would fill
equally. If this is really true, you can treat the position
of the parts as a noise factor and try to adjust other
factors to minimize or eliminate this effect. This is where
DOE becomes particularly powerful and leads to robust processes
and design.
For my experiment, the part was located at four different
90 degree positions on the button to test whether position
played a role, and if its effect could be minimized by
altering the other factors. I placed grids at the 12, 3,
6, and 9 o’clock positions. I was careful to ensure that
the 12 o’clock position was always located in the same
location for each flask, so as not to introduce a variable.
To do this, I marked the part at the 12 o’clock position
and aligned it with a mark on the flask (figure 6).
Once the parts were sprued, the flasks were invested using
the same batch of investment. I ensured that all parameters
were kept the same for each batch, since I could not invest
all the flasks with one mix cycle.
After drying, the flasks were sent through the burnout
cycle following the normal production process. The flasks
were cast, devested, and pickled in caustic soda to remove
investment. Each tree was en-graved with the flask number
after casting to preserve the order of the casts.
To obtain results, I counted how many squares were completely
filled in each grid at each position. I then plugged in
the results on the spreadsheet in preparation for analysis
(figure 7).
Analyzing the Results
Without doing any fancy math, one conclusion I can draw from
the results is that I can get complete fill in all positions
at the higher metal temperature (1,975°C), regardless of the
levels of the other factors. This is an obvious effect. What
is difficult to figure out is the possible subtle effects of
the other factors. This is where the math enters, which is
rather straightforward.
To find out the influence of a factor, take the results
from all the low level runs of that factor, add them together,
and subtract that number from the sum of the results from
all the high level runs of that factor. Then take that
result and divide it by the number of runs at the high
level, in this case eight.
For example, take the influence of the velocity factor
at the 12 o’clock position. The calculation is:
(86+97+100+100+64+88+100+100)-(76+28+100+100+71+59+100+100)
/8 = 12.625
This result shows me that running the experiment at the
higher velocity level (500 rpm) increased my fill rate
by 12.6 percent. This could be a substantially significant
influence.
You can run these formulas for all factors at all positions.
Doing so will give you the results shown in figure 8. Here
are some of my conclusions for the three remaining factors:
Metal temperature. What you can see from the results,
which confirm my earlier observations of the raw data,
is that metal temperature has the highest influence on
fill. You can also see that the influence at the 12 and
6 o’clock positions is almost the same, and that the influence
is highest at the 3 o’clock position, which is on the back
end of the swing. So while metal temperature had the greatest
influence, the amount of that influence was driven by flask
position.
Oven temperature. If you look at the oven temperature
influence, you can see that the numbers are low, and negative
in the case of the 12 and 3 o’clock positions. The only
number of any significance is at the 9 o’clock position.
This could be indicating that this position is the last
to fill, and with the higher oven temperature, the metal
is still molten enough to do some filling after the other
positions are filled.
Acceleration. The acceleration influence is very interesting,
as it seems to show little influence at the 3 and 9 o’clock
positions, but a negative influence at the 12 o’clock position,
meaning a higher acceleration produces less fill at that
position. The spread between the 12 and 6 o’clock positions
is almost 17 percent. To explain this, you might theorize
that the metal is not really leaving the crucible at a
true horizontal, but has a downward motion to it. Increasing
the acceleration increases the downward angle of the metal
stream (or the mold is actually lifting off the cradle
from the kick), causing more filling at the 6 o’clock position.
Another theory is that just the opposite is occurring:
The stream is riding high, hitting the upper edge of the
opening and causing less fill at the 12 o’clock position.
You could theorize about what is happening, but the important
thing to remember is that something is occurring that could
be influencing your fill characteristics—something that
you wouldn’t be able to see without this experimenting
tool.
Graphing Your Results
You can also graph your results, as visuals tend
to show influencing factors more clearly. This is where
statistical software, such as Minitab, can be very useful.
Minitab creates a factorial grid for you. You start by
inputing your factors and levels. After the experiment,
you input your results and the software calculates all
the graphs showing the influences. Figures 9 and 10 show
my results graphed by this software.
As you can see, the graphs show the influences very clearly.
On the Main Effects Plot (figure 9), a horizontal line
indicates no influence from the factor at the tested levels.
This can be seen in the graph for oven temperature. The
steeper the plotted line, the greater the influence of
that factor, as is the case with metal temperature.
For the Interaction Plot (figure 10), crisscrossing lines
indicate that the two factors influence each other. If
the lines run parallel to each other, or don’t cross between
the two levels, then they don’t influence each other.
If you look at oven temperature versus metal temperature,
you see no interaction: I achieved 100 percent fill at
the 1,975°C metal temperature, regardless of whether the
oven was at 1,450°F or 1,550°F. If you look at oven temperature
versus acceleration, on the other hand, you see that at
the higher acceleration (1), increasing oven temperature
increased fill. However, at the lower acceleration (4),
increasing oven temperature actually decreased fill, a
phenomena that is difficult to explain and would benefit
from replication of the experiment.
One thing that you don’t want to do is extrapolate past
the levels examined in the experiment, which would result
in false conclusions. These results are only valid for
these levels.
For the final graph (figure 11), I added the results of
all the positions to get a total amount of fill for each
run. This was then plotted to see if anything changed.
When compared with the Main Effects Plot for the 12 o’clock
position, you can see two changes. There is now a slight
influence from oven temperature, though not much, and a
reverse influence from acceleration. (You could theorize
that increasing the acceleration has a diminishing effect
at some point due to the creation of excess turbulence,
which impedes filling.) You can also compare the total
to the other three positions to see what changes are evident.
Conclusions
The power of using statistically designed experiments is that
from one set of experiments you can gather a wealth of data
that otherwise would be difficult to collect and analyze. For
example, this experiment showed me that oven temperature is
not as critical as one might think in filling a fine piece.
If this is the case, there are several advantages to staying
at the lower oven temperature: less wear and tear on the equipment
and materials, better cast surface due to reduced reaction
with investment, easier investment re-moval, and shorter process
time. These factors combined would result in a higher quality
piece at lower cost. This type of data allows jewelry manufacturers
to make informed decisions regarding our casting processes.
Its usefulness has been proven in many industries—and now it’s
time to utilize it in ours. In today’s global environment,
it is a proven tool for staying competitive.
Metallurgist Tino Volpe is the technical manager for Tiffany & Co.
in Cumberland, Rhode Island.
Experimenting Tips
The following tips will help you set up an experiment
that will yield quantitative data and scientific results:
Think about what factors you want to study. In
the experiment presented in this article, I examined how
system temperatures and centrifugal forces influence fill,
but there are many other factors that you could consider
in a casting DOE. Tree design, mold permeability, and straight
arm casting versus broken arm casting are interesting potential
factors.
Measure results quantitatively. Qualitative
or subjective results such as “looks good” won’t fit the
formulas very well. You could have results such as the
number of pieces rejected or accepted, but if the judge
is human, your data could be tainted.
All subjectivity can be avoided if you set up the experiment
to produce measurable results. In this case, I used a 10
square by 10 square grid to measure fill because it is
easy to count how many squares filled, which can be represented
as a percentage.
Watch your constants. It's essential
that all factors other than those you are studying are
kept constant throughout the experimental run, so as not
to influence the results.