As opposed to the ``ideal'' world of computer simulation, where
numbers are exact, the real fabrication process has to deal with
fluctuations and variations in all parameters. Therefore, for a given
process and given process parameters, the measured electrical
characteristics of the fabricated devices vary within certain
ranges.
These variations are modeled as appearing at the input of the
fabrication units, e.g., a statistical variation is assigned to
a process parameter, with the system itself assumed to behave
completely deterministic.
The parameters of these statistical variations
can be obtained from long-term measurements in the fabrication unit.
Given a model of the fabrication process - either a sequence of
process simulators of a global model -, a large number of
experiments can be generated from the nominal values and the
statistical parameters
.
In order to optimize the yield of the
product, the fraction of experiments that leads to results that
fall into the manufacture acceptance
window has to be maximized.
By adjusting the nominal process parameter values, the design is
centered with respect to the output distribution.
In practical applications, the system is represented by a response
surface model, and experiments are generated by using the Monte Carlo
method
.
Figure 2.5 shows the experimental loop used for design
centering.