We used SimCentral26 simulation software. Robustness
is important when the results are obtained by inputting random data into
the process simulator. With SimCentral, we receive an error flag and can
discard the results when the solution is divergent and cannot be
obtained. In an iterative process, a sequential simulator takes a long
time to converge; but a convergent solution can be obtained in a short
time with SimCcntral. In addition, X can be set as needed in ADoE
because the specifications can be freely changed. Furthermore,
SimCentral can be performed using JavaScript and can be automated by
association with Python programs, which means that the proposed method
can be easily combined with SimCentral.
We succeeded in searching for X that achieved target ranges of Y using
139 simulations (D-optimal design required 50, and Bayesian optimization
required 89). As a comparison method, X candidates were selected in a
random search. In this case, the target ranges of the Y could not be
achieved at all. Table 3 shows the result of 139 simulations, and Figure
4 shows the behavior of Y for 139 simulations. In Bayesian optimization,
the number of samples within the target range of Y increased with the
number of trials. The accuracy of the GP model improved as the number of
samples increased with the number of trials. The Y8result, which has a narrow range, has a low probability and is not
considered by P all. Therefore, it was possible to
treat each variable equally by performing range scaling, as shown in Eq
(13).
Table 3. The results of Y that achieved targets