3.2. Validation of the calibration models
In order to see if the calibration models are able to predict the
ethanol concentration during new process runs, each calibration model
obtained from a cultivation run was validated with the data from the
other two cultivations which have different initial concentrations. Fig.
7 (a) and Fig.7 (b) presents the predicted (using data from BC1 for the
calibration model) as well as the off-line measurements of ethanol
concentration as a function of time during BC2 and BC3 respectively.
Fig. 7 (c) and Fig.7 (d) presents the predicted (using data from BC2 for
the calibration model) as well as the off-line measurements of ethanol
concentration as a function of time during BC1 and BC3 respectively.
Fig. 7 (e) and Fig.7 (f) presents the predicted (using data from BC3 for
the calibration model) as well as the off-line measurements of ethanol
concentration as a function of time during BC1 and BC2 respectively.
Fig. 7 indicates that the predicted ethanol concentration using the MBC
approach corresponds well with the off-line measurements during all
cultivations. Furthermore, the ethanol production phase and ethanol
consumption phase are clearly indicated by the predicted ethanol values
without a significant time delay (in comparison with the off-line
values). In S. cerevisiae batch cultivation, the metabolic shift
of the yeast cells (shifting from ethanol production to ethanol
consumption) is a critical point which indicates a significant change in
its metabolism and can be observed using the gas sensor array. Ethanol
prediction with the MBC approach was compared with the ethanol
prediction using the CCM approach. For this reason RMSEP and SEP were
calculated and the results are shown in Table 3.
Table. 3 reveals that, by using the MBC approach, the SEP of prediction
is below 7 % in all cases besides when BC2 is used for calibration and
the data set from BC1 is used for validation (SEP = 9.3 %). However,
even though in the model-based calibration approach, during the
calibration procedure of the chemometric model no off-line measurement
were used, the prediction corresponds very well with the simulated
values. Furthermore, in comparison with the predictions from using the
classical calibration method, lower prediction errors are obtained.
The larger errors of prediction from the classical calibration method
are because not so many off-line samples are used for calibration.
Therefore, by increasing the number of off-line samples, more accurate
predictions might be obtained. However this would be a time consuming
and expensive approach.
When using the MBC approach, the percentage error highly depends on the
kinetic parameters of the simulation model which are obtained from the
optimization algorithm. If they are close to the real values, the
process model will describe the process sufficiently accurate.
Therefore, improving the optimization process can even lead to even
lower percentage errors. Of course, the percentage errors also depend on
how good the off-line samplings were performed and how accurate they
are. However, the prediction error for ethanol concentrations are below
10 % in all three cultivations which is considered decent for a
bioprocess especially when considering that no off-line values were used
to achieve this result.