the components in the various culture medium formulations and the

initial cell concentration were the input variables and the cell concentration at any instance
was the output variable of the FBN model.
2. Materials and methods
2.1. Database
Raw data from 426 experiments originally carried out for a genetic
algorithm-based search for an optimal culture medium for the marine
dinoflagellate K. veneficum were used [5]. This database comprised 6 different batches
(generations) of 71 experiments each. Each experiment
had a different composition of the culture medium and was carried
out in triplicate.
The above database was supplemented with new batch experiments
carried out using extreme formulations of culture media, that is formulations with
exceptionally high or nil amounts of certain constituents.
The responses of these extreme formulations and the other data were
used in training the FBN to enhance its predictive capability. The extreme dataset spanned
three levels of the initial cell concentration in
the culture medium (i.e. 1 × 104, 2 × 104 and 3.3 × 104 cells mL-1)
and the following in independent experiments: (1) the L1 medium
with nitrate concentration ranging from 0 to 882 μM; (2) the L1 medium without phosphate;
(3) the culture medium formulation “MAX”
that used the maximum concentrations of all the nutrients as shown
in Table 1; and (4) seawater without any added nutrients.
Inocula were grown as batches under a 12:12 h light–dark cycle at
18 ± 1 °C. Filter-sterilized (0.22 μm membrane filter) L1 medium prepared using natural
Mediterranean Sea water was used as the control
medium in all experiments. Cells in their exponential growth phase
were used for inoculation. Multiwell plates (2-mL per well) were used
as growth vessels. Cultures were illuminated from the top using Philips
TLD 36 W/54 fluorescent lamps at an average irradiance of 200 μmol
photons m-2 s-1. All experiments were carried out in triplicate.
Growth was measured by cell counts every three days

The ANN input variables consisted of a k × s matrix with s (=

1398)
columns of the initial concentrations of 25 nutrients selected from
those
commonly found in seawater-based algal culture media [5]; the
initial
cell concentration (C0); and the time of measurement (t). The
ranges
of the values for the inputs used in establishing the ANN model
and
the output data are shown in Table 1. The target output variable
was
the dimensionless cell concentration (Ct/C0) in the culture at any
time
t. Here, Ct is the cell concentration at time t. As the values of the
input
variables ranged up to eight orders of magnitude (Table 1), the
input
and target variables were normalized using the following
equations:

the components in the various culture medium formulations and the

initial cell concentration were the input variables and the cell concentration at any instance
was the output variable of the FBN model.
2. Materials and methods
2.1. Database
Raw data from 426 experiments originally carried out for a genetic
algorithm-based search for an optimal culture medium for the marine
dinoflagellate K. veneficum were used [5]. This database comprised 6 different batches
(generations) of 71 experiments each. Each experiment
had a different composition of the culture medium and was carried
out in triplicate.
The above database was supplemented with new batch experiments
carried out using extreme formulations of culture media, that is formulations with
exceptionally high or nil amounts of certain constituents.
The responses of these extreme formulations and the other data were
used in training the FBN to enhance its predictive capability. The extreme dataset spanned
three levels of the initial cell concentration in
the culture medium (i.e. 1 × 104, 2 × 104 and 3.3 × 104 cells mL-1)
and the following in independent experiments: (1) the L1 medium
with nitrate concentration ranging from 0 to 882 μM; (2) the L1 medium without phosphate;
(3) the culture medium formulation “MAX”
that used the maximum concentrations of all the nutrients as shown
in Table 1; and (4) seawater without any added nutrients.
Inocula were grown as batches under a 12:12 h light–dark cycle at
18 ± 1 °C. Filter-sterilized (0.22 μm membrane filter) L1 medium prepared using natural
Mediterranean Sea water was used as the control
medium in all experiments. Cells in their exponential growth phase
were used for inoculation. Multiwell plates (2-mL per well) were used
as growth vessels. Cultures were illuminated from the top using Philips
TLD 36 W/54 fluorescent lamps at an average irradiance of 200 μmol
photons m-2 s-1. All experiments were carried out in triplicate.
Growth was measured by cell counts every three days

The ANN input variables consisted of a k × s matrix with s (=

1398)
columns of the initial concentrations of 25 nutrients selected from
those
commonly found in seawater-based algal culture media [5]; the
initial
cell concentration (C0); and the time of measurement (t). The
ranges
of the values for the inputs used in establishing the ANN model
and
the output data are shown in Table 1. The target output variable
was
the dimensionless cell concentration (Ct/C0) in the culture at any
time
t. Here, Ct is the cell concentration at time t. As the values of the
input
variables ranged up to eight orders of magnitude (Table 1), the
input
and target variables were normalized using the following
equations:
the components in the various culture medium formulations and the
initial cell concentration were the input variables and the cell concentration at any instance
was the output variable of the FBN model.
2. Materials and methods
2.1. Database
Raw data from 426 experiments originally carried out for a genetic
algorithm-based search for an optimal culture medium for the marine
dinoflagellate K. veneficum were used [5]. This database comprised 6 different batches
(generations) of 71 experiments each. Each experiment
had a different composition of the culture medium and was carried
out in triplicate.
The above database was supplemented with new batch experiments
carried out using extreme formulations of culture media, that is formulations with
exceptionally high or nil amounts of certain constituents.
The responses of these extreme formulations and the other data were
used in training the FBN to enhance its predictive capability. The extreme dataset spanned
three levels of the initial cell concentration in
the culture medium (i.e. 1 × 104, 2 × 104 and 3.3 × 104 cells mL-1)
and the following in independent experiments: (1) the L1 medium
with nitrate concentration ranging from 0 to 882 μM; (2) the L1 medium without phosphate;
(3) the culture medium formulation “MAX”
that used the maximum concentrations of all the nutrients as shown
in Table 1; and (4) seawater without any added nutrients.
Inocula were grown as batches under a 12:12 h light–dark cycle at
18 ± 1 °C. Filter-sterilized (0.22 μm membrane filter) L1 medium prepared using natural
Mediterranean Sea water was used as the control
medium in all experiments. Cells in their exponential growth phase
were used for inoculation. Multiwell plates (2-mL per well) were used
as growth vessels. Cultures were illuminated from the top using Philips
TLD 36 W/54 fluorescent lamps at an average irradiance of 200 μmol
photons m-2 s-1. All experiments were carried out in triplicate.
Growth was measured by cell counts every three days

The ANN input variables consisted of a k × s matrix with s (=

1398)
columns of the initial concentrations of 25 nutrients selected from
those
commonly found in seawater-based algal culture media [5]; the
initial
cell concentration (C0); and the time of measurement (t). The
ranges
of the values for the inputs used in establishing the ANN model
and
the output data are shown in Table 1. The target output variable
was
the dimensionless cell concentration (Ct/C0) in the culture at any
time
t. Here, Ct is the cell concentration at time t. As the values of the
input
variables ranged up to eight orders of magnitude (Table 1), the
input
and target variables were normalized using the following
equations:
n the FBN training process, MSE, the mean squared error between
the experimental data and the corresponding predicted data, was calculated and then
propagated back through the network in each cycle. The
algorithm adjusted the weightings between the input layer, the hidden
layer and output neurons in order to reduce the error. This procedure
was repeated until the error between the experimental data and the
predicted values satisfied the typical criteria for Bayesian regularization.
The weights (Wij and Zj) and bias (b1j and b2) values in Eq. (4) were obtained after
the training phase of the FBN model had been comple
n the FBN training process, MSE, the mean squared error between
the experimental data and the corresponding predicted data, was calculated and then
propagated back through the network in each cycle. The
algorithm adjusted the weightings between the input layer, the hidden
layer and output neurons in order to reduce the error. This procedure
was repeated until the error between the experimental data and the
predicted values satisfied the typical criteria for Bayesian regularization.
The weights (Wij and Zj) and bias (b1j and b2) values in Eq. (4) were obtained after
the training phase of the FBN model had been comple
n the FBN training process, MSE, the mean squared error between
the experimental data and the corresponding predicted data, was calculated and then
propagated back through the network in each cycle. The
algorithm adjusted the weightings between the input layer, the hidden
layer and output neurons in order to reduce the error. This procedure
was repeated until the error between the experimental data and the
predicted values satisfied the typical criteria for Bayesian regularization.
The weights (Wij and Zj) and bias (b1j and b2) values in Eq. (4) were obtained after
the training phase of the FBN model had been comple
Fig. 3 compares the model predicted values of the Ct/C0 and the measured data. The
comparisons are shown for the training runs, the test
runs and the combined (training and test) runs. There was excellent
agreement between the predicted and the measured data in the training
runs (Fig. 3). In the validation test runs, the data were event distributed
around the parity line (y = x) nearly all the predicted and the measured
values agreed within about ±75% of the measured data (R = 0.9138).
Using the combined data sets of the training runs that the test runs,
the measured and the predicted values of Ct/C0 were strongly correlated
(R N 0.984) and evenly distributed around the parity line (Fig. 3). This
confirmed the ability of the FBN model to predict new data. This is clearly seen in
Fig. 4, where the predicted and the observed values of Ct/C0
versus the observation order are shown.
Fig. 3 compares the model predicted values of the Ct/C0 and the measured data. The
comparisons are shown for the training runs, the test
runs and the combined (training and test) runs. There was excellent
agreement between the predicted and the measured data in the training
runs (Fig. 3). In the validation test runs, the data were event distributed
around the parity line (y = x) nearly all the predicted and the measured
values agreed within about ±75% of the measured data (R = 0.9138).
Using the combined data sets of the training runs that the test runs,
the measured and the predicted values of Ct/C0 were strongly correlated
(R N 0.984) and evenly distributed around the parity line (Fig. 3). This
confirmed the ability of the FBN model to predict new data. This is clearly seen in
Fig. 4, where the predicted and the observed values of Ct/C0
versus the observation order are shown.
Fig. 3 compares the model predicted values of the Ct/C0 and the measured data.
The comparisons are shown for the training runs, the test
runs and the combined (training and test) runs. There was excellent
agreement between the predicted and the measured data in the training
runs (Fig. 3). In the validation test runs, the data were event distributed
around the parity line (y = x) nearly all the predicted and the measured
values agreed within about ±75% of the measured data (R = 0.9138).
Using the combined data sets of the training runs that the test runs,
the measured and the predicted values of Ct/C0 were strongly correlated
(R N 0.984) and evenly distributed around the parity line (Fig. 3). This
confirmed the ability of the FBN model to predict new data. This is clearly seen
in Fig. 4, where the predicted and the observed values of Ct/C0
versus the observation order are shown.