Application Note                                                                                              Zaptron Systems, Inc.

Multi-target Optimization in Glutamate Fermentation

Introduction:

The optimization of fermentation process in glutamate production is similar to the optimization of fermentation in penicillin production described elsewhere. However, this is a 4-target, 3-class and 14-dimensional optimization problem.

Optimization Targets:

There are four (4) targets that need to be optimized in the process. They are:

1) Conversion rate (from glucose to glutamic acid) – should be high
2) Productivity – should be high
3) Yield – should be high
4) Length of fermentation time period – should be short

Data Conditioning

Each batch (tank) was considered as one sample. Data from 80 production batches (tanks) were used as training set. To simplify data processing, data of each fermentation period, which is usually 30 or more hours long, was divided into three segments:

	First segment - data from 0 to 12 hours
	Second segment -data from 12 to 24 hours
	Third segment - data from longer than 24 hours (to the fermentation end)

Some of data, including PH value, ventilation rate and temperature, were then averaged in each of the three segments to generate the averaged features for each segment. Data samples were further classified into 3 classes according to the range of features:

	Class 1: the glucose-to-sugar conversion rate is larger than 50%, and the time period of fermentation is less than 34 hours.
	Class 2: the glucose-to-sugar conversion rate is smaller than 49.5%, and the time period of fermentation is larger than 34.5 hours.
	Class 3: samples have other parameters

Feature Selection:

Features that have been selected by using MasterMiner include relevant operation parameters, such as temperature, ventilation, PH value, and etc., and the physical-chemical data (as time series data) of liquors, such as glucose concentration, OD value and etc.  The OD is an optical property of sugar (for example glucose). Since a change in OD value indicates a change in substance concentration or composition, the OD vale is used as a parameter for monitoring the process.

The complete set of features for the overall fermentation process is listed below:

x1 - transparency of liquor
x2 - glucose concentration at starting point
x3 - PH of liquor
x4 - PH value at the final stage of germ plantation
x5 - increase of (change in) OD value
x6 - PH value in first segment (averaged)
x7 - PH value in second segment (averaged)
x8 - PH value in third segment (averaged)
x9 - ventilation rate (cubic meters/min) in first segment (averaged)
x10 - ventilation rate (cubic meters/min) in second segment (averaged)
x11 - ventilation rate (cubic meters/min) in third segment (averaged)
x12 - temperature in first segment
x13 - temperature in second segment
x14 - temperature in third segment

Findings:

By PLS (partial least square) regression, it has been found that no linear relationships existed among features or between feature and target. It therefore seems impossible to change any single feature to improve the targets. It appears that we have to define the optimal zone directly in the multi-dimensional hyperspace. By MasterMiner software, at last we found that the optimal zone is near a hyper-plane. After comparing the practical operation data and this hyper-plane, we have adjusted four features {x4, x6, x7, x8} to slightly lower than original values, and other parameters slightly higher. The results are rather good.

After finding the boundary of the optimal zone, MasterMiner also offers an operational advisory to the fermentation technicians by adding and testing a number of virtual samples to the optimal zone. A test sample generated by MasterMiner that falls inside this optimal zone is considered as an optimal sample, and it could be expected to lead to the optimization of the 4 targets in fermentation. Three such optimal samples by MasterMiner, called prediction values, are listed in the following table: (the mean value is used).

Table of Some Optimal Samples by MasterMiner

#	X1	X2	X3	X4	X5	X6	X7	X8	X9	X10	X11	X12	X13	X14
1	.37	-.14	.47	-.09	-.078	-.11	-.05	-.24	24	20	18	-.27	1.34	.59
2	.84	-.13	.30	.03	-.016	.09	-.04	-.20	33	30	46	.39	1.08	.64
3	.56	-.01	.18	-.14	-.026	.01	-.03	-.12	-3	6	71	-.15	.74	.35

The above prediction value are roughly in agreement with the result obtained by another method whereby single features were adjusted to search for the best operational direction. However, some features, such as X2, are not in agreement in these two methods.

In addition, MasterMiner offered the following advisory to production:

	PH value should (in this concrete case) be slightly reduced
	Temperature should be slightly increased.
	Ventilation rate should be slightly increased.