CARL L. WILSON Tennessee Eastman Co., Division of Eastman Kodak Co., Kingsport, Tenn.
Production
of
a New Chemical
A Designed Experiment This study in existing plant equipment provides a clear example in which the lowest product cost is not obtained under conditions which give highest yield
AN
DEVELOPING a process for producing a new chemical in the laboratory or pilot plant, unit operations similar to those used in plant-scale production are ordinarily used. Usually the process recommended for use on a plant scale results from considerable experimentation in the pilot plant. In some cases, however, the operating conditions developed in the laboratory may be used in proceeding directly to plant-scale operation, omitting the pilot-plant step. Experimental work can be done in a production plant. The purpose of this article is to describe such a program and to show its advantages. In this study a designed experiment on production of a new chemical was carried out in equiprrient used for producing another material by a similar process. The process for producing the new chemical had been investigated adequately in the laboratory, using factorial experiments. The step from laboratory to plant-scale represented an increase of approximately a thousandfold; therefore, optimum process conditions on the plant scale might be quite different from those developed on the laboratory scale. Even in scaling-up from the pilot plant to a production unit it is common to find that while there is a similarity in the relationship of process variables, the optimum conditions are often changed. The main objective of this production run was to manufacture a given quantity of satisfactory material for use in market evaluation. A second important objective of the run was to produce information to determine the most economical operating conditions for the existing plant and for the designing of a new plant. In planning a production experiment the quality of the product _ . requirement ^ must be given careful consideration. Experimental variations in conditions
which they will be varied the committee approach is recommended. The committee should consist of members with practical and theoretical knowledge of the product and the process as well as skill in the design of experimentsfor example, production superintendent, production supervisor, chemist, development superintendent, chemical engineer, and statistician. In other problems it may be advisable to substitute a physicist, a mechanical engineer, an electrical engineer. an accountant or others with requirpd special skills. A chemical process normally involves several steps from raw materials to refined product. Each step has a number of conditions which can be varied and the results (effects) of these variables can be observed. As the length of this production run liniitrd thc number of possible experiments, the most important variables were selected for the firqt studies. A few key variables which can be isolated from the rest of the process are preferred. In producing the new chemical, it had been found in the laboratory that the step involving a condensation reaction affected the total cost of the product more than any other. Therefore, these variables were considered-temperature, reactor, kind of solvent, solvent concentration, operator, pressure, agitation,
Table I. The Actual Responses for Each of the Runs Are Shown for Yield, Production, and Cost. The Variability from Cycle to Cycle a t Each Set of Conditions Can Be Noted Run Run Cycle 1 2 3 4 5 Cycle 1 2 3 4 5
must be made within limits that produce responses but not off-grade material. In selecting the proper variables and the limits over
504
feed rate, purity of feed, temperature of feed, kind of catalyst, and catalyst concentration. The temperature and the feed rate were selected as the first variables to study because their importance was shown in previous laboratory work. The yield increased as the temperature decreased, and the productivity in a given vessel increased as the temperature increased. Laboratory data indicated that the process in the plant should be run at 25" C , using a feed rate of 3.0 gallons per minute. A two-factor, two-level factorial experiment with a center point was designed (Figure 1). The four corners are all the possible combinations of the two factors at the two levels. The center point is halfway between the two levels of temperature and feed rate. This center point is the recommended operating point obtained from laboratory data, and the four surrounding points explore the advantages to be gained by moving the operating point in any direction in the plane. The center point provides a means of comparing the results 01 all five operating conditions us. constant conditions at the center point. The levels of the factors are determined on the basis of experience. If the Ievels of the factors are svt too far apart, it is possible that quality problems
INDUSTRIAL AND ENGINEERING CHEMISTRY
Yield Response, % 1 2 4
71 73 78 72
74 79 82 83
67 65 74 70
84 85 79 81
Cost Response, $/lb. 64 69 69 76
Production Response, lb./day 1 2 3 4
2655 2405 2390 2580
1895 1940 1788 1960
3278 3240 2860 3160
2243 2200 2040 2460
2454 2495 2120 2300
1 2 3 4
0.433 0.456 0.474 0.438
0.531 0.512 0.559 0.501
0.393 0.401 0.407 0.393
0.455 0.459 0.495 0.433
0.432 0.450 0.502 0.462
0.450 73.5 2508
0.526 79.5 1896
0.398 69.0 3160
0.460 82.2 2236
0.462 69.5 2342
Average Cost, S/lb.
Yield, % Production Ib./day
STATISTICS IRI CHEMICAL PROCESSES
(3)
(4) 24
30 Feed, GPM
36
Figure 1. The five sets of conditions carried out in the experiment are shown numerically in the order in which they are to b e run
will be encountered or undesirable variations in production and yield will be obtained. If the levels of the factors are set too close, no measurable effects will be obtained. When an experiment is designed, careful consideration must be given to the choice of the results (effects) of the variations in operating conditions which will be studied. These results, which are called responses, are the dependent variables. The responses considered were: cost, yield, production rate, color, hardness, chemical analysis, and other physical properties. A careful study was made in the effect of the experiment on cost, yield, and production rate. Other responses were observed to ensure that they remained within specified limits. Each of the five sets of conditions was maintained for 24 hours. Therefore, the yield and production rate determined for the 24hour period were used as responses in the experiment. No leveling period after changing conditions was necessary in this case. The following cost equation was developed from an economic study of the process:
c
= [F
on the information board (Figure 2) in the same order as the sequence of runs in Figure 1. A small replica of Figure 1, shown at the left of the running averages, makes it easy to determine the conditions which gave each response. The information board (Figure 2) presents the three responses, the effects with their 95% confidence limits, the change in mean with its 95y0confidence limit, and the standard deviation of each response. The 95yo confidence limits are expressed as a plus or minus value with each effect. For example:
was termed a cycle, and four cycles were completed during the production run. Data from repetition of the cycles provided a measure of the variability of the process. The plan of using short experiments under each set of conditions in four cycles aided in removing time trends which might have interfered in determining effects if each of the five sets of conditions had been carried out four consecutive days in a single cycle. The evolutionary operation method of calculation using standard forms developed by Box and Hunter (2) was used.
95% Confidence
Effect Temperature effect (cost) Change in mean (cost)
Limits
Conclusion
rt0.015 rt0.0132
Effect is significant at 95y0 level Effect i s not significant at 95% level
The results of the calculations were listed on an information board such as that described by Box and Hunter (2). The information board is advantageous as it shows all the main responses together. The responses obtained on each set of operating conditions for each of the four cycles are given in Table I. The data were calculated at the completion of each cycle and posted on the information board. The calculation method revealed the effects of the factors (temperature and feed rate) and gave a mathematical probability of the significance of the effects. Calculation of the results a t the end of each cycle reveals the status of the experiment on a continuous basis and allows action to be taken in changing experimental strategy at the earliest possible time. The average response for each set of conditions at the end of cycle 4 is shown
Cycle 4
The effect of a factor is the change in response produced by a change in the factor from the low level to the high level. The calculation of the effect is made in such a way that the sign before the effect indicates the direction to move the level of the factor to maximize the response. If a minimum response is desired, the level of the factor is moved in the direction opposite to the sign. The interaction effect ( T X F ) shows if the effect of one factor is different at different levels of the other factor. The confidence limits indicate whether or not the interaction effect is significant. If there is a significant effect, the factors are said to interact. The change in mean gives a measure of the effect of operating by the experimental plan compared to conducting all runs at the center point conditions. Figure 3,A presents the yield responses with the effects and their 95y0confidence
Phase 1 v)
25
P
+ VP/Y] c P
e
2.4
C = cost, 46 per lb. P = daily production, lb. F = fixed cost ($700 per day) V = variable cost, $0.12 per lb. of product a t 10Oyq yield (materials plus operating costs) Y = yield, fraction This equation produced cost data quickly and removed other causes of cost variation from the data. T o determine the sequence of operating conditions, the technique employed in “evolutionary operation” (7) was used. Five sets of conditions were carried out in numerical sequence, starting with experiment 1 (Figure 1). A complete set of the five conditions
-0.063 +0.009
3.0
Frcrd
3.6
OL
95% Error Limits
0.0148
For Averages
Effects With 95% Error Limits
II
1
I1
3.59
Feed rate
-0.065
-C
.015
+1.1
Temp.
-0.063
j=
.015
-11.6
+0.001
j=
.015
T
x
F
Change In Mean
Standard Deviation
-1.6 I
1
+0.009
.0132 +1.2 0.0148
i:
* * j=
104
+239
* * *
-80
-c
3.59
+579
3.59
+685
3.59 I
3.19
3.59
/I 104 104 104 ll
93
104
Figure 2. This information board presents the three responses, the effects with 9570 confidence limits, the change in mean with its 95y0 confidence limit, and the standard deviation of each response VOL. 52, NO. 6
JUNE 1960
505
limits. The indicated feed rate effect of +1.1 was obtained by adding the two responses at 2.4 gallon-per-minute feed and subtracting the sun1 of the two responses at 3.6 gallon-per-minute feed and dividing by 2. The confidence limits of 4 3 . 5 9 indicate that the 1.1 effect is not significant. The temperature effect of -11.6 was obtained by adding the two responses at 30' C. and subtracting the sum of the two responses at 20' C. and dividing by 2. The confidence limits of h 3 . 5 9 indicate that the - 11.6 temperature effect is significant. The 9570 confidence limits of the temperature effect are -11.6 to h 3 . 5 9 or -15.2 to -8.0. The indicated interaction ( T X F ) effect of - 1.6 was determined by adding
69.0 e
69.5 e
304
A 755
82.2 e
79.5 e
24
30
36
Feed, G.F!M 3160 0
"1
B
2508
I
2236 0
,
3.0
2.4
3.6
F e e d , G.PM.
0.462
0.398 e
e
C 0.450 e
25
0.460
0.526 20
2.4
3.0
+
INDUSTRIAL AND ENGINEERING CHEMISTRY
30
2o
+
-1
I
-&-
I
1 " " 24
30
36
4 I2
Feed . G P M
Figure 4. The original design and new design are given with the points in common shown
shifted in the direction of reduced cost. LVhen no further reduction in cost is obtained, then additional factors may be added or substituted for one or both of the present factors. The search for optimum conditions should be a continuous one, studying the most important factors first. The experiment also provided a basis for recommending a change in design of the condensation reactor to increase capacity, yield, and to lower costs. These data will be used in designing a new plant when the product obtains market acceptance. This technique may be used to determine new optimum conditions when equipment or processes are changed. Statistically designed experiments enable existing production equipment to be used to produce a new chemical for market evaluation and at the same time to produce information to determine optimum operating conditions and provide data for a new design. There are other experimental designs and methods of calculation, some of which can be more efficient in producing information from the limited time available. However, the technique of evolutionary operation is of value in this type of production experiment due to its simplicity of calculation and clarity of presentation. The requirements of the problem must be considered in selecting the best experimental design.
A
experiment. Under these new conditions the cost of the product was reduced $0.05 per pound with a 26% increase in productivity. The experimental data suggested that even lower costs might be achieved. Further production runs were made at the experimental conditions (Figure 4). The area of experimentation has been
Changeinmean
.
B
1.1 i 3.59 -11.6 i.3.59 - 1.6 3z 3.59 1.213.19
3.6
Yield response in per cent Production response in pounds per d a y C. Cost response in dollars per pound
506
Feed rote Temperature
T X F
Figure 3. The average response at the end of four cycles i s shown in information board form B.
Effects with 95% error limits
e
Feed, G.P.M
A.
the two responses at the ends of a diagonal pointing upward to the right and subtracting the responses of the runs at the ends of a diagonal pointing upward to the left and dividing by 2. The confidence interval of =!=3.59 indicates that the - 1.6 interaction effect is not significant. The change in mean effect of +1.2 was determined by adding the four periphery runs and subtracting four times the center run and dividing by 5 (the number of runs). The confidence limits of 1 3 . 1 9 indicate that the +1.2 change-in-mean efyect is not significant at the 957& level. Figure 3,B presents the production rate responses with the effects and their 95$& confidence limits. The effects of feed rate and temperature were both significant at the 957, level. A higher production rate was obtained by increasing both feed and temperature. The interaction ( T X F ) was significant at the 95% level. The increase in production was greater when the feed rate was increased with the temperature at 30' C. than when it was at 20' C. Also, there was a greater increase in production when the temperature was increased with the feed at 3.6 gallons per minute than with the feed at 2.4 gallons per minute. Figure 3,C presents cost responses with the effects and their 95% confidence limits. Cost response is the most important response in this problem. Yield and production responses were used in developing cost responses. Effects of feed rate and temperature on cost were both significant at the 950/;, level. Increases in temperature and feed rate rrduced cost. The interaction ( T X F ) did not shoiv a significant effect on cost. The change in mean did not show a significant effect on the cost. Therefore, the cost of operating at the five different sets of conditions in the experiment was not significantly different from the cost had all runs been made at the center point conditions. The lowest cost was obtained at 30' C and 3.6 gallon-per-minute feed. This feed rate and temperature gave the lowest cost and the highest production although the lowest yield. The quality of the product was within all s p e d cation requirements throughout the
+579 =J= 104 f685 zk 104 f239 4 104 - 80h93
C -0.065 i.0.015 -0.063f0.015 f O . 0 0 1 f 0.015 +0.009~0.0132
literature Cited (1) Box, G. E. P., Afifil. Sfatistics 6, No. 2, 2-23 (1957). (2) Box, G. E. P., Hunter, J. S., Technornetrics 1, No. 1, 77-95 (1959). RECEIVED for review September 15, 1959 ACCEPTED March 22, 1960 Division of Industrial and Engineering Chemistry, 136th Meeting, ACS, Atlantic City, N. J., September 1959.
