- Email: [email protected]
Horizontal cross-flow bubble column reactors: CFD and validation by plant scale tracer experiments
Chemical Engineering Science 62 (2007) 5495 – 5502 www.elsevier.com/locate/ces
Horizontal cross-flow bubble column reactors: CFD and validation by plant scale tracer experiments P.A.A. Klusener, G. Jonkers, F. During, E.D. Hollander, C.J. Schellekens, I.H.J. Ploemen, A. Othman, A.N.R. Bos ∗ Shell Global Solutions International B.V., P.O. Box 38000, 1030 BN Amsterdam, The Netherlands Received 16 June 2006; received in revised form 30 March 2007; accepted 31 March 2007 Available online 13 April 2007
Abstract Unique gas and liquid phase radioactive tracer experiments have been performed in horizontal cross-flow bubble column reactors of a commercial ethylbenzene oxidation plant. These experiments validated CFD modelling predictions, which in turn formed the basis for evaluation and selection of potential measures to reduce “oxygen starvation” predicted to occur for certain configurations and operating conditions. After implementation in commercial reactors lower operating temperatures, improved selectivity and thus better plant yields were achieved. 䉷 2007 Shell Global Solutions International B.V. Published by Elsevier Ltd. All rights reserved. Keywords: Bubble column; Radioactive tracer; CFD; Ethyl benzene; Propylene oxide; PO; SMPO
1. Introduction Propylene oxide is a versatile chemical intermediate used in a wide range of industrial and commercial products. Current world production is over 6 million metric tons a year. While other processes exist, the Shell Chemicals companies have derived a strong competitive advantage by using and continually developing their proprietary SM/PO technology, a process in which propene and ethylbenzene (EB) are converted into propylene oxide (PO) and styrene monomer (SM), respectively. Worldwide there are now five world-scale SM/PO plants based on Shell Technology, the most recent one started up in China in 2006. The first step in this multi-step process is the air-oxidation of ethylbenzene to ethyl-benzene hydroperoxide (EBHP). This is performed by Shell in cross-flow operation in a series of large horizontal bubble column reactors with a very low aspect ratio: height 4–6 m, length 15–25 m. Typically, a reactor train consists of 4–5 of such horizontal columns, which are equipped with baffles and heating/cooling coils. Air is introduced via ∗ Corresponding author. Tel.: +31 20 6302887.
E-mail address: [email protected] (A.N.R. Bos).
separate middle and side sparger systems. The gas outlet stream contains besides unconverted oxygen a very significant amount of EB from evaporation/stripping and this EB is recovered in a condensing column and recycled to the reactor train. The literature on horizontal cross-flow bubble columns is rather limited, see e.g. Tilton and Russell (1982), Zuiderweg and Bruinzeel (1971); Krzysztofoski et al. (1986). Pohorecki et al. (2001) did mention tracer experiments in similar commercial reactors, but did not disclose any details. Within Shell we have developed computational fluid dynamics (CFD) models as well as integrated reactor models with complex reaction kinetics for these rather unconventional bubble columns. To validate some of the model predictions it was decided to perform radioactive tracer experiments in one of the commercial plants during normal operation. Two completely different types of tracer experiments, with completely different purposes, have been performed: (1) liquid phase and (2) gas phase tracer experiments. 2. Liquid phase tracer experiments The peroxidation of EB to EBHP is autocatalytic implying that a minimum concentration of the product EBHP is necessary
0009-2509/$ - see front matter 䉷 2007 Shell Global Solutions International B.V. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.ces.2007.03.044
5496
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
Fig. 1. Liquid phase tracer injection and detection points.
to initiate the reaction and that the reaction goes faster on increasing EBHP concentration. On the other hand, the higher EHBP concentration leads to increased EBHP decomposition, by-product formation and lower selectivity. For this reason backmixing is desirable in the front-end of the first reactor, whereas plug flow behaviour is desirable in the remaining reactor train. To facilitate this, the reactors are segmented by a few transversal baffles. However, in the first reactor of the train a lower number of such baffles are installed. The two main incentives for investigating the actual mixing in the reactor train are to determine: 1. The effectiveness of the transversal baffles in facilitating plug flow (“staging” by forming compartments in series) and subsequently the scope for improved selectivity by further reducing the deviations from plug flow. 2. The effectiveness of installing a lower number of such baffles in promoting backmixing in the first reactor and hence the scope for improvement by (further) enhancing backmixing here. Fig. 1 illustrates where the tracer injections and detections in the actual plant were positioned. The experimentally obtained RTD curves (count rate vs. time) were fitted with both the tanksin-series model and the axial dispersion model, see standard reaction engineering textbooks. Both models have basically two parameters: the mean residence time and the parameter describing the spread in residence time, i.e. the number of tanks N and the Péclet number, respectively. In principle, could be calculated for (accurately) known flow rates, volumes, levels and gas hold-ups. However, here is considered to be an unknown model parameter and is one of the two fit parameters. This is not only because of the presence of significant EB evaporation but also because the plant measured or calculated values in practice are often not very accurate. The time axis was set to zero at the time that the injected bolus of activity passed the detectors positioned at the reactor inlet. 2.1. Results of RTD model parameter estimation Fig. 2 shows the results of fitting the RTD data of R102 with the tanks-in-series model for N = 3, 4, 5 and 6. For each N the
average residence time was optimised. It can be clearly seen that this model cannot quite fit the data: either the left-hand or the right-hand side of the measured RTD curve can be modelled well, but not the full curve with one set of parameters. Fig. 3 shows the result of fitting the data of R102 using the axial dispersion model, using open–open boundary conditions to allow easy fitting in the time domain. A remarkably good fit was obtained and the resulting Péclet number was 6.6. In Fig. 4 the tracer had been injected into the inlet of R101 with measurement, respectively, at the outlets R101 and R102. Unfortunately, no detector was present in the outlet of R104 during these experiments, and for R104 we only have the signal recorded by the detector placed at the circumference of R104, i.e. mounted on the reactor vessel itself rather than on the outlet piping. So this is not really the RTD over the whole reactor train. Consequently, this R104 signal can only be used qualitatively for interpretational purposes. 2.2. Translation to closed–closed boundary conditions It is well-known that for a sufficiently high Péclet number (say Pe > 20) the differences between “open–open” and “closed–closed” (or Danckwerts) boundary conditions become relatively small, but for very low Pe (say Pe < 5) the differences can become considerable. Since the fitted values are intermediate, we therefore have estimated the Péclet numbers associated with the “closed–closed” boundary condition model. This was done as by equating the variance of the “open–open” (oo) model to that of the “closed–closed” (cc) model: 2 2Pecc − 2 + 2e−Pecc 1 2 8 × = + . Peoo 1 + 2/Peoo Pe2cc Pe2oo (1) The factor (1 + 2/Pe)2 arises to properly account for the inherently different normalisation times in the dimensionless equations for the variances as functions of the Péclet numbers. We believe this method is quite accurate except for very low Péclet numbers (say < 3; for Pe < 1 the method will surely fail). As expected, the Péclet numbers found for the “closed–closed” boundary conditions are somewhat lower: e.g. Peoo = 7.5 with “open–open” conditions becomes Pecc = 6.7 for “closed–closed” conditions. The latter Péclet numbers were used to correlate to the tanks-in-series model. 2.3. Discussion of the liquid phase results Table 1 summarises the results; also the “optimum” tanks-inseries fit results are given. In the last column the experimental mean residence time (exp ) is given, somewhat arbitrarily calculated from reactor liquid volume divided by the outlet flow, i.e. implicitly assuming EB evaporation to be instantaneous at the inlet) . In general, a good fit with exp was found. The cc values for the “closed–closed” model match the experimental figures much better than those of the “open–open” model, indicating that indeed the “closed–closed” axial dispersion model is more appropriate for the EB oxidation reactors.
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
Fig. 2. Experimental RTD for R102 and tanks in series model fits showing the inadequacy of this model.
Fig. 3. Experimental RTD for R102 and axial dispersion model fits.
Fig. 4. Measured and fitted RTD for injection in inlet R101 and detection in outlet R101, R102 and inside R104 and axial dispersion model fits.
5497
5498
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
Table 1 Summary of liquid phase RTD measurements Enclosed reactor system
Axial dispersion model “Open–open” fit
“Closed–closed” estimation
oo
cc
Peoo
(s) R101 R102 R104 R101 + 02 R101 + 02 + 03 + 04∗d R104∗e Correctedf R101 + 02 + 03 + 04
Tanks-in-series model
a
Pecc
b
(s)
Dax (m2 /s)
Exp
“Tank” fit Ncalc
c
N
N
exp
(s)
(s)
883 925 1084 2019 4209 643
7.5 6.6 6.6 14.4 22.4 3.4
1118 1207 1414 2299
6.7 5.7 5.7 13.6
0.088 0.096 0.080 0.084
3.8 3.4 3.4 7.3
5 4 4 8 12 3
1040 1134 1330 2229 4488 867
1063 1131 1361 2194
4650
25.6
5013
24.7
0.085
12.9
13
4951
4786
a
cc = oo (1+ b Derived from
2/Pefit ). Peoo using Eq. (1). c N = Pe /2 + 0.5. cc d Outlet R104 not measured; detector measuring inside of reactor! e Side detector data from injection in R104. f Calculated by subtracting the values for R104∗ (side detector) from R101 + 02 + 03 + 04∗ and ding those of R104 (outlet detector). The “closed–closed” estimates were subsequently calculated from the so obtained Peoo values.
For the full train of four reactors, only an estimate of the RTD can be made because no measurement results are available for injection into R101 and measurement at the outlet of R104. Hence, the measured value in Table 1 of Pe = 22.4 is too low. Assuming that Péclet numbers are approximately cumulative for sufficiently high Péclet, the Pe number for the full reactor train was calculated from the side R104 measurement at R101 injection and the Pe number obtained from R104 injection. This procedure provides a Peoo of 25.6 from which Pecc = 24.7 was derived; the latter corresponds to approximately 13 tanks-inseries (also obtained when fitting N). R101 contains fewer baffles and hence fewer compartments; therefore it might be expected (and actually it was hoped for) to have more backmixing in R101. However, the Péclet number of R101 is slightly higher—i.e. slightly less backmixing—than the number found for R102: Pe = 6.7 vs. Pe = 5.7. This would suggest that the transversal baffles do not significantly influence the backmixing, i.e. no significant staging is achieved with these baffles. However, to assess the effect of the number of baffles one should also look at the dispersion coefficients Dax rather than only at the Péclet numbers because the residence times of the reactors increase significantly from R101 to R104 due to EB evaporation. The variation of the Pe and Dax values as given in Table 1 is of the same order as our estimated experimental accuracy. However, if we do assume the differences to be statistically significant, two lines of reasoning may be applied, leading to different conclusions. On the one hand, Dax for R101 is slightly higher than for R104 and R101+102 and is also slightly higher than for the complete reactor train, indicating a slightly higher dispersion for R101. So this would suggest higher R101 to have slightly higher backmixing, but this leaves the higher Dax for R102 unexplained. On the other hand, one could assume that due to the decreasing liquid velocity Dax would have a decreasing trend from R101 to R104 if these vessels were the
same. In that case, a higher value for R101 than for R101 could be expected if the lower number of baffles did not have effect. The fact that Dax for R101 is found to be lower rather than higher suggests that in R101 the lower number of baffles has slightly reduced the backmixing. This second line of reasoning is supported by the fact that, if one assumes Dax linearly increases with the liquid velocity, the results of R102 and R104 are fully consistent (Pe equal despite differences in ). In that case, one would conclude from the higher Péclet number for R101 that the lower number baffles in R101 has, if anything, decreased rather than increased the backmixing.
3. Gas phase tracer experiments It is important that oxygen is well-distributed, as a main side reaction of EB oxidation is decomposition of the product EBHP, which is enhanced under the so-called “oxygen starved” conditions. Therefore, air is introduced along the bottom of the reactors via middle and side spargers. The airflow through each of these can be regulated separately. Because the selectivity observed in the commercial plants falls up to a few percent short of what is achieved at ideal laboratory conditions, it was hypothesised that this may be due to “starvation” in poorly aerated regions of the large reactors. Starvation is defined as the presence of reaction zones in which the dissolved oxygen concentration—and thus the EB oxidation rate—is virtually zero. For a better understanding and quantification of this starvation phenomenon we modelled the complex hydrodynamics including mass transfer and chemical reaction with CFD. For this we used the discrete particle model of FLUENT 6.× along with a k. turbulence model. To reduce the computational effort a 2D geometry model was applied. A uniform bubble size was
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
5499
Fig. 5. CFD predictions: (a) liquid velocity field (m/s); (b) gas hold-up (kg/m3 ); (c): O2 concentration in the liquid (mol/m3 ); (d) O2 consumption rate (kmol/s m3 ).
used because preliminary simulations with different bubble size distributions had showed no significant effects on the hydrodynamics. In-house derived mass transfer correlations were used to describe the gas to liquid oxygen transfer. Simplified reaction kinetics for O2 consumption was adapted. An instantaneous EB flash was taken into account. The key CFD predictions are illustrated in Fig. 5, showing the presence of very strong liquid circulation patterns in the transverse direction and a fast upward flow of gas bubbles along quite narrow paths with very high gas hold up. For this specific case the gas is predicted to mainly rise through two relatively narrow paths near the walls, and gas injected in the middle spargers is predicted to predominantly rise sideways, i.e. the liquid in the centre of the reactor is not effectively aerated and is predicted to suffer from starvation (Figs. 5c and d). To validate the model predictions and to assess the scope for improvement, we performed gas phase radioactive tracer experiments in one of our commercial plants during normal production. Tracer injections (41 Ar) were done separately in either the air feed at the side or the middle spargers. Using 14 detectors mounted on the reactor wall, see Fig. 6, we could obtain a tracking of the gas tracer (from which among other things the gas rise velocities and the circulation patterns were determined. In Fig. 7 the raw data of the tracer signals are shown for both runs. In Fig. 8 the breakthrough times per detector are shown graphically.
Fig. 6. Schematic drawing of the commercial reactor with the 14 detector locations (heat exchange coils not shown).
3.1. Discussion of gas phase results at standard operating conditions First, and maybe foremost, these results confirm the key prediction of the CFD model and the assumed cause of the
Counts/sec
5500
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
7000 6000 5000 4000 3000 2000 1000 0
5 6 7
Counts/sec
11 2000 1800 1600 1400 1200 1000 800 600 400 200 0
Side injection
4
9 8
13
15
8
17
19
21
middle injection
2
3
10
4 10
12
14
16
25
20
22
10 8
5
23
9
18
Time (s)
Fig. 7. Detector responses for side and middle tracer injections.
Fig. 9. Bubble trajectories of gas from the middle and side spargers: comparison CFD prediction (blue and red lines) and tracer experiments (yellow and green dots) calculated with Eq. (2).
Fig. 10. Response of detectors 4 and 12 (opposites) after injection in the right side sparger.
Fig. 8. Breakthrough times per detector.
“starvation”: gas injected in the middle sparger is detected flowing upwards near the side of the vessel. From the data average gas velocities of 2.5 and 0.8 m/s were calculated for side and middle injection, respectively. For reference, the natural single bubble rise velocity is only of the order of ∼ 0.1 m/s. The significantly different velocities found for the side and middle injected gas tracer pulses indicate separate, and largely segregated, paths for these gas flows. Remarkably, this was also predicted by our CFD model, see Fig. 9, that shows the particle tracking of gas from the side and middle spargers indicated in red and blue, respectively. Quantitatively, the experimentally determined breakthrough times at the gas/liquid interface, 2 and 4 s for side and middle injection,
respectively, match very well those modelled with the CFD model. The detectors at the side of the vessel opposite to where injection took place show much higher “breakthrough times” 13–15 s (see Fig. 8). The responses of detectors 9 and 14, located at opposite positions at the gas/liquid interface, confirm the prediction that the gas from the side predominantly flows towards the interface along the side of the reactor and not via the middle. In the latter, detectors 9 and 14 would have observed the tracer more or less at the same time. This is furthermore confirmed by the fact that the tracer released from the (right-hand) side sparger is observed by the detectors 4–10 in increasing order of time. The response of detector 12 located opposite to detector 4 shows a curve with a maximum that is similar to the second maximum of the response curve of detector 4 suggesting a recycling period of ∼ 18 s, see Fig. 10. In order to better comprehend the recorded signals, a simple data-acquisition model accounting for solid angle of acceptance and attenuation (homogeneous fluid density) effects for a point
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
-source (representative for all injected activity being present in one ‘air bubble’) in the reactor has been set up. Relative signal responses have been computed starting from detector 4 for a vertical and an along-wall source trajectory. A vertical displacement indicates an intense signal at detector 4, quickly decreasing in intensity and increasing in width for detectors 5–7 and then slowly increasing in intensity and decreasing in width for detectors 8–10. For an equidistant movement along the reactor wall, all detectors would record signals with similar intensities and widths. The observed intensity and width distribution of the detector responses for side sparger injection is in agreement with the displacement of limited volume of air more or less in vertical direction and in not in line with a homogeneous air distribution. The relationship between -source intensity and signals registered by the individual detectors has been used to back-project the position of the air jet flow by assuming this flow is a moving point -source onto the radial towards the detector. The relationship used is Id# = C · exp(i − i · xid# )/(xs−d# )2 ,
5501
As described above, a point -source (representative for all injected activity contained in one “air bubble” reflecting nondispersive plug flow) was used to compute the flow path of the “air bubble” inside the reactor. It should be noted, however, that the recorded signals do show considerable broadening, which may be caused by: (a) the conical-widening field-of-view of the detectors; (b) that not all activity is contained in one bubble, but in a collection of bubbles; (c) dispersion. In this respect our simplifying approach should be considered as no more than an attempt to optimise our qualitative analysis exploiting the available experimental data. In summary, despite the crude method applied for computation of the flow paths from the radiotracer data, the overall agreement between the results of the simplifying point -source interpretative model and the CFD simulations is striking, including the largely segregated flow paths of the gas from the side and middle spargers.
(2)
where C is an arbitrary computational calibration constant (m2 ), Id# the maximum signal registered by detector number # (c/s), i the mass attenuation coefficient for the medium (i) (m−1 ), xid# the thickness of the medium (i) in view direction of the detector (m), and xs−d# the source-to-detector-# distance (m), which is equal to i xid# . The computed experimental flow paths are depicted in Fig. 9 along with the CFD prediction of the bubble tracks. The flow path for side sparger injection indicates that after initially flowing somewhat towards the reactor wall, the air flows more or less vertically towards the level interface. The computed velocities are high near detector 5 (∼ 5 m/s) slowing down to ∼ 2 m/s in between detectors 5 and 8, while it increases to 3–6 m/s in the gas cap (detector 9 to 10). The small fraction flowing from the injection point into the bottom section of the reactor moves very close along the reactor wall. The computed velocity in traversing from detector 4 to 2 is ∼ 0.5 m/s. The shape of the recorded signal for detector 2 supports the flow closely along the wall. The flow path for middle sparger injection has initially a substantial horizontal component; above the side sparger it starts to flow more vertically. The velocities between detector 2 and 3 are high (∼ 6 m/s) slowing down in the middle part to ∼ 1 m/s, and increasing to ∼ 6 m/s in the gas cap. The average gas velocity for side sparger injection is significantly higher than for middle sparger injection; under normal operating conditions approximately 2 vs. 1 m/s, respectively. This again strongly suggests that air bubbles from the side sparger and middle sparger follow distinct paths through the reactor: air bubbles emerging from the side sparger predominantly do not mix with air bubbles from the middle sparger, each are largely following their own, distinct path towards the liquid/gas interface. In comparison with the “jet” stream predicted by CFD the positions of the experimental flow path points are somewhat skewed towards the reactor wall.
3.2. Results at varying air distribution ratios Four dedicated tracer experiments were done at two different distribution ratios in which the air flow to the middle and side spargers was increased whilst keeping the total airflow constant. When the gas flow to the side spargers was reduced by 20% the oxygen uptake increased by roughly 1%. When the gas flow to the side sparger was further decreased (60% extra to middle sparger), the oxygen consumption decreased slightly compared to normal operation. In that latter case, the breakthrough times observed for side sparger injections increased from 1.3 to 2.3 s and the velocities decreased from 1.7 to 1.1 m/s. The breakthrough times observed for middle sparger injections increase from 4.3 to 5.3 s, without showing a significant effect on the average velocity in the middle of the reactor. Compared to the normal case, the air replacement from the side to the middle spargers results in an overall longer residence time of the air. Also for the changed air ratio cases the flow paths were computed using the same approach as described above. Below we briefly elaborate on this. Extra air in the middle sparger enters the reactor somewhat deeper; in the 60% case the middle flows end higher up in the fluid and then quickly moves nearer the reactor wall. This would lead to a less efficient use of reactor volume and more starvation explaining the observed decrease in oxygen uptake. In the 20% change case the middle and side flows seem to be more separated. This would result in a better aeration of the fluid and subsequently less starvation. This might explain the observed increase in oxygen uptake. 3.3. Plant changes As a consequence of this work we developed, again with the aid of the gas tracer experiments validated CFD model, a number of potential measures to reduce the “starvation” predicted to occur for certain configurations and severe operating
5502
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
single reactor are around 6 (moderate backmixing). For the complete reactor train a Péclet number of 25 was found. • The application of lower number of transversal baffles in the first reactor does not seem to have a significant effect. • With these liquid RTD results, the scope for improvement could be assessed with dedicated proprietary integrated plant models. The following can be concluded from gas phase radiotracer measurements:
Fig. 11. Comparison of commercial reactor performance before and after remedial action (hardware changes).
conditions. Options included repositioning of the gas inlets and installing additional baffles at specific locations, see also our patent application Hollander et al. (2006). After implementation of the optimum configuration in a number of commercial reactors, the reaction temperature could be lowered significantly (when compared at equal oxygen consumption rates), see Fig. 11. This translates to improved yields and/or increased production rates. 4. Conclusions Rather unique gas and liquid phase reactive tracer experiments have been performed in a commercial horizontal bubble column reactor. The following can be concluded from the liquid phase radiotracer measurements: • As opposed to the tanks-in-series model, the axial dispersion model can fit the liquid phase residence time distribution curve very well. The associated Péclet numbers for a
• The experiments confirmed the key CFD prediction that the major part of the gas injected at the middle does indeed largely rise sideways, i.e. mostly near the wall rather than through the middle. This is believed to be the main cause for potential oxygen starvation. • The bubbles injected at the side rise about twice as fast as the bubbles injected near the centre, implying that these two—close to each other—“plumes” are largely segregated. This as well as the derived bubble trajectories and circulation patterns are consistent with CFD model predictions. • These conclusions are consistent with the theory that— especially under severe conditions—there can be a starvation zone in the middle region. • This work has led to concrete plant changes resulting in a lower reaction temperature, improved selectivity and plant yields. References Hollander, E.D., Klusener, P.A.A., Ploemen, I.H.J., Schellekens, C.J., 2006. Horizontal reactor vessel. Patent WO2006024655. Krzysztoforski, A., Wojzit, Z., Pohorecki, R., Baldyga, J., 1986. Industrial contribution to the reaction engineering of cyclohexane oxidation. Industrial Engineering Chemistry Process Design and Development 25 (4), 894–898. Pohorecki, R., Baldyga, J., Moniuk, W., Podgorska, W., Zdrojkowski, A., Wierzchowski, P.T., 2001. Kinetic model of cyclohexane oxidation. Chemical Engineering Science 56, 1285–1291. Tilton, J.N., Russell, T.W.F., 1982. Designing gas-sparged vessels for mass transfer. Chemical Engineering 89 (24), 61–68. Zuiderweg, F.J., Bruinzeel, C., 1971. Liquid mixing and oscillations in sparged horizontal cylindrical vessels. Proceedings of the European Symposium of Chemical Reaction Engineering, Fourth Meeting, 1968, pp. 183–189.
Horizontal cross-flow bubble column reactors: CFD and validation by plant scale tracer experiments P.A.A. Klusener, G. Jonkers, F. During, E.D. Hollander, C.J. Schellekens, I.H.J. Ploemen, A. Othman, A.N.R. Bos ∗ Shell Global Solutions International B.V., P.O. Box 38000, 1030 BN Amsterdam, The Netherlands Received 16 June 2006; received in revised form 30 March 2007; accepted 31 March 2007 Available online 13 April 2007
Abstract Unique gas and liquid phase radioactive tracer experiments have been performed in horizontal cross-flow bubble column reactors of a commercial ethylbenzene oxidation plant. These experiments validated CFD modelling predictions, which in turn formed the basis for evaluation and selection of potential measures to reduce “oxygen starvation” predicted to occur for certain configurations and operating conditions. After implementation in commercial reactors lower operating temperatures, improved selectivity and thus better plant yields were achieved. 䉷 2007 Shell Global Solutions International B.V. Published by Elsevier Ltd. All rights reserved. Keywords: Bubble column; Radioactive tracer; CFD; Ethyl benzene; Propylene oxide; PO; SMPO
1. Introduction Propylene oxide is a versatile chemical intermediate used in a wide range of industrial and commercial products. Current world production is over 6 million metric tons a year. While other processes exist, the Shell Chemicals companies have derived a strong competitive advantage by using and continually developing their proprietary SM/PO technology, a process in which propene and ethylbenzene (EB) are converted into propylene oxide (PO) and styrene monomer (SM), respectively. Worldwide there are now five world-scale SM/PO plants based on Shell Technology, the most recent one started up in China in 2006. The first step in this multi-step process is the air-oxidation of ethylbenzene to ethyl-benzene hydroperoxide (EBHP). This is performed by Shell in cross-flow operation in a series of large horizontal bubble column reactors with a very low aspect ratio: height 4–6 m, length 15–25 m. Typically, a reactor train consists of 4–5 of such horizontal columns, which are equipped with baffles and heating/cooling coils. Air is introduced via ∗ Corresponding author. Tel.: +31 20 6302887.
E-mail address: [email protected] (A.N.R. Bos).
separate middle and side sparger systems. The gas outlet stream contains besides unconverted oxygen a very significant amount of EB from evaporation/stripping and this EB is recovered in a condensing column and recycled to the reactor train. The literature on horizontal cross-flow bubble columns is rather limited, see e.g. Tilton and Russell (1982), Zuiderweg and Bruinzeel (1971); Krzysztofoski et al. (1986). Pohorecki et al. (2001) did mention tracer experiments in similar commercial reactors, but did not disclose any details. Within Shell we have developed computational fluid dynamics (CFD) models as well as integrated reactor models with complex reaction kinetics for these rather unconventional bubble columns. To validate some of the model predictions it was decided to perform radioactive tracer experiments in one of the commercial plants during normal operation. Two completely different types of tracer experiments, with completely different purposes, have been performed: (1) liquid phase and (2) gas phase tracer experiments. 2. Liquid phase tracer experiments The peroxidation of EB to EBHP is autocatalytic implying that a minimum concentration of the product EBHP is necessary
0009-2509/$ - see front matter 䉷 2007 Shell Global Solutions International B.V. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.ces.2007.03.044
5496
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
Fig. 1. Liquid phase tracer injection and detection points.
to initiate the reaction and that the reaction goes faster on increasing EBHP concentration. On the other hand, the higher EHBP concentration leads to increased EBHP decomposition, by-product formation and lower selectivity. For this reason backmixing is desirable in the front-end of the first reactor, whereas plug flow behaviour is desirable in the remaining reactor train. To facilitate this, the reactors are segmented by a few transversal baffles. However, in the first reactor of the train a lower number of such baffles are installed. The two main incentives for investigating the actual mixing in the reactor train are to determine: 1. The effectiveness of the transversal baffles in facilitating plug flow (“staging” by forming compartments in series) and subsequently the scope for improved selectivity by further reducing the deviations from plug flow. 2. The effectiveness of installing a lower number of such baffles in promoting backmixing in the first reactor and hence the scope for improvement by (further) enhancing backmixing here. Fig. 1 illustrates where the tracer injections and detections in the actual plant were positioned. The experimentally obtained RTD curves (count rate vs. time) were fitted with both the tanksin-series model and the axial dispersion model, see standard reaction engineering textbooks. Both models have basically two parameters: the mean residence time and the parameter describing the spread in residence time, i.e. the number of tanks N and the Péclet number, respectively. In principle, could be calculated for (accurately) known flow rates, volumes, levels and gas hold-ups. However, here is considered to be an unknown model parameter and is one of the two fit parameters. This is not only because of the presence of significant EB evaporation but also because the plant measured or calculated values in practice are often not very accurate. The time axis was set to zero at the time that the injected bolus of activity passed the detectors positioned at the reactor inlet. 2.1. Results of RTD model parameter estimation Fig. 2 shows the results of fitting the RTD data of R102 with the tanks-in-series model for N = 3, 4, 5 and 6. For each N the
average residence time was optimised. It can be clearly seen that this model cannot quite fit the data: either the left-hand or the right-hand side of the measured RTD curve can be modelled well, but not the full curve with one set of parameters. Fig. 3 shows the result of fitting the data of R102 using the axial dispersion model, using open–open boundary conditions to allow easy fitting in the time domain. A remarkably good fit was obtained and the resulting Péclet number was 6.6. In Fig. 4 the tracer had been injected into the inlet of R101 with measurement, respectively, at the outlets R101 and R102. Unfortunately, no detector was present in the outlet of R104 during these experiments, and for R104 we only have the signal recorded by the detector placed at the circumference of R104, i.e. mounted on the reactor vessel itself rather than on the outlet piping. So this is not really the RTD over the whole reactor train. Consequently, this R104 signal can only be used qualitatively for interpretational purposes. 2.2. Translation to closed–closed boundary conditions It is well-known that for a sufficiently high Péclet number (say Pe > 20) the differences between “open–open” and “closed–closed” (or Danckwerts) boundary conditions become relatively small, but for very low Pe (say Pe < 5) the differences can become considerable. Since the fitted values are intermediate, we therefore have estimated the Péclet numbers associated with the “closed–closed” boundary condition model. This was done as by equating the variance of the “open–open” (oo) model to that of the “closed–closed” (cc) model: 2 2Pecc − 2 + 2e−Pecc 1 2 8 × = + . Peoo 1 + 2/Peoo Pe2cc Pe2oo (1) The factor (1 + 2/Pe)2 arises to properly account for the inherently different normalisation times in the dimensionless equations for the variances as functions of the Péclet numbers. We believe this method is quite accurate except for very low Péclet numbers (say < 3; for Pe < 1 the method will surely fail). As expected, the Péclet numbers found for the “closed–closed” boundary conditions are somewhat lower: e.g. Peoo = 7.5 with “open–open” conditions becomes Pecc = 6.7 for “closed–closed” conditions. The latter Péclet numbers were used to correlate to the tanks-in-series model. 2.3. Discussion of the liquid phase results Table 1 summarises the results; also the “optimum” tanks-inseries fit results are given. In the last column the experimental mean residence time (exp ) is given, somewhat arbitrarily calculated from reactor liquid volume divided by the outlet flow, i.e. implicitly assuming EB evaporation to be instantaneous at the inlet) . In general, a good fit with exp was found. The cc values for the “closed–closed” model match the experimental figures much better than those of the “open–open” model, indicating that indeed the “closed–closed” axial dispersion model is more appropriate for the EB oxidation reactors.
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
Fig. 2. Experimental RTD for R102 and tanks in series model fits showing the inadequacy of this model.
Fig. 3. Experimental RTD for R102 and axial dispersion model fits.
Fig. 4. Measured and fitted RTD for injection in inlet R101 and detection in outlet R101, R102 and inside R104 and axial dispersion model fits.
5497
5498
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
Table 1 Summary of liquid phase RTD measurements Enclosed reactor system
Axial dispersion model “Open–open” fit
“Closed–closed” estimation
oo
cc
Peoo
(s) R101 R102 R104 R101 + 02 R101 + 02 + 03 + 04∗d R104∗e Correctedf R101 + 02 + 03 + 04
Tanks-in-series model
a
Pecc
b
(s)
Dax (m2 /s)
Exp
“Tank” fit Ncalc
c
N
N
exp
(s)
(s)
883 925 1084 2019 4209 643
7.5 6.6 6.6 14.4 22.4 3.4
1118 1207 1414 2299
6.7 5.7 5.7 13.6
0.088 0.096 0.080 0.084
3.8 3.4 3.4 7.3
5 4 4 8 12 3
1040 1134 1330 2229 4488 867
1063 1131 1361 2194
4650
25.6
5013
24.7
0.085
12.9
13
4951
4786
a
cc = oo (1+ b Derived from
2/Pefit ). Peoo using Eq. (1). c N = Pe /2 + 0.5. cc d Outlet R104 not measured; detector measuring inside of reactor! e Side detector data from injection in R104. f Calculated by subtracting the values for R104∗ (side detector) from R101 + 02 + 03 + 04∗ and ding those of R104 (outlet detector). The “closed–closed” estimates were subsequently calculated from the so obtained Peoo values.
For the full train of four reactors, only an estimate of the RTD can be made because no measurement results are available for injection into R101 and measurement at the outlet of R104. Hence, the measured value in Table 1 of Pe = 22.4 is too low. Assuming that Péclet numbers are approximately cumulative for sufficiently high Péclet, the Pe number for the full reactor train was calculated from the side R104 measurement at R101 injection and the Pe number obtained from R104 injection. This procedure provides a Peoo of 25.6 from which Pecc = 24.7 was derived; the latter corresponds to approximately 13 tanks-inseries (also obtained when fitting N). R101 contains fewer baffles and hence fewer compartments; therefore it might be expected (and actually it was hoped for) to have more backmixing in R101. However, the Péclet number of R101 is slightly higher—i.e. slightly less backmixing—than the number found for R102: Pe = 6.7 vs. Pe = 5.7. This would suggest that the transversal baffles do not significantly influence the backmixing, i.e. no significant staging is achieved with these baffles. However, to assess the effect of the number of baffles one should also look at the dispersion coefficients Dax rather than only at the Péclet numbers because the residence times of the reactors increase significantly from R101 to R104 due to EB evaporation. The variation of the Pe and Dax values as given in Table 1 is of the same order as our estimated experimental accuracy. However, if we do assume the differences to be statistically significant, two lines of reasoning may be applied, leading to different conclusions. On the one hand, Dax for R101 is slightly higher than for R104 and R101+102 and is also slightly higher than for the complete reactor train, indicating a slightly higher dispersion for R101. So this would suggest higher R101 to have slightly higher backmixing, but this leaves the higher Dax for R102 unexplained. On the other hand, one could assume that due to the decreasing liquid velocity Dax would have a decreasing trend from R101 to R104 if these vessels were the
same. In that case, a higher value for R101 than for R101 could be expected if the lower number of baffles did not have effect. The fact that Dax for R101 is found to be lower rather than higher suggests that in R101 the lower number of baffles has slightly reduced the backmixing. This second line of reasoning is supported by the fact that, if one assumes Dax linearly increases with the liquid velocity, the results of R102 and R104 are fully consistent (Pe equal despite differences in ). In that case, one would conclude from the higher Péclet number for R101 that the lower number baffles in R101 has, if anything, decreased rather than increased the backmixing.
3. Gas phase tracer experiments It is important that oxygen is well-distributed, as a main side reaction of EB oxidation is decomposition of the product EBHP, which is enhanced under the so-called “oxygen starved” conditions. Therefore, air is introduced along the bottom of the reactors via middle and side spargers. The airflow through each of these can be regulated separately. Because the selectivity observed in the commercial plants falls up to a few percent short of what is achieved at ideal laboratory conditions, it was hypothesised that this may be due to “starvation” in poorly aerated regions of the large reactors. Starvation is defined as the presence of reaction zones in which the dissolved oxygen concentration—and thus the EB oxidation rate—is virtually zero. For a better understanding and quantification of this starvation phenomenon we modelled the complex hydrodynamics including mass transfer and chemical reaction with CFD. For this we used the discrete particle model of FLUENT 6.× along with a k. turbulence model. To reduce the computational effort a 2D geometry model was applied. A uniform bubble size was
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
5499
Fig. 5. CFD predictions: (a) liquid velocity field (m/s); (b) gas hold-up (kg/m3 ); (c): O2 concentration in the liquid (mol/m3 ); (d) O2 consumption rate (kmol/s m3 ).
used because preliminary simulations with different bubble size distributions had showed no significant effects on the hydrodynamics. In-house derived mass transfer correlations were used to describe the gas to liquid oxygen transfer. Simplified reaction kinetics for O2 consumption was adapted. An instantaneous EB flash was taken into account. The key CFD predictions are illustrated in Fig. 5, showing the presence of very strong liquid circulation patterns in the transverse direction and a fast upward flow of gas bubbles along quite narrow paths with very high gas hold up. For this specific case the gas is predicted to mainly rise through two relatively narrow paths near the walls, and gas injected in the middle spargers is predicted to predominantly rise sideways, i.e. the liquid in the centre of the reactor is not effectively aerated and is predicted to suffer from starvation (Figs. 5c and d). To validate the model predictions and to assess the scope for improvement, we performed gas phase radioactive tracer experiments in one of our commercial plants during normal production. Tracer injections (41 Ar) were done separately in either the air feed at the side or the middle spargers. Using 14 detectors mounted on the reactor wall, see Fig. 6, we could obtain a tracking of the gas tracer (from which among other things the gas rise velocities and the circulation patterns were determined. In Fig. 7 the raw data of the tracer signals are shown for both runs. In Fig. 8 the breakthrough times per detector are shown graphically.
Fig. 6. Schematic drawing of the commercial reactor with the 14 detector locations (heat exchange coils not shown).
3.1. Discussion of gas phase results at standard operating conditions First, and maybe foremost, these results confirm the key prediction of the CFD model and the assumed cause of the
Counts/sec
5500
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
7000 6000 5000 4000 3000 2000 1000 0
5 6 7
Counts/sec
11 2000 1800 1600 1400 1200 1000 800 600 400 200 0
Side injection
4
9 8
13
15
8
17
19
21
middle injection
2
3
10
4 10
12
14
16
25
20
22
10 8
5
23
9
18
Time (s)
Fig. 7. Detector responses for side and middle tracer injections.
Fig. 9. Bubble trajectories of gas from the middle and side spargers: comparison CFD prediction (blue and red lines) and tracer experiments (yellow and green dots) calculated with Eq. (2).
Fig. 10. Response of detectors 4 and 12 (opposites) after injection in the right side sparger.
Fig. 8. Breakthrough times per detector.
“starvation”: gas injected in the middle sparger is detected flowing upwards near the side of the vessel. From the data average gas velocities of 2.5 and 0.8 m/s were calculated for side and middle injection, respectively. For reference, the natural single bubble rise velocity is only of the order of ∼ 0.1 m/s. The significantly different velocities found for the side and middle injected gas tracer pulses indicate separate, and largely segregated, paths for these gas flows. Remarkably, this was also predicted by our CFD model, see Fig. 9, that shows the particle tracking of gas from the side and middle spargers indicated in red and blue, respectively. Quantitatively, the experimentally determined breakthrough times at the gas/liquid interface, 2 and 4 s for side and middle injection,
respectively, match very well those modelled with the CFD model. The detectors at the side of the vessel opposite to where injection took place show much higher “breakthrough times” 13–15 s (see Fig. 8). The responses of detectors 9 and 14, located at opposite positions at the gas/liquid interface, confirm the prediction that the gas from the side predominantly flows towards the interface along the side of the reactor and not via the middle. In the latter, detectors 9 and 14 would have observed the tracer more or less at the same time. This is furthermore confirmed by the fact that the tracer released from the (right-hand) side sparger is observed by the detectors 4–10 in increasing order of time. The response of detector 12 located opposite to detector 4 shows a curve with a maximum that is similar to the second maximum of the response curve of detector 4 suggesting a recycling period of ∼ 18 s, see Fig. 10. In order to better comprehend the recorded signals, a simple data-acquisition model accounting for solid angle of acceptance and attenuation (homogeneous fluid density) effects for a point
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
-source (representative for all injected activity being present in one ‘air bubble’) in the reactor has been set up. Relative signal responses have been computed starting from detector 4 for a vertical and an along-wall source trajectory. A vertical displacement indicates an intense signal at detector 4, quickly decreasing in intensity and increasing in width for detectors 5–7 and then slowly increasing in intensity and decreasing in width for detectors 8–10. For an equidistant movement along the reactor wall, all detectors would record signals with similar intensities and widths. The observed intensity and width distribution of the detector responses for side sparger injection is in agreement with the displacement of limited volume of air more or less in vertical direction and in not in line with a homogeneous air distribution. The relationship between -source intensity and signals registered by the individual detectors has been used to back-project the position of the air jet flow by assuming this flow is a moving point -source onto the radial towards the detector. The relationship used is Id# = C · exp(i − i · xid# )/(xs−d# )2 ,
5501
As described above, a point -source (representative for all injected activity contained in one “air bubble” reflecting nondispersive plug flow) was used to compute the flow path of the “air bubble” inside the reactor. It should be noted, however, that the recorded signals do show considerable broadening, which may be caused by: (a) the conical-widening field-of-view of the detectors; (b) that not all activity is contained in one bubble, but in a collection of bubbles; (c) dispersion. In this respect our simplifying approach should be considered as no more than an attempt to optimise our qualitative analysis exploiting the available experimental data. In summary, despite the crude method applied for computation of the flow paths from the radiotracer data, the overall agreement between the results of the simplifying point -source interpretative model and the CFD simulations is striking, including the largely segregated flow paths of the gas from the side and middle spargers.
(2)
where C is an arbitrary computational calibration constant (m2 ), Id# the maximum signal registered by detector number # (c/s), i the mass attenuation coefficient for the medium (i) (m−1 ), xid# the thickness of the medium (i) in view direction of the detector (m), and xs−d# the source-to-detector-# distance (m), which is equal to i xid# . The computed experimental flow paths are depicted in Fig. 9 along with the CFD prediction of the bubble tracks. The flow path for side sparger injection indicates that after initially flowing somewhat towards the reactor wall, the air flows more or less vertically towards the level interface. The computed velocities are high near detector 5 (∼ 5 m/s) slowing down to ∼ 2 m/s in between detectors 5 and 8, while it increases to 3–6 m/s in the gas cap (detector 9 to 10). The small fraction flowing from the injection point into the bottom section of the reactor moves very close along the reactor wall. The computed velocity in traversing from detector 4 to 2 is ∼ 0.5 m/s. The shape of the recorded signal for detector 2 supports the flow closely along the wall. The flow path for middle sparger injection has initially a substantial horizontal component; above the side sparger it starts to flow more vertically. The velocities between detector 2 and 3 are high (∼ 6 m/s) slowing down in the middle part to ∼ 1 m/s, and increasing to ∼ 6 m/s in the gas cap. The average gas velocity for side sparger injection is significantly higher than for middle sparger injection; under normal operating conditions approximately 2 vs. 1 m/s, respectively. This again strongly suggests that air bubbles from the side sparger and middle sparger follow distinct paths through the reactor: air bubbles emerging from the side sparger predominantly do not mix with air bubbles from the middle sparger, each are largely following their own, distinct path towards the liquid/gas interface. In comparison with the “jet” stream predicted by CFD the positions of the experimental flow path points are somewhat skewed towards the reactor wall.
3.2. Results at varying air distribution ratios Four dedicated tracer experiments were done at two different distribution ratios in which the air flow to the middle and side spargers was increased whilst keeping the total airflow constant. When the gas flow to the side spargers was reduced by 20% the oxygen uptake increased by roughly 1%. When the gas flow to the side sparger was further decreased (60% extra to middle sparger), the oxygen consumption decreased slightly compared to normal operation. In that latter case, the breakthrough times observed for side sparger injections increased from 1.3 to 2.3 s and the velocities decreased from 1.7 to 1.1 m/s. The breakthrough times observed for middle sparger injections increase from 4.3 to 5.3 s, without showing a significant effect on the average velocity in the middle of the reactor. Compared to the normal case, the air replacement from the side to the middle spargers results in an overall longer residence time of the air. Also for the changed air ratio cases the flow paths were computed using the same approach as described above. Below we briefly elaborate on this. Extra air in the middle sparger enters the reactor somewhat deeper; in the 60% case the middle flows end higher up in the fluid and then quickly moves nearer the reactor wall. This would lead to a less efficient use of reactor volume and more starvation explaining the observed decrease in oxygen uptake. In the 20% change case the middle and side flows seem to be more separated. This would result in a better aeration of the fluid and subsequently less starvation. This might explain the observed increase in oxygen uptake. 3.3. Plant changes As a consequence of this work we developed, again with the aid of the gas tracer experiments validated CFD model, a number of potential measures to reduce the “starvation” predicted to occur for certain configurations and severe operating
5502
P.A.A. Klusener et al. / Chemical Engineering Science 62 (2007) 5495 – 5502
single reactor are around 6 (moderate backmixing). For the complete reactor train a Péclet number of 25 was found. • The application of lower number of transversal baffles in the first reactor does not seem to have a significant effect. • With these liquid RTD results, the scope for improvement could be assessed with dedicated proprietary integrated plant models. The following can be concluded from gas phase radiotracer measurements:
Fig. 11. Comparison of commercial reactor performance before and after remedial action (hardware changes).
conditions. Options included repositioning of the gas inlets and installing additional baffles at specific locations, see also our patent application Hollander et al. (2006). After implementation of the optimum configuration in a number of commercial reactors, the reaction temperature could be lowered significantly (when compared at equal oxygen consumption rates), see Fig. 11. This translates to improved yields and/or increased production rates. 4. Conclusions Rather unique gas and liquid phase reactive tracer experiments have been performed in a commercial horizontal bubble column reactor. The following can be concluded from the liquid phase radiotracer measurements: • As opposed to the tanks-in-series model, the axial dispersion model can fit the liquid phase residence time distribution curve very well. The associated Péclet numbers for a
• The experiments confirmed the key CFD prediction that the major part of the gas injected at the middle does indeed largely rise sideways, i.e. mostly near the wall rather than through the middle. This is believed to be the main cause for potential oxygen starvation. • The bubbles injected at the side rise about twice as fast as the bubbles injected near the centre, implying that these two—close to each other—“plumes” are largely segregated. This as well as the derived bubble trajectories and circulation patterns are consistent with CFD model predictions. • These conclusions are consistent with the theory that— especially under severe conditions—there can be a starvation zone in the middle region. • This work has led to concrete plant changes resulting in a lower reaction temperature, improved selectivity and plant yields. References Hollander, E.D., Klusener, P.A.A., Ploemen, I.H.J., Schellekens, C.J., 2006. Horizontal reactor vessel. Patent WO2006024655. Krzysztoforski, A., Wojzit, Z., Pohorecki, R., Baldyga, J., 1986. Industrial contribution to the reaction engineering of cyclohexane oxidation. Industrial Engineering Chemistry Process Design and Development 25 (4), 894–898. Pohorecki, R., Baldyga, J., Moniuk, W., Podgorska, W., Zdrojkowski, A., Wierzchowski, P.T., 2001. Kinetic model of cyclohexane oxidation. Chemical Engineering Science 56, 1285–1291. Tilton, J.N., Russell, T.W.F., 1982. Designing gas-sparged vessels for mass transfer. Chemical Engineering 89 (24), 61–68. Zuiderweg, F.J., Bruinzeel, C., 1971. Liquid mixing and oscillations in sparged horizontal cylindrical vessels. Proceedings of the European Symposium of Chemical Reaction Engineering, Fourth Meeting, 1968, pp. 183–189.