Optimization of the Mizoroki-Heck reaction using Design of Experiment (DoE)

Article abstract of DOI:10.1021/op058013q

The influence of the variables concentration, temperature, time, and amounts of base and palladium on the yield of a Heck reaction was investigated using statistical methods (DoE) in a systematic sequential study. This revealed that temperature, concentration, and amount of palladium were the most important variables with a factor interaction between temperature and concentration. Through this approach, the yield of the Heck reaction was improved from 57% to 98% (89% isolated) with concomitant reduction of palladium loading from 3% to 0.5%.

Full text of DOI:10.1021/op058013q

Organic Process Research & Development 2006, 10, 64−69
Optimization of the Mizoroki-Heck Reaction Using Design of Experiment (DoE)
Varinder K. Aggarwal,*,† Anne C. Staubitz,† and Martin Owen‡
UniVersity of Bristol, School of Chemistry, Cantock’s Close, Bristol, BS8 1TS, UK, and GlaxoSmithKline, SteVenage, UK
Abstract:
The influence of the variables concentration, temperature, time,
and amounts of base and palladium on the yield of a Heck
reaction was investigated using statistical methods (DoE) in a
systematic sequential study. This revealed that temperature,
concentration, and amount of palladium were the most impor-
tant variables with a factor interaction between temperature
and concentration. Through this approach, the yield of the Heck
reaction was improved from 57% to 98% (89% isolated) with
concomitant reduction of palladium loading from 3% to 0.5%.
Introduction
We recently required large amounts of 3-phenylcinna-
maldehyde, which, because of its high cost, we elected to
synthesize using the Heck reaction of iodobenzene with
cinnamaldehyde. Such a reaction had been reported previ-
ously by Cacchi et al. but gave a mixture of starting material
1 and product 3 in a ratio of 41:59 (Scheme 1).1
Figure 1. Comparison between the one-factor-at-a-time study
and a full factorial study.
Scheme 1. Heck reaction for the preparation of 3
Under their conditions we obtained similar results, but
attempts to increase conversion by increasing the amount of
palladium catalyst, base, and PTC were fruitless. Since a
large number of variables can impact the success of this
reaction (and palladium reactions in general), we decided to
optimize the process by statistical methods (Design of
Experiment, DoE).2 Such methods are often practiced in
industry for reaction optimization but very rarely in aca-
demia, where a one-factor-at-a-time approach is generally
used. Although this intuitive approach can directly reveal
how certain factors affect yield, it covers the parameter space
very poorly and also reveals nothing about the interaction
of certain factors. In contrast, using a DoE approach, the
reaction space can be explored more efficiently, looking at
as many dimensions as there are variables at the same time.
Thus, a larger response space volume is covered, and at the
same time factor interactions can be revealed (Figure 1).
Also, several responses such as yield, residual starting
material, and generated impurities can be modeled simulta-
neously at no extra experimental cost.
factorial study) which in many cases is refined by a more
detailed second nonlinear study, such as a response surface
study (RSM). By placing the settings for the variables to be
examined of a factorial study at the extremes (low and high
values) of the part of the parameter space one wishes to
explore, a very large “volume” of this space is covered.
We thought that the variables concentration, temperature,
time, amount of catalyst, and amount of base could all have
an impact on yield (Table 1).3 The rationale behind the
selection of the ranges of the factors tested was that they
should be sufficiently wide in order to get a clear picture of
how these factors influence the conversion. Furthermore,
when the values for the signal cover a wide range, it is more
likely that the signal/noise ratio is acceptable even if the noise
level is high. On the other hand, this range was restricted
by practical considerations:
In a sequential experimental design, the initial design will
usually be a screening study (also referred to as a linear
(1) The Volume of the solVent could not be more than
5 mL because above this value sufficient heat transfer to
the vessel in the parallel reactor could not be guaranteed.
(2) The temperature was set between 50.0 °C and
90.0 °C, to ensure that the reaction took place and stayed
below the upper limit of the catalyst stability.4
* To whom correspondence should be addressed. E-mail: v.aggarwal@
bristol.ac.uk.
† University of Bristol.
‡ GlaxoSmithKline.
(1) Amorese, A.; Arcadi, A.; Bernocchi, E.; Cacchi, S.; Cerrini, S.; Fedeli, W.;
Ortar, G. Tetrahedron 1989, 45, 813.
(2) For an introduction into DoE, see for factorial designs: (a) Box, G. E. P.;
Hunter, W. G.; Hunter, J. S. Statistics for Experiments - An Introduction
to Design, Data Analysis, and Model Building; Wiley: New York, 1978.
For response surface models, see: (b) Box, G. E. P.; Draper, N. R. Empirical
Model-Building and Response Surfaces; Wiley: New York, 1987.
(3) Solvent, type of catalyst, and base might also have an effect. There are
methods such as principle component analysis (PCA) to convert categorical
variables into numerical variables, and had the current study not been
successful, these would have been studied.
64
•
Vol. 10, No. 1, 2006 / Organic Process Research & Development
10.1021/op058013q CCC: $33.50 © 2006 American Chemical Society
Published on Web 11/30/2005

25 experiments are in principle required to resolve all effects
by factors and factor interactions unambiguously.6 However,
as indicated above only a fraction (half or quarter) of these
are required to generate a sufficient level of understanding
of the factors and their interactions.
Table 1. Ranges chosen for factors in the screening study
(half factorial)
factor
values chosen
initial values
volume/mL
T/°C
catalyst/equiv
base/equiv
time/h
1.6 to 5.0
2.0
60.0
0.030
1.7
50.0 to 90.0
0.005 to 0.050
1.5 to 4.0
Results and Discussion
We elected to carry out a half factorial design with
4.0 to 9.0
3.5
additional centre points.7 These are design points that are
placed in the center of the “volume” of the parameter space
one wishes to explore and are not included in the model.
Rather, they serve to determine the signal-to-noise ratio and
can show whether curvature occurs, i.e., when they are not
on the straight line connecting two design points, but
significantly above or below.
(3) Catalyst loading: As the problem with this reaction
might be the precipitation of the catalyst, the lower boundary
was set very low. The upper value also reflected that the
catalyst is expensive and therefore should be minimized as
well.
(4) Quantity of base: The base had been varied by Cacchi
et al., so the values chosen for this reflect a distribution of
those used in the paper.5
These considerations led to a set of 20 reactions which
were performed in parallel and yields were determined by
GC-MS (calibrated). The results of this study can be
visualized in a half normal plot8 (Figure 2) where a
statistically significant effect is shown by a factor or factor
interaction straying from the indicator line.
(5) Time: It was not necessary to reduce the reaction time
as this factor does not present a limitation in the laboratory.
However, by including the factor time, information is gained
on when the reaction is complete and the possible degradation
of product. To establish when the samples should be taken,
the reaction was run under the original conditions and
samples were taken after 2 h 45 min, 4 h 30 min, and 8 h.
There was no further conversion observed after 8 h. To be
sure that in more extreme cases the end of the reaction could
be observed, a longer time of 9 h was taken as the longest
reaction time.
As strong factor interactions were suspected, a half
factorial screening design was deemed appropriate. A full
factorial on the other hand did not seem necessary because,
with the half factorial, only three and four factor interactions
are unresolvable. These are generally rare and are not likely
to be significant. To get a clearer picture of the precision
and the accuracy of the linear model, four centre points were
run, with two together at the beginning of the study and two
together at the end (to guard against reactor or time related
trends).
All reactions were run in a parallel reactor. This not only
allowed us to run a large number of reactions at the same
time (though each run was constrained to be at the same
temperature) but also helped to reduce any experimental bias,
because the experimental setup was essentially the same for
each experiment.
For statistical parameter optimizations it is vital that all
experiments are performed under exactly the same conditions
apart from those intentionally varied. In this case, this was
done by using standardized reaction vessels in a parallel
reactor and by always using the same batch of chemicals.
Another important factor is randomization, wherever possible
(e.g., in allocation to vessels or order of experiments and
chemical analysis), to eliminate bias. All reactions were
performed on a 1.0 mmol scale with respect to the aldehyde.
With five variables and high and low values for each of them,
From this it is clear that the most important factor is
temperature (furthest away from the indicator line), followed
by catalyst loading, concentration, and an interdependence
of temperature and concentration (Figure 2). The finding of
a significant interaction between two variables demonstrates
the superiority of DoE compared to a one-factor-at-a time
study where the interdependence of factors would have been
missed. The amount of base and time of reaction9 had no
effect on yield. Interaction between variables can also be
represented graphically by plotting yield against one of the
variables overlaid with a second variable (Figure 3). If the
lines run in parallel, there is no interaction between the two
variables (temperature and catalyst loading), but if the lines
divert, there is an interaction (volume and temperature).
Figure 4 shows the calculated model (2FI ) Two Factor
Interaction Model) for the yield in a cubic graph or a 3D
surface graph. The four center-points gave yields between
44% and 50% which were above the value predicted by the
linear 2FI model, thus indicating significant curvature. This
was evidence that while the 2FI could be used as a rough
guide to predict trends in the reaction, a more detailed study
allowing for nonlinear terms would be needed to model the
real situation more accurately. The fact that the center points
give a rather broad range for the yield indicates that the
reaction is very sensitive in this area to only subtle changes
in the reaction conditions.
The following general conclusions were made:
(1) A higher temperature (variable D) results in higher
yields, especially when the volume (variable A) is high
(6) As the number of variables under consideration increase, the possible
permutations of high and low settings increase rapidly.
(7) The calculations and randomisations for this optimisation were done using
a commercially available computer program: DesignExpert, version 6.0.10;
StatEase, Inc.: 2021 East Hennepin Ave., Suite 480, Minneapolis, MN
55413, USA, 2003.
(8) A half normal plot is a tool to visualize whether a factor or factor interaction
has a significant effect. The red line is used to denote the factors that do
not lead to a significant effect and follow the normal distribution. The further
a factor falls to the right away from the line, the more pronounced the effect.
(9) Reactions were run for 4 and 9 h and showed no further conversion at the
first time point. Product decomposition was not evident.
(4) Initially, we suspected that the catalyst stability would be reduced at higher
temperatures. However, we were proven wrong, as the results of the factorial
and later of the RSM study clearly showed.
(5) The rationale behind the values chosen in the paper is not clear since they
vary with different bases.
Vol. 10, No. 1, 2006 / Organic Process Research & Development
•
65

Figure 2. Most important factors determining the yield of the Heck reaction.
Figure 3. Interaction graphs for the half factorial screening study.
(factor interaction AD).
conditions (see RSM study below) extending time had an
(2) Higher catalyst loading (variable B) gives higher
yields, an effect which is more pronounced at higher
temperatures. This effect is dependent on the volume and
more pronounced if the volume is high.
(3) Equivalents of base (variable C) are virtually nonin-
fluential (i.e., not significant) under all conditions tested and
will not be included in further studies. Time (variable E) on
the other hand was retained as a variable even though it was
not significant in this study. Since the reaction did not go to
completion, we wished to see if under different reaction
effect, i.e., higher conversions or product degradation.
(4) The curvature found is significant. This is a clear
indication that the linear assumptions made in a factorial
model are not descriptive of the real response surface. Thus,
any model derived from this can only give a rough indication
for point predictions within the chosen boundaries. However,
the overall trends as indicated by the values at the corners
of the cube are valid (Figure 4).
We concluded from this study that high temperatures were
required, probably higher than those we had set initially. At
66
•
Vol. 10, No. 1, 2006 / Organic Process Research & Development

Figure 4. Cubic and 3D view of the model for the half factorial model.
this stage it seemed by extrapolation of the model that even
better results might be obtained if temperature, dilution, and
catalyst loading were increased. However, we were keen to
keep loading of the costly Pd catalyst to a minimum. In
general, high catalyst loading gave higher yields than low
loading, but this effect was smaller at higher dilution. Thus,
it was possible that, with further dilution and by operating
at the optimum temperature, catalyst loading could be
reduced without impacting negatively on yield.
Therefore, we initiated a second study, exploring higher
dilutions together with higher temperatures, while again
examining catalyst loading and time as a factor.
Since factorial studies are intrinsically linear (they only
have two levels), they connect all points with straight lines
and cannot model a mathematical optimum. When the center
points are indicative of significant curvature, then the
approximate linear model of the reaction can be refined by
a more detailed second nonlinear study, such as a response
surface study (RSM). RSM designs are particularly useful
for process optimizations and are often used to augment a
factorial design, if it shows significant curvature, to get a
detailed image of the response surface. However, in this case,
it was clear that the optimum area to operate this reaction
was outside the one chosen for the factorial, as the highest
yield was found in a corner of the cubic graph. Therefore,
the RSM design chosen was set up to be analyzed indepen-
dently from the screening study, i.e., without using those
values for the model. The points for the RSM model were
chosen according to a central composite design (CCD). This
design provides estimations for all coefficients in a quadratic
model, while still employing relatively few experiments. To
estimate coefficients for second-order terms using a factorial
design, it would be necessary to carry out a 3k factorial
design, with k ) number of variables tested. This would lead
to a large number of experiments, even if k were to be only
a small number. To reduce this experimental expenditure,
Figure 5. Distribution of experiments for an RSM study in
the parameter space.
the CCDs are composed of 2k points in the design space
augmented by 2k points on the k axis, equally spaced at (R
units, and n center points (Figure 5). In the present study
we used a CCD, which comprises a factorial design with
additional axial points and six center points, leading to a set
of 30 experiments. It is important that n should be a
reasonably high number as this allows for good estimation
of curvature and pure experimental error variance. While this
is also true for pure factorial designs with center points, it
becomes more important for estimation of higher order terms.
The factors studied in the second study were volume,
temperature, catalyst loading, and time (Table 2):
(1) The Volume of the solVent clearly proved to be one of
the most important factors. Its range was set between 5 and
10 mL.10
(2) The temperature was set between 90.0 °C and
120.0 °C, as high-temperature had proved to be beneficial.
(3) Catalyst loading: From the factorial it was clear that
higher catalyst loading contributed to higher conversions.
But looking at the cubic graph, it seemed also possible to
achieve good results even with lower catalyst loading and it
Vol. 10, No. 1, 2006 / Organic Process Research & Development
•
67

Figure 6. Graphical analysis for the product.
Table 2. Values for the optimization (RSM) study
For the present study the value for R was set to be 1, so
that only three different temperatures instead of five had to
be taken into account (face centered CCD) minimizing the
number of times the parallel reactor block had to be used.
For points which differed only in respect to time, the same
sample was measured twice at the given times in order to
reduce the number of actual experiments. While this will
affect the validity of the statistical analysis (typically it will
underestimate the background noise thus identifying more
effects as significant than should truly be the case), it was
practically very desirable and thought unlikely to appreciably
distort the model. Analysis and calibration were done in the
same manner as that for the screening study.
Viewing the quadratic model for the yield as cubic and
3D surface graphs (Figure 6), the following conclusions can
be made:
(1) The higher the temperature, the better the yield,
especially if the volume is large.
(2) The higher the catalyst loading, at high volumes, the
higher the yield. BUT, if volume and temperature are high,
even a low catalyst loading leads to excellent yields.
best values from
the factorial optimization
(76%)
factor
values chosen
volumea)/mL
temperature/°C
catalyst/equiv
time/h
5.0 to 10.0
5.0
90.0
0.050
8.2
90.0 to 120.0
0.005 to 0.050
8.0 to 24.0
a This is given relative to the half factorial study for better comparison.
However, due to the dimensions of the reaction vessels in the carousel, every
reaction was run on half the quoted scale.
was desirable to find out how low it was possible to go while
still obtaining good yields. Therefore, the setting was again
between 0.005 equiv and 0.05 equiv.
(4) Time was included as a factor since, although in the
laboratory there is no strong need to reduce the time taken,
it is desirable to achieve complete conversion, yet higher
temperatures sometimes lead to decomposition of products
or starting materials. Therefore, the low level was set at 8 h,
and the high level, at 24 h.
Table 3. Verification experiments
catalyst
loading/
equiv
volume/
mL
expected
yield/%
GC-MS
isolated
yield/%
conditions
T/°C
yield/%a
original conditions
60
0.03
4
10
10
33b (59)c
56
59
58
39
41
40
(base, 1.7 equiv; t, 24 h)
conditions judging the trends
shown in the half factorial
(base, 1.5 equiv; t, 24 h)
120
120
0.05
>76
99
99
99
91
93
90
predicted best conditions after RSM
(base, 1.5 equiv; t, 24 h)
0.005
98d
98
98
98
88
90
89
a Calibrated. b From the 2FI model. c Isolated yield of Cacchi et al. d RSM model.
68
•
Vol. 10, No. 1, 2006 / Organic Process Research & Development

(3) Time is again not very important, apart from the fact
that some reactions were incomplete after 8 h.
Now, when reactions were conducted at 120 °C, very high
yields were obtained, even at 0.5% catalyst loading, provided
reactions were conducted at high dilution.
The models generated by both the screening and the
optimization studies are approximations of reality. They are
used to predict the favored settings and then these settings
can be implemented in order to validate the model.
on reaction optimization from literature conditions of a
closely related (or the same) substrate. However, the literature
conditions in our case were far from optimum. The high
temperature, low catalyst loading, and high dilution that were
found to be optimum are beginning to emerge as a common
set of conditions for the ligand-free Heck reaction, and we
believe these should now be the starting point for most
substrates.
Experimental Section
To substantiate the findings from both, the factorial and
the RSM study, three verification experiments were per-
formed three times: the original conditions, the conditions
thought best after performing the half factorial design, and
the conditions found in the RSM study (Table 3). Each of
these experiments was measured three times by GC-MS in
order to get a clearer picture of the precision of the analysis.
The optimum conditions arrived at through the statistical
approach (high temperature, low catalyst loading, high
dilution) can be rationalized in chemical terms. The high
temperature gives faster rates of all the steps involved in
the Heck reaction, and the low catalyst loading with high
dilution prolongs the longevity of the Pd catalyst, thus
resulting in full conversion. Indeed, the low catalyst loading
with high dilution will reduce the likelihood for the Pd (0)
nanoparticles,11 which may be the catalytically active species,
to conglomerate and precipitate out as unreactive Pd black.
In fact, the use of “homeopathic” quantities of Pd(OAc)2
has recently been advocated in related reactions.12
3,3-Diphenylacrylaldehyde 314 (Optimized Procedure):
A mixture of NaOAc (123.0 mg, 1.5 mmol), (nBu)4NBr
(322.4 mg, 1.0 mmol), iodobenzene (168 µL, 1.5 mmol),
and trans-cinnamaldehyde (126 µL, 1.0 mmol) in DMF
(7.5 mL) was heated to 120 °C. Pd(OAc)2 (1.1 mg, 5 µmol)
was dissolved in DMF (2.5 mL). After the catalyst had
dissolved completely, this solution was added to the reaction
mixture. After 24 h, the reaction mixture was cooled to rt
and poured on NaHCO3half-sat. (80 mL) and extracted with
EtOAc (3 × 40 mL). The combined organic phases were
washed with brine (40 mL), dried over Na2SO4, and
concentrated in the rotary evaporator. To remove further
residual DMF, the crude mixture was dried at 60 °C under
high vacuum. Then the product was purified by flash
chromatography (petrol ether/EtOAc ) 20/1). The product
1
was isolated as a yellow oil. H NMR (400 MHz, CDCl3):
9.53 (d, J ) 8.3 Hz, 1H, dCHsCHO), 7.54-7.19 (2 m,
10H, ArsH), 6.60 (d, J ) 8.3 Hz, 1H, dCHsCHO); 13C
NMR (100 MHz, CDCl3): 193.2, 162.0, 139.7, 136.7, 130.6,
130.4, 129.6, 128.6, 128.3, 127.3; MS (EI): 208 (M+, 72),
207 (100), 179 (35), 178 (53), 177 (10), 176 (13), 165 (16),
152 (13), 102 (32), 77 (10).
Conclusions
In conclusion, we have shown how DoE can be effectively
used to rapidly (3 weeks) optimize a complex chemical
process, the Heck reaction, involving five variables. Two
variables, concentration and temperature, were found to be
dependent on each other. The optimum conditions using low
catalyst loading were found to require high temperatures and
low concentrations. We believe that this results in rapid
reaction rates with minimization of precipitation of Pd black
through conglomeration of Pd (0) nanoparticles. Although
these conditions are close to the Jefferey’s conditions13
employed for other Heck reactions, the sensitivity of such
reactions to substrate structure invariably leads one to embark
Acknowledgment
We thank GSK (in particular Dan Tray and Marion
Chatfield for useful discussions and Andy Reason for
sponsoring a 2 day Experimental Design Workshop for the
Chemical Education Community). A.S. thanks the Fonds der
chemischen Industrie (FCI) for a Kekule´ fellowship.
Supporting Information Available
Tables showing individual results of runs, details of
statistical analysis, expanded discussions of peripheral find-
ings, experimental procedure, analytical data. This material
is available free of charge via the Internet at http://
pubs.acs.org.
(10) As all reactions had to be done in the parallel reactor, this meant that to
achieve this dilution the complete scale (and thus the volumes) had to be
divided by two. However, to give some comparability to the factorial design,
the tables will show the values in a molar volume scale (mL/mmol 18).
For simplicity, this will be hinted at in the Supporting Information.
(11) (a) Reetz, M. T.; Westermann, E. Angew. Chem., Int. Engl. Ed. 2000, 39,
165. (b) Reetz, M. T.; Lohmer, G. Chem. Commun. 1996, 1912.
(12) (a) de Vries, A. H. M.; Mulders, J. M. C. A.; Mommers, J. H. M.;
Henderickx, H. J. W.; de Vries, J. G. Org. Lett. 2003, 5, 3285. (b) Reetz,
M. T.; de Vries, J. G. Chem. Commun. 2004, 1559.
Received for review August 12, 2005.
OP058013Q
(13) For a discussion of Jefferey’s conditions, see: (a) Jeffery, T.; David, M.
Tetrahedron Lett. 1998, 39, 5751. (b) Jefferey, T. Tetrahedron 1996, 52,
10113.
(14) First characterization (microanalysis, mp): (a) Kohler, E. P.; Larsen, R. G.
J. Am. Chem. Soc. 1935, 57, 1448. For NMR, mass, IR, see: (b) Kurosawa,
K.; Tsujita, T. Bull. Chem. Soc. Jpn. 1981, 54, 2391.
Vol. 10, No. 1, 2006 / Organic Process Research & Development
•
69