Systematic Exploration of a Multi-Promoter Catalyst Composition
Space with Limited Experiments: Non-Oxidative Propane
Dehydrogenation to Propylene
Christian Kunkel, Frederik Ruther, Frederic Felsen, Charles W. P. Pare, Aybike Terzi,
Robert Baumgarten, Esteban Gioria, Raoul Naumann d’Alnoncourt, Christoph Scheurer,
Frank Rosowski,*and Karsten Reuter*
Cite This: ACS Catal. 2024, 14, 9008−9017
Read Online
ACCESS Metrics & More Article Recommendations *
sı Supporting Information
ABSTRACT: Promoters are indispensable for the optimized
performance and lifetime of industrial catalysts. Present-day
systems nevertheless benefit only from a small number of different
promoters, identified and often only locally optimized in laborious
empirical research. Here, we present an accelerated discovery
approach that globally explores a multipromoter design space with
only limited experiments. Cornerstones are an efficient iterative
design-of-experiment (DoE) planning of the measurements and a
throughput maximization through a parallelized testing protocol.
With less than 100 experiments conducted within weeks, we
identify a competitive promoter chemistry for the nonoxidative
propane dehydrogenation to propylene over alumina-supported Pt.
This discovery rests on an achieved deep understanding of the positive and negative actions of multiple promoters on the reaction
yield and deactivation. The iterative DoE strategy successively querying batches of experiments proves to be a powerful general
concept for data-efficient hypothesis validation and insight-based adaptation of design spaces.
KEYWORDS: high-throughput experimentation, automated laboratory, design of experiment, active learning,
non-oxidative propane dehydrogenation, heterogeneous catalysis, promoters
■INTRODUCTION
Promoters are a prevalent ingredient to industrial catalysis,
enhancing performance, improving selectivity, or mitigating
deactivation.
1−4
Various modes of action are known, such as
modifying the electronic structure, facilitating reactant
adsorption, or favoring specific reaction pathways.
5−8
These
actions extend over both the active material and the support
9,10
and depend sensitively on both type and concentration of the
employed promoter(s). Lacking a full mechanistic under-
standing of these complex actions, the additive selection of
suitable promoters and simultaneous optimization of concen-
trations are, in practice, a time-consuming empirical endeavor.
Rather than systematically exploiting the full combinatorial
multipromoter design space, industrial catalyst systems there-
fore typically feature only one or two promoters, at
concentrations that are only locally optimized. Due to limited
experimental budgets, the optimization is often even carried
out sequentially, i.e., the first promoter concentration is varied
around the optimized unpromoted system, the second
promoter concentration then varied around the thus
“optimized” one-promoter system, and so forth.
The nonoxidative propane dehydrogenation (PDH) to
propylene is a prime example for this common situation. As
one of the most important feedstock chemicals, e.g., polymer
manufacturing, propylene demand is expected to approach 200
megatons by 2030,
11
which cannot be met with the existing
cracking processes.
12
Industrial scale catalysts for the PDH
reaction usually build on oxidic support materials with ≤1 wt
% Pt or 18−20 wt % CrOxas active metals.
10,13
With the large-
scale application in mind, high surface area alumina (Al2O3) is
a preferred support choice following its thermal and
mechanical stability and comparably high density.
13
Unfortu-
nately, the process suffers from side reactions toward lower
alkanes, as well as severe coke formation, which leads to a fast
deactivation of the catalyst under operation conditions.
11,14,15
A working process therefore often incorporates regular
Received: March 21, 2024
Revised: May 17, 2024
Accepted: May 17, 2024
Published: May 29, 2024
Research Articlepubs.acs.org/acscatalysis
© 2024 The Authors. Published by
American Chemical Society 9008
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
This article is licensed under CC-BY 4.0
regeneration steps to remove the formed coke and restore the
initial activity of the catalytic system�up to a point, where
irreversible long-term deactivation (due to persistent sintering)
poses a challenge. To improve on this situation, a variety of
promoters have been studied with the goal to maximize
product yields or the overall efficiency of the process.
12,13,16
A
now commercially applied process (Oleflex) comprises the
ternary catalyst formulation Pt−Sn−K that emerged from such
studies throughout the last decades.
17−19
Indeed, more than 40
(!) years passed from the finding of Pt to be active in the
dehydrogenation of paraffins,
20−22
over first investigations on
the promoting effect of Sn,
17
to the first industrial application
starting from paraffins (Pacol-process),
23
and finally the
development of the Oleflex process.
Even so, not even this process achieves to push the yield
limitations of this system toward an operation at the
thermodynamic limit.
12
In order to ultimately achieve this
ambitious goal, the identification of new high-performance
multipromoter combinations and a deeper chemical under-
standing of the promoting effects are considered as key
factors.
12,16,24
Ideally, such promoters also suppress the
challenging reversible fast deactivation of the catalysts, leading
to a smart process that relies on fewer regeneration steps or, in
the ideal case, results in a continuous process without
regeneration at all. Here, we address this with an accelerated
discovery approach for complex multipromoted systems. This
approach systematically explores the corresponding high-
dimensional design space spanned by the employed promoter
species and their concentrations. Central pillars are efficient
experiment planning based on an iterative design of experiment
(DoE) approach as well as an experiment setup that holistically
considers the entire workflow from synthesis to catalytic
testing. The prior planning enables systematic promoter space
exploration at a number of samples explicitly to be studied that
is comparable to previous local and sequential optimizations.
The latter experiment setup then maximizes the throughput of
these studies by automating and efficiently parallelizing the
limiting steps in the workflow, e.g., through a dedicated time-
resolved testing protocol for a multireactor setup. At restricted
human intervention, the complete synthesis-to-testing for a
batch of eight samples can thus be achieved every 5 days.
We initially minimize an interfering influence of free support
sites and analyze the action of six potential promoters (Fe, Co,
Ga, Mn, K, and Ca; see below) on the PDH at Pt on θ−Al2O3
in the limit of surface-covering promotion. The global view of
this multipromoter design space reached after testing only five
batches reveals a Pareto front between the two central
performance descriptors, propylene yield and fast deactivation.
Subsequent batches designed to specifically assess the separate
actions of the promoters trace this Pareto dilemma down to
the individual promoter level. An improvement in one
performance descriptor brought about by the addition of any
promoter always comes at the cost of a deterioration in the
other descriptor, and there are no significant multipromoter
interactions that would allow to overcome this limitation.
Intriguingly, the six promoters fall into two groups with similar
actions though. While one group improves product yield at
increased fast deactivation, the other group reduces deactiva-
tion at reduced product yield. Naturally, best compromise
performance gains can therefore be obtained at the boundaries
of the design space, where at a reduced number of different
promoters, optimally tuned concentrations of ultimately only
one representative of each promoter group allow to maximally
compensate these opposing actions. Leveraging this deep
understanding, we therefore switch to a lower-dimensional
design space, where we only retain a most promising promoter
of each group and finally explore additional support site effects
by allowing for total concentrations below the surface-coverage
promotion limit. The global view of this bipromoted (Mn and
Ca)-space reached after only two further batches shows that
the intrinsic support activity essentially only shifts the Pareto
front. It does it to such an extent though that a performance
comparable to the established Pt−Sn−K reference can be
reached. With only a very limited number of experiments
performed within a matter of weeks, our accelerated discovery
approach thus rationally leads to a new promoter formulation
that is competitive to the one arrived at over decades of
empirical research.
■RESULTS
Delineation of the Multipromoter Design Space. In
the selection of the six promoter species investigated in our
approach, we can draw on an ever increasing variety of single-
promoter studies of Pt-based PDH catalysts in literature.
12,13
The reported elements with a potentially beneficial effect on
the catalysis are alkali and earth-alkali metals like K or Ca, 3d
transition metals like Mn, Fe, or Co, as well as main group
metals like Al, Ga, or Sn. The increased selectivities observed
upon the addition of alkali or earth-alkali metals are thereby
mainly assigned to their potential to moderate the strong acid
sites of the Al2O3support.
9,16,25−30
In contrast, the promoting
effect of the metals is more discussed in terms of an improved
Pt dispersion
31−34
or the formation of intermetallic or oxidic
compounds.
6,7,12,35−41
Aiming to cover all of these suspected
actions, we select two likely reducible (Fe, Co), two hardly
reducible (Ga, Mn), and two moderating elements (K, Ca) as
the promoters spanning the design space for a multipromoter
PDH system. From the two known Oleflex promoters (Sn and
K), the function of Sn is quite well understood in terms of Pt/
Sn nanoalloy or intermetallic formation. We therefore
deliberately excluded it from our study in favor of less well-
known promoters that also might form alloys. The role of K,
on the other hand, is less clear. Reports on its moderation of
the acidity of the alumina support consider only the lowest
concentrations. We therefore deliberately kept K in the study
to see if it has an additional effect if the concentration exceeds
the threshold needed for the moderation of acid sites.
Independent of this selection, we still use a Pt−Sn−K
formulation with concentrations mimicking the Oleflex system
as a benchmark catalyst to put the observed catalytic
performances to scale; see experimental section and the
Supporting Information for more details.
Pt and the promoters are applied to the thermally pretreated
θ−Al2O3support by incipient wetness impregnation. With
the objective to first analyze the Pt-promoter interactions at
minimized additional interactions with free support sites, we
pursue a surface-covering promotion concept and only
consider catalysts where the total loading of Pt and all
promoters arithmetically equals about one monolayer on the
surface of the support (determined to amount to 6 atoms per
nm2; see Supporting Information). Henceforth, we report the
respective surface concentrations Xi(i= Pt, Fe, Co, Ga, Mn, K,
and Ca) as fractions of this targeted total surface
concentration. We then have ∑iXi= 1.0 and the conversion
to nominal weight loadings is provided in Table S1. For Pt, we
investigate an economically viable range of 0.02 ≤XPt ≤0.06,
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9009
amounting to nominal Pt loadings between 0.10 and 0.29 wt
%. As pre-experiments indicated excessive coke formation and
reactor blocking, the maximum surface fractions of Fe and Co
need to be limited to 0.1, with an additionally imposed
multicomponent constraint XFe +XCo ≤4·XPt. At the same
time, we suspect nontrivial performance variations in the limit
of individually vanishing promoter concentrations. To ensure a
sufficient smoothness for the intended DoE-based exploration
of the design space, we therefore also set minimum surface
fractions of 0.02 for Fe and Co and 0.05 for all other
promoters. In consequence, this defines maximum concen-
trations for Ga, Mn, K, and Ca of 0.75.
Experiment Planning. Even though we formally vary the
concentrations of Pt and all six promoters, the mixture
constraint (∑iXi= 1.0) reduces this to a six-dimensional
design space.
42,43
Due to uncertainty in the experimental
synthesis, only variations with a minimum change of surface
fractions will reliably yield a new catalyst. If we loosely use 0.02
(0.05) as this minimum step size for Pt, Fe, and Co (Ga, Mn,
K, and Ca) and account for the above-mentioned multi-
component constraint for XFe +XCo, this leads to a total of
7265 unique catalysts that can in principle be synthesized in
the defined design space. While not excessively large in
absolute numbers, mapping this space exhaustively is
essentially not tractable. Even in the parallel experiment
setup further detailed below with a throughput of eight
samples per 5 days, this would amount to a continuous
experimentation over more than 12 years. We therefore follow
a DoE philosophy to determine the response surface in terms
of propylene yield and deactivation over the design space from
a set of carefully planned measurements.
For an initial exploration that avoids an overly reliance on
assumptions like a smoothly varying response surface or the
absence of significant promoter interactions, we first developed
a problem adapted screening, addressing at the same time
blocking and mixing constraints, that yields a space-filling
distribution of 35 catalyst compositions. This algorithm first
constructs a space-filling mixture design of 27 compositions
with an ensured minimum separation between the two points.
Eight additional compositions are then purpose-selectedly
added after analysis of this initial distribution. Nine randomly
selected compositions from this design are replicated, either
fully resynthesized and tested or retested in a different reactor
tube. Together with 12 reference measurements (blind test,
unpromoted Pt, only promoters, bare support, Pt−Sn−K, also
replicated), this yields a total of 56 samples that are measured
in seven batches. Full details on the design construction are
given in the Supporting Information, and the full design,
replications, and references are reproduced in Tables S4−S7.
Time-Efficient Parallel Catalytic Testing. Catalyst
testing focuses on the reaction yield Ypropylene, as a combination
of the conversion of propane in a continuous flow system and
product selectivity. For the very dynamic PDH reaction, which
shows a fast (reversible) deactivation over time, we specifically
measure the transient behavior of Ypropylene over two PDH
cycles with a time on stream (TOS) of 24 h each and a 2 h
regeneration phase using a stepped oxygen supply in between.
Figure 1 shows a corresponding trace for an unpromoted Pt
reference catalyst with a 0.06 surface fraction (0.29 wt %
loading). The strong transient in particular in the initial
formation phase with high conversion at limited selectivity
challenges an efficient parallel testing. The need to resolve this
transient in short time steps clashes with the intrinsic time
demand of the gas analysis through gas chromatography (GC).
Even a highly optimized GC-measurement with a simple
separation task including C1 to C4 alkanes and olefins still
takes around ∼4 min. Simply alternating these measurements
between the eight tubes in our parallel reactor setup would
then only yield data in ∼32 min steps.
We therefore develop a time-resolved parallel testing
(TRPT) protocol fully described in the Supporting Informa-
tion, in which the testing in the different reactor tubes starts
subsequently, each with a 1 h delay. This way, the first reactor
is heated to reaction temperature and the transient behavior of
the PDH catalyst can be monitored at high resolution, while all
other reactors are still inactive. After 1 h, the second reactor is
heated and this sample is exclusively tested for 1 h. This
procedure is continued until all reactors are at reaction
temperature. From then on, all reactors are monitored with
alternating GC-measurements at a concomitant lower
resolution. However, at these later times, the deactivation,
then primarily through coking and sintering, also proceeds on
longer time scales, which do not need such a high resolution;
cf. Figure 1. The smoothness of Ypropylene at this later TOS also
mitigates the measurement break that necessarily results from
the TRPT, where the first seven reactors are not tested for a
certain period of time after they have already reached the
reaction temperature.
We can therefore robustly fit the TOS data of each cycle
with a simple sum of two exponentially decaying functions, as
motivated by the two deactivation regimes. As illustrated in
Figure 1, integrating this function over each cycle then yields
Y1and Y2as a meaningful descriptor for the integral yield in
the main product propylene. In order to also capture the fast
deactivation behavior, we additionally define a second
descriptor D1/2 =Y1/2
[17.5h;25h]/Y1/2
[10h;17.5h] as the ratio of the
integral yields of the subtraces over the time range from 17.5 to
25 h and 10 to 17.5 h; see Supporting Information for full
Figure 1. Extraction of catalytic performance descriptors from
measured reaction yield traces. The reaction yield toward propylene
is shown over two PDH cycles of 24 h each with a 2 h regeneration
period in between (to restore the diminished activity after the fast
deactivation in the initial cycle). For clarity, the propylene formation
rate rpropylene [mmol/h] (right y-axis) is normalized to reaction yield Y
in % of the amount of propane dosed per time (left y-axis). Shown are
data for an unpromoted Pt reference catalyst (XPt = 0.06), for a Pt−
Sn−K reference catalyst (XPt = 0.09, XSn = 0.07, XK= 0.52, cf.
Supporting Information for details), and for a here identified Pt−
Mn−Ca catalyst (XPt = 0.04, XMn = 0.25, XCa = 0.25); see text. The
dotted horizontal line marks the thermodynamic limit for the reaction
yield. Integration of the fitted interpolation function (blue line) over
the 24 h TOS of each cycle iyields the integral yield descriptor Yifor
propylene (in % of the total amount of propane dosed over the same
period of time). The descriptor Difor the fast deactivation over each
cycle results from the fraction of the integrated subtraces in the time
windows [17.5 h; 25 h] and [10 h; 17.5 h]. The missing data ranges
result from the parallel testing protocol described in the main text.
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9010
details of the descriptor definitions as well as of a series of
dedicated replication experiments that confirm a replication
accuracy (root-mean-square error) of ±0.5/0.7% in Y1/2 and of
±2.2/2.8% in D1/2. High Y1/2 and high D1/2 values are thus
favorable, encoding catalysts with a high product yield and
slow deactivation. We specifically employ Y2and D2of the
second cycle as the DoE response functions, as we expect the
performance over the second PDH cycle to be more
representative for the longer term catalyst function. As
apparent from Figure 1, the Pt−Sn−K reference catalyst
indeed significantly improves over the unpromoted Pt catalyst,
with respect to both descriptors. The reaction yield is much
higher and the rate of deactivation is decreased.
Global Exploration of the Design Space. We already
achieve high-quality fits of the measured response data Y2,jand
D2,jfor the j= 1, ..., 35 catalyst compositions of our space-
filling design with Scheffe-form linear models appropriate for
mixture constraints
44
= + + + +
+ +
Y X X X X X
X X
j j j j j j
j j
2, Pt Pt, Fe Fe, Co Co, Ga Ga, Mn Mn,
KK, Ca Ca,
(1)
with the fit parameters βPt/Fe/Co/Ga/Mn/K/Ca and an analogue
model for D2,j. These models can be further improved by
adding the most decisive blending terms βxyXx,jXy,jamong the
seven considered elements (Pt and six promoters). The latter
is motivated from a chemical perspective, as pairwise
interactions might occur between a few but likely not all of
the incorporated elements. Specifically, we compute the 10
most promising models for each response, when allowing up to
3 blending terms to be added, and finally obtain robust
predictions for Y2and D2by averaging over the 10 model
predictions. Full details of the models and the fitting procedure
are given in the Supporting Information. The predictive quality
of the models is reflected in a low root-mean-square error of
0.8 (2.9)% for Y2(D2).
On the basis of the established model, Figure 2a provides an
overview over the performance descriptors of all 7265 unique
catalysts in the design space. The 35 explicitly measured
compositions are highlighted and distribute over a large part of
the range of achievable descriptors. This confirms the original
space-filling design concept and suggests that the model
predictions are largely based on interpolation. Unfortunately,
the achieved global overview over the possibilities in the
considered multipromoter design space indicates a Pareto front
between reaction yield and deactivation. More active catalysts
(higher Y2) deactivate faster (lower D2). This also holds for the
unpromoted Pt reference catalyst and the empty support, the
data of which are also included in Figure 2a. The empty
support is, of course, less active but barely deactivates over the
cycle. Intriguingly, however, this product yield of the empty
support is nevertheless actually far from being zero; i.e., the
support exhibits a significant intrinsic activity. This contribu-
tion is obviously suppressed in the present limit of surface-
covering promotion, as illustrated by the reduced Y2perform-
ance descriptor of a representative promotor-only reference
catalyst with nearly full surface coverage also included in
Figure 2a.
According to the model and in the considered range of
concentrations, none of the tested promoter combinations is
thus able to break the yield-deactivation trade-off suspected
from single-promoter studies.
13,16,45,46
In order to validate this
important but slightly disappointing insight, we devise a further
batch of experiments to specifically probe the tentative Pareto
front. On the basis of the predicted performance descriptors,
eight compositions are chosen accordingly, cf. Supporting
Information, with Figure 2b displaying their location all along
the Pareto front. Figure 2b also shows their performance
descriptors that are subsequently obtained in the measure-
ments. While there are in part significant deviations between
predicted and measured performance, the overall insight into
the existence of a Pareto front is fully confirmed.
Unfortunately, in the present surface-covering promotion
limit, this Pareto front runs at quite some distance from the
established Pt−Sn−K reference catalyst, cf. Figure 2a. While
the considered promoters can drastically slow down
deactivation, with D2values equal to or substantially higher
than the one of the Pt−Sn−K catalyst, this comes at the prize
of a strongly reduced yield. Indeed, the largest fraction of all
7265 multipromoted catalysts is in fact predicted to have
substantially higher D2descriptors than the unpromoted Pt
reference catalyst. Instead, very few are predicted to exhibit
higher yields. Ascribing the latter mostly to the lacking yield
contribution from empty support sites, it is thus imperative to
Figure 2. Overview over the multipromoter catalyst design space. (a) Performance descriptors for reaction yield Y2and deactivation D2for the
7265 unique catalysts in the design space as predicted by the regression model (blue dots). The descriptors for the 35 explicitly measured
compositions are shown as black crosses. Also shown are the descriptors for the unpromoted Pt catalyst (red cross) and the Pt−Sn−K reference
catalyst (brown cross) already discussed in Figure 1, as well as for the empty support (pink cross) and a support largely covered with promoters
only (green cross, XFe/Co = 0.08, XGa/Mn/K/Ca = 0.2); see text. (b) Same as (a), but including the data for an additional batch of eight experiments to
validate the Pareto front apparent from the model prediction. Orange dots are the model predictions, orange crosses the subsequently measured
corresponding experimental values.
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9011
evaluate the action of the individual promoters and identify
those with the strongest effects per concentration. These
promoters would then be prime candidates for a low surface-
covering promoter formulation, i.e., a formulation that achieves
a slowed down deactivation similar to those in Figure 2a, while
simultaneously blocking only a minimum number of support
sites.
Analyzing Individual Promoter Effects. Figure 3 shows
the variation of the two primary descriptors Y2and D2along
Cox component effect directions
42,47
for Pt and the six
promoters. In an independent factor experiment, the main
effect of one promoter would be assessed by varying its surface
fraction, while keeping all other components unchanged.
Within a mixture constraint, components cannot be varied
independently of each other though. Cox component effect
directions are then a good approximation of an independent
one-factor variation by making appropriate changes to the
other components that lead to minimum aliasing. Full details
on the construction of the Cox directions around a reference
point Sin the center of the design space, as well as an explicit
experimental validation of the deduced effects are given in the
Supporting Information. Note in particular that here and
henceforth we exploit the availability of the eight additional
experimental measurements done to analyze the Pareto front
above and refit the DoE-based predictive model to the
accordingly enlarged experimental database of 43 composi-
tions. The predictive quality of the models is again reflected in
a low root-mean-square error of 1.1 (3.3)% for Y2(D2).
The main effect of the active metal Pt seen in Figure 3 is as
intuitively expected. More Pt leads to an increase in the
reaction yield at a concomitantly increasing deactivation.
Unfortunately, an analogue yield-deactivation trade-off is
individually observed for all investigated promoters, too. An
improvement in one performance descriptor always comes at
the cost of deterioration in the other. Intriguingly, however, the
promoters are separated into two classes. An increasing
fraction of Ga or Mn leads to an increase in reaction yield,
at a simultaneous decrease of D2. In contrast, the addition of all
other promoters (Fe, Co, Ca, and K) stabilizes the system with
increasing D2, while reducing the reaction yield.
Overall, these different actions are fully consistent with
previous qualitative assignments from single-promoter studies
in literature.
27,29,31,35,36,38
However, our present quantitative
analysis in the full multipromoter space reveals quite
pronounced differences in the magnitude of the effects. In
particular, already smallest changes of the Fe or Co fractions
have drastic effects on both Y2and D2, cf. their steep slopes
shown in Figure 3, whereas all other promoters have a much
weaker effect. Interestingly, the magnitude of the effect on one
performance descriptor can also be quite different from that on
the other one. The moderate beneficial effect of more Ga on
the reaction yield is, for example, accompanied by a dramatic
reduction of D2. In this respect, Mn is a much more promising
promoter. It achieves an even slightly higher increase of Y2as
Ga, but it worsens the deactivation behavior much less. A
similar case can be observed in the other promoter group that
slows deactivation with increasing D2at higher fractions. Here,
Ca and K lead to similar D2increases, but Ca reduces the
reaction yield much less.
Insight-Guided Extension of the Design Space. In
particular the identified opposing actions of the two groups of
promoters underscore the hitherto barely exploited potential of
multipromoter systems, where in an optimum mixture the
unavoidable deteriorating effects of promoter(s) of one group
are maximally compensated by the beneficial action of
promoter(s) of the other group. At the varying strengths of
actions apparent in Figure 3, this compensation could also be
achieved at quite differing concentrations. A small surface
fraction of a strongly acting promoter may then compensate
the effect of a larger surface fraction of a weakly acting
promoter. At the same time, similar actions of the various
promoters within one group identified above question the
minimum concentrations we initially imposed for each
promoter in the mixture constraint. While the latter allowed
to establish a reliable DoE model for a first overview and the
present analysis, precisely this analysis now suggests an
extension of the design space to where individual promoter
concentrations go down to zero. This way, a minimum
presence of promoters whose action is equally achieved by
another promoter is not unnecessarily enforced, while at the
same time, higher maximum surface fractions of other
promoters can be explored under the given mixture constraint.
Seeing the beneficial action, in particular, of Mn on the
reaction yield, one hope is that such higher maximum fractions
in a concomitantly extended design space could particularly
improve the reachable performance descriptors toward Y2that
are more competitive with the Pt−Sn−K reference catalyst.
We correspondingly set the lowest possible surface fraction
for all promoters to 0.0 and concomitantly increase the
maximum possible fractions of Ga, Mn, K, and Ca to 0.9.
Under the given mixture constraint and the same step widths
in fractions as before, this leads to an increased design space
comprising 19,836 unique catalysts. Figure 4 summarizes the
performance descriptors predicted by the regression model for
this extended space. The range of predicted performance
descriptors in Figure 4 is now significantly increased,
confirming the expectations from the effect analysis, and
extends in particular to much higher reaction yields.
A Pareto front is still visible. However, in particular toward
larger Y2, this front changes slope and less deactivation occurs
upon increased reaction yield. This thus improves upon the
originally deduced Pareto front in the smaller design space and
already reduces the performance gap to the Pt−Sn−K catalyst.
Figure 3. Individual promoter effects. Main effects of Pt and the six
promoters on (a) reaction yield descriptor Y2and (b) deactivation
descriptor D2. Within the given mixture constraints, these individual
effects are approximately obtained along Cox directions varying the
surface fraction Xiof the corresponding component around a
reference point Si(XPt = 0.04, XFe/Co = 0.06, XGa/Mn/K/Ca = 0.21) in
the middle of the design space, while making appropriate
compensating changes to the other surface fractions; see text.
Confidence intervals at the 95% level around each predicted mean
effect are provided with a lighter transparent shading.
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9012
We have to recognize though that the model predictions in this
rightmost part of the Pareto front in Figure 4 are likely less
reliable. The corresponding catalyst compositions are at the
outer boundary of the design space, where deviations from the
simple behavior assumed in our DoE model are to be expected.
To assess the corresponding uncertainty in the predicted
position of the Pareto front, we subjected a series of promising
candidate compositions with highest predicted Y2to
experimental testing. Specifically and using a space-filling
algorithm analogous to the one employed for the initial
exploration of the design space, cf. Supporting Information for
details, we choose a batch of eight candidates from the pool of
100 such highest predicted-yield compositions. This pool, the
selected batch, and the outcome of the measurements are also
displayed in Figure 4. As expected, the actually measured
performance descriptors do vary quite a bit from the
predictions. Notwithstanding, half of the candidates from the
batch indeed exhibit reaction yields that are significantly larger
than the one of the Pt reference catalyst. This confirms the
understanding that higher Y2values are possible at the
boundaries of the now enlarged design space. In fact, two of
the measured candidates even show yields that are comparable
to the one of the Pt−Sn−K reference catalyst (Y2s of 19.6 and
21.0 vs 19.9%, respectively). Unsurprising from the perspective
of the effect analysis, they contain essentially either only Mn or
Mn and Ca.
Support Site Effects in the (Pt, Mn, and Ca) Space.
The understanding gained so far separates the six promoters
into two groups of similar actions, underscores the necessity to
go up to higher surface concentrations in individual promoters,
and motivates lower surface-covering formulations that then
allow for the additional yield contribution from empty support
sites. In a final step, we implement this understanding by
switching to the low-dimensional (Pt, Mn, and Ca) design
space, in which we give up the mixture constraint and consider
any total concentrations up to full surface coverage. With Mn
and Ca, this retains the most promising promoter from each
group, according to the effect analysis. Relying on the
previously introduced space-filling algorithm, we design two
batches of measurements to determine a polynomial model
with full details provided in the Supporting Information.Figure
5shows the performance descriptors of these measured
catalysts together with the predictions of the determined
model for the entire design space. The models show a low
root-mean-square error of 1.9 (1.9)% for Y2(D2). As before, a
Pareto front between the reaction yield and deactivation can
clearly be discerned. However, confirming the expected
support site contribution, this front is now shifted to higher
Y2. As a result, there are multiple promoter formulations with a
predicted performance competitive to that of the equally not
fully surface-covering Pt−Sn−K reference. In fact, when
comparing the explicit reaction yield traces of the first and
second PDH cycle shown in Figure 1 of one of such top-
performing Pt−Mn−Ca catalysts that was part of the
measurement batch, we even observe smaller conditioning
effects. The Y1and D1of this catalyst are already almost the
same as its Y2and D2, whereas quite some reduction in both
performance descriptors is observed between the first and
second cycle for the Pt−Sn−K reference.
■CONCLUSIONS
With less than a hundred experiments performed within a
matter of weeks, we have identified a promising biopromoted
Pt catalyst for the PDH that shows a performance in terms of
reaction yield and deactivation resistance that is competitive to
a Pt−Sn−K formulation as employed in the industrial Oleflex
process. With a large-scale application in mind, this catalyst
rests on a commercially available alumina support, an
industrially viable Pt loading, and a simple synthesis protocol
employing incipient wetness impregnation. The discovered
Mn−Ca promotion demonstrates that there is nothing unique
about the hitherto established Sn−K chemistry. In fact, while
still showing a slightly inferior absolute performance, the new
Pt−Mn−Ca system actually already exhibits a recovery after
Figure 4. Identification of higher yield catalysts in an extended design
space. Performance descriptors for reaction yield Y2and deactivation
D2as predicted by the regression model for 19 836 unique catalysts in
an extended design space that allows individual promoter concen-
trations to go to zero (olive dots), while maintaining the surface-
covering promotion limit; see text. Additionally shown are the 7265
unique catalysts in the original design space summarized already in
Figure 2 (blue dots), the Pareto front deduced from these original
data, as well as the descriptors for the unpromoted Pt catalyst (red
cross) and the Pt−Sn−K reference catalyst (brown cross). Eight
promising compositions are suitably chosen (black dots) from the top
100 candidates (gray dots) with highest Y2in this extended design
space. Their subsequently measured corresponding experimental
values are shown as black crosses.
Figure 5. Overview over the rationally motivated (Pt, Mn, and Ca)-
design space without surface-covering constraint. Performance
descriptors for reaction yield Y2and deactivation D2for 607 unique
catalysts in the design space as predicted by the regression model
(blue dots). The descriptors for the explicitly measured compositions
are shown as black crosses. Also shown are the descriptors for the
unpromoted Pt catalyst (red cross), the Pt−Sn−K reference catalyst
(brown cross), and the tentative Pareto front deduced before for
surface-covering promotion, cf. Figure 2.
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9013
the first regeneration that is clearly better than that of the Pt−
Sn−K reference. These findings constitute breakthroughs at
the discovery stage. As is established practice throughout, this
will now be followed by an optimization stage, where the
compositions are fine-tuned, different synthesis routes leading
to the same composition are explored (e.g., order of
impregnation) and where it then makes sense to assess the
longer-term stability over three and more cycles. There is little
doubt that further performance gains will be achieved in this
optimization stage, but such results might then no longer be
published in the open literature.
However, even more remarkable than this discovery itself is
in our view the way it was achieved. Initially, promoter action
on catalyst performance was globally evaluated in a defined
high-dimensional multipromoter space spanning six mindfully
chosen promoters and suppressing additional support site
effects through surface-covering promotion. This overview
could be reached in a data- and time-efficient manner through
careful experiment planning based on an iterative DoE
approach utilizing space-filling algorithms and the develop-
ment of a testing protocol for a multireactor setup that
maximized the experimental throughput. Catalyst performance
could quantitatively be compared through the definition of
suitable yield and deactivation descriptors as well as stringently
maintained reaction conditions. Following an insight-based
iterative strategy, the obtained results motivated further
analyses, which were supported by additional custom DoE
designs and concomitant measurement batches. The observa-
tion of a Pareto front between yield and deactivation led to an
effect analysis that traced the trade-off down to the individual
promoter level, clearly identifying both positive and negative
actions of the individual species. The resulting classification of
the promoters into two groups with opposing actions on yield
and deactivation led to an extension of the initial design space
that revealed the most promising performance when reducing
the total promoter number to ultimately one representative of
each group. Abandoning the surface-covering promotion to
also benefit from the intrinsic support site activity finally led to
the global exploration of the low-dimensional (Pt, Mn, and
Ca) design space that yielded the competitive formulation.
Most importantly, the identification of this binary promoter
search space was rationally achieved and is the result of the
previous designs unraveling the synergistic action of this hardly
reducible and one moderating element as well as the only
insignificant interactions with the other promoters. Directly
restricting the exploration to only binary promoter spaces
would have implied assuming these insights and would hardly
have reduced the experimental burden (at the present
throughput capacity essentially by about 3 weeks).
This rational and systematic discovery approach stands in
strong contrast to the scattered empirical catalysis literature,
focusing on positive effects of individual promoters. The
provided global understanding of trends, possibilities, and
limitations in the multipromoter space is also much more than
a mere global optimization as often strived for in closed-loop
automated lab approaches.
48−54
Apart from enabling the
efficient identification of promising multipromoter formula-
tions, this understanding provides guidance on which systems
would particularly merit deeper subsequent mechanistic
studies and creates leads for insight-based modifications of
the design space.
In contrast to widely known classic DoE, iterative DoE as
pursued here should methodologically rather be seen as a
highly data-efficient active-learning technique such as Bayesian
optimization (BO). Like BO, it can query new experiments to
refine the response model based on various optimality criteria.
Able to robustly handle noise in the data and able to work with
larger data batches, iterative DoE is in fact ideally suited for
accelerated discovery within automated catalysis laboratories at
present throughput capacities. Most importantly, and as
demonstrated here, it can naturally deal with hypothesis- or
insight-based adaptions of the design space between iterations.
As such, it is in our view a far better suited approach to realize
the human-in-the-loop paradigm of process automation. Based
on the hitherto acquired data and knowledge, the involved
human experts can make key decisions to suitably adapt the
(otherwise automated) workflow for the next batch of
experiments, rather than stoically executing an originally
defined optimization in a fixed design space in a closed-loop
setup.
■METHODS
Synthesis. The catalysts were prepared in a home-built half
automated setup by incipient wetness impregnation. Commer-
cially available γ−alumina (BASF SE) with a particle size
fraction of 300−500 μm was used as a support. Before
impregnation, the support was thermally treated at 1000 °C for
12 h employing a heating rate of 5 °C min−1and effectively
achieving a phase transition to θ−Al2O3. In a first step, 2 mL
of different combinations of 2 M stock solutions of the metal
precursors was then impregnated on 4 g of support using a
dosing rate of 0.2 mL min−1. The support was contained in a
glass vessel, placed on a multisampling orbital shaker
(Heidolph Instruments GmbH). The impregnated support
was homogenized for 30 min at 1600 rpm. Using a multistage
syringe pump (model Fusion 200-X, Chemyx Inc.), eight
impregnations were performed in parallel. After that, the
supported metal promoters were dried (2 h, 70 °C) and
calcined (4 h, 500 °C, 10 °C min−1) in a horizontal tubular
furnace using 1 L min−1of a mixture of 20% O2in He. In a
second step, the catalysts were impregnated with an aqueous
85 mM solution of
Pt(NH ) (NO )
3 4 3 2
following the same
procedure as described above. Finally, the catalysts were again
dried (2 h, 70 °C) and calcined (2 h, 300 °C, 5 °C min−1).
After the thermal treatment, the furnace cooled down
naturally. All reagents (
·Fe(NO ) 9H O
3 3 2
99.9%,
·Co(NO ) 6H O
3 2 2
99.9%,
·Ca(NO ) 4H O
3 2 2
99.0%,
·Ga(NO ) 8H O
3 3 2
98.0%, K(NO3) 99.0%,
·Mn(NO ) 4H O
3 2 2
97.0%, and
Pt(NH ) (NO ) 50.0
3 4 3 2
% Pt) were obtained from
Sigma-Aldrich and used as received without further purifica-
tion. H2O HPLC grade (Honeywell Riedel-de-Haen, Ger-
many) was employed for the preparation of the metallic
solutions.
55
As described in the main text, all catalyst
preparations aim to achieve full total surface coverage. Based
on results for Al
56
and extrapolating to the larger elements
employed here, we estimate a surface concentration of 6
atoms/nm2to build a closed monolayer, cf. Supporting
Information.
Catalytic Testing. The generation of catalytic performance
data was conducted using an eight-fold parallel test setup
equipped with gas dosing, reactor heating, and gas analysis
units. Each of the prepared samples (500 mg, with the exact
weight documented) was filled in a quartz tube reactor (dinner =
3.8 mm) with a fixed bed embedded into a filling of steatite as
the inert material. Above and below the catalytic bed, quartz
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9014
wool was used to fix the bed heights within the reactor and
ensure an isothermic temperature. The reactor temperatures
were individually controlled and continuously monitored. The
PDH reaction was run at 600 °C with a reactant gas flow
comprising 10 mL min−1propane, 5 mL min−1hydrogen, and
2 mL min−1nitrogen per reactor (WHSV = 2.4 kgPropane/h/
kgCat, GHSV = 720 LPropane/h/LCat or GHSV = 1224 Lfeed/h/
LCat).
Two PDH cycles were studied with a TOS of 24 h each.
Prior to both reaction cycles, a reduction step was used as
pretreatment of the catalyst at 600 °C for 1 h using 5 mL
min−1of hydrogen in 50 mL min−1. Afterward, the reactors
were cooled to 225 °C, which was determined as temperature
with no observable catalytic activity. In between the two PDH
cycles, a regeneration step was conducted using a stepped
oxygen supply at a fixed temperature of 500 °C (5% O2and
20% O2using synthetic air diluted with nitrogen for 1 h each,
heat-up with ΔT≈17 K min−1in regeneration stream). Before
and after the catalytic test, as well as in between changing feed
conditions, the samples were flushed with nitrogen. The
analysis of the outlet gas stream was conducted using an online
GC (GC 7890 A, Agilent) equipped with a flame ionization
and a thermal conductivity detector for the identification and
quantification of the reactants and reaction products like n-
butane, 1-butene, propylene, ethane, ethylene, and methane.
Besides, nitrogen was used as internal standard in the
calculation of the gas compositions, in order to consider the
gas expansion. In addition, the reactant gas composition was
monitored continuously throughout the DoE using a bypass
reactor. The test procedure was fully automated using the
commercial software hte control; raw data evaluation was
performed using the software tool myhte. The reproducibility
of the catalytic test including the proof of independence of the
sample placement throughout the eight reactors was given by
performing reproduction experiments; see Figure S2 in the
Supporting Information.
From the outlet concentrations (ci), the propane conversion
Xpropane and the selectivity Sitoward the quantified reaction
products (i= propylene, ethane, ethylene, methane, butane,
and butene) were calculated using a product based approach
following eqs 2 and 3, with the number of carbon atoms
represented by NC.
[ ] =
+ ·
·
=
i
k
j
j
j
j
j
j
j
j
j
j
y
{
z
z
z
z
z
z
z
z
z
z
( )
X
c
c c
% 1 100
ii
N
propane
propane
propane 1
6
3
iC,
(2)
[ ] = ·
· + ·
·
=
S
c N
c c N
%3 ( ) 100
i
i i
ii i
C,
propane 1
6
C,
(3)
From this, the yield Ypropylene toward the desired product
propylene is derived, combining the relative conversion of the
hydrocarbon reactant and the selectivity to the olefin product
[ ] = ·
+ ·
=
( )
Y X
c
c c
%
ii
N
propylene propane
propylene
propane 1
6
3
iC,
(4)
The strongly time-dependent Ypropylene (tTOS) are naturally
given as a percentage of the constant amount of propane dosed
per time. This molar flow rate of propane rpropane,in into the
reactor can be calculated from the employed volumetric flow
rate
V
propane,in
and the molar volume at standard conditions Vm,
cf. Supporting Information
=
=
·
·
=
Ä
Ç
Å
Å
Å
Å
Å
Å
Å
Å
É
Ö
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ä
Ç
Å
Å
Å
Å
Å
Å
É
Ö
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ä
Ç
Å
Å
Å
Å
Å
Å
É
Ö
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ä
Ç
Å
Å
Å
Å
Å
Å
É
Ö
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ä
Ç
Å
Å
Å
Å
Å
Å
É
Ö
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
r
V
V
mmol
h
10 60
1000 0.02241
26.7 mmol
h
propane,in
propane,in
m
mL
min
min
h
mL
L
L
mmol
(5)
The time-dependent molar flow rate of propylene then
follows simply as
= [ ]·
Ä
Ç
Å
Å
Å
Å
Å
Å
Å
Å
É
Ö
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
Ñ
r t Y t r( ) mmol
h( ) %
propylene TOS propylene TOS propane,in
(6)
■ASSOCIATED CONTENT
*
sı Supporting Information
The Supporting Information is available free of charge at
https://pubs.acs.org/doi/10.1021/acscatal.4c01740.
Experimental details on synthesis and catalytic testing of
the multipromoter and reference catalysts, experimental
designs of the initial exploration and insight-guided
extended design space, details on replication and
reference catalysts, TRPT protocol, details of perform-
ance descriptors, regression models and their perform-
ance for the exploration of the multipromoter mixture
design space and the Pt−Mn−Ca design space,
exploration of the Pareto front, catalyst component
effect along the Cox direction, and study workflow
(PDF)
■AUTHOR INFORMATION
Corresponding Authors
Frank Rosowski −BasCat�UniCat BASF JointLab,
Technische Universität Berlin, D-10623 Berlin, Germany;
BASF SE, Catalysis Research, D-67065 Ludwigshafen,
Germany; Email: [email protected]
Karsten Reuter −Fritz-Haber-Institut der Max-Planck-
Gesellschaft, D-14195 Berlin, Germany; orcid.org/0000-
0001-8473-8659; Email: [email protected]
Authors
Christian Kunkel −Fritz-Haber-Institut der Max-Planck-
Gesellschaft, D-14195 Berlin, Germany
Frederik Ruther −BasCat�UniCat BASF JointLab,
Technische Universität Berlin, D-10623 Berlin, Germany
Frederic Felsen −Fritz-Haber-Institut der Max-Planck-
Gesellschaft, D-14195 Berlin, Germany
Charles W. P. Pare −Fritz-Haber-Institut der Max-Planck-
Gesellschaft, D-14195 Berlin, Germany
Aybike Terzi −BasCat�UniCat BASF JointLab, Technische
Universität Berlin, D-10623 Berlin, Germany
Robert Baumgarten −BasCat�UniCat BASF JointLab,
Technische Universität Berlin, D-10623 Berlin, Germany
Esteban Gioria −BasCat�UniCat BASF JointLab,
Technische Universität Berlin, D-10623 Berlin, Germany
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9015
Raoul Naumann d’Alnoncourt −BasCat�UniCat BASF
JointLab, Technische Universität Berlin, D-10623 Berlin,
Germany; orcid.org/0000-0002-9946-4619
Christoph Scheurer −Fritz-Haber-Institut der Max-Planck-
Gesellschaft, D-14195 Berlin, Germany
Complete contact information is available at:
https://pubs.acs.org/10.1021/acscatal.4c01740
Author Contributions
C.K. and F.Rue. contributed equally. C.K., F.F., and C.S.
designed the experiments. C.K. and C.P. curated data,
analyzed, and conceptualized the results (if not otherwise
stated). R.N.-d’A. and F.Rue. conceptualized the experimental
TRPT strategy. F.Rue. and A.T. performed the catalytic
measurements. F.Rue. conducted the literature search. R.B.,
E.G., and R.N.-d’A. conceptualized and carried out the
synthesis. F.R., R.N.-d’A., C.S., and K.R. initially conceived
the study. All authors contributed in conceptualizing the study.
C.K., F.Rue., and K.R. wrote the manuscript.
Funding
Open access funded by Max Planck Society.
Notes
The authors declare no competing financial interest.
■ACKNOWLEDGMENTS
This work was conducted in the framework of the BasCat�
UniCat BASF JointLab between BASF SE, Technische
Universität Berlin (TU Berlin), and Fritz-Haber-Institut
(FHI) der Max-Planck-Gesellschaft. Funding by the Deutsche
Forschungsgemeinschaft (DFG, German Research Founda-
tion) under Germany’s Excellence Strategy�EXC 2008−
390540038�UniSysCat is acknowledged. We thank Abbas El-
Jamal (TU Berlin) for diligent efforts and commitment in
operating the synthesis setup and Jan Meißner (TU Berlin) for
the technical support running the catalytic tests.
■REFERENCES
(1) Pease, R. N.; Taylor, H. S. Promotor Action in Catalysis. J. Phys.
Chem. 1920,24, 241−265.
(2) Ertl, G. Elementary Steps in Heterogeneous Catalysis. Angew.
Chem., Int. Ed. 1990,29, 1219−1227.
(3) Hutchings, G. J. Promotion in Heterogeneous Catalysis: A Topic
Requiring a New Approach? Catal. Lett. 2001,75, 1−12.
(4) Schlögl, R. Heterogeneous Catalysis. Angew. Chem., Int. Ed.
2015,54, 3465−3520.
(5) Han, Z.; Li, S.; Jiang, F.; Wang, T.; Ma, X.; Gong, J. Propane
Dehydrogenation Over Pt−Cu Bimetallic Catalysts: The Nature of
Coke Deposition and the Role of Copper. Nanoscale 2014,6, 10000−
10008.
(6) Wu, Z.; Bukowski, B. C.; Li, Z.; Milligan, C.; Zhou, L.; Ma, T.;
Wu, Y.; Ren, Y.; Ribeiro, F. H.; Delgass, W. N.; Greeley, J.; Zhang, G.;
Miller, J. T. Changes in Catalytic and Adsorptive Properties of 2 nm
Pt3Mn Nanoparticles by Subsurface Atoms. J. Am. Chem. Soc. 2018,
140, 14870−14877.
(7) Wegener, E. C.; Bukowski, B. C.; Yang, D.; Wu, Z.; Kropf, A. J.;
Delgass, W. N.; Greeley, J.; Zhang, G.; Miller, J. T. Intermetallic
Compounds as an Alternative to Single-atom Alloy Catalysts:
Geometric and Electronic Structures from Advanced X-ray Spectros-
copies and Computational Studies. ChemCatChem 2020,12, 1325−
1333.
(8) Cybulskis, V. J.; Bukowski, B. C.; Tseng, H.-T.; Gallagher, J. R.;
Wu, Z.; Wegener, E.; Kropf, A. J.; Ravel, B.; Ribeiro, F. H.; Greeley, J.;
Miller, J. T. Zinc Promotion of Platinum for Catalytic Light Alkane
Dehydrogenation: Insights into Geometric and Electronic Effects.
ACS Catal. 2017,7, 4173−4181.
(9) Chen, S.; Pei, C.; Sun, G.; Zhao, Z.-J.; Gong, J. Nanostructured
Catalysts toward Efficient Propane Dehydrogenation. Acc. Mater. Res.
2020,1, 30−40.
(10) Zhang, Y.; Zhou, Y.; Shi, J.; Zhou, S.; Sheng, X.; Zhang, Z.;
Xiang, S. Comparative Study of Bimetallic Pt-Sn Catalysts Supported
on Different Supports for Propane Dehydrogenation. J. Mol. Catal. A:
Chem. 2014,381, 138−147.
(11) Carter, J. H.; Bere, T.; Pitchers, J. R.; Hewes, D. G.;
Vandegehuchte, B. D.; Kiely, C. J.; Taylor, S. H.; Hutchings, G. J.
Direct and Oxidative Dehydrogenation of Propane: From Catalyst
Design to Industrial Application. Green Chem. 2021,23, 9747−9799.
(12) Chen, S.; Chang, X.; Sun, G.; Zhang, T.; Xu, Y.; Wang, Y.; Pei,
C.; Gong, J. Propane Dehydrogenation: Catalyst Development, New
Chemistry, and Emerging Technologies. Chem. Soc. Rev. 2021,50,
3315−3354.
(13) Sattler, J. J. H. B.; Ruiz-Martinez, J.; Santillan-Jimenez, E.;
Weckhuysen, B. M. Catalytic Dehydrogenation of Light Alkanes on
Metals and Metal Oxides. Chem. Rev. 2014,114, 10613−10653.
(14) Araujo-Lopez, E.; Vandegehuchte, B. D.; Curulla-Ferré, D.;
Sharapa, D. I.; Studt, F. Trends in the Activation of Light Alkanes on
Transition-Metal Surfaces. J. Phys. Chem. C 2020,124, 27503−27510.
(15) Werghi, B.; Wu, L.; Ebrahim, A. M.; Chi, M.; Ni, H.; Cargnello,
M.; Bare, S. R. Selective Catalytic Behavior Induced by Crystal-Phase
Transformation in Well-Defined Bimetallic Pt-Sn Nanocrystals. Small
2023,19, 2207956.
(16) Shan, Y.; Hu, H.; Fan, X.; Zhao, Z. Recent Progress in Catalytic
Dehydrogenation of Propane Over Pt-Based Catalysts. Phys. Chem.
Chem. Phys. 2023,25, 18609−18622.
(17) Burch, R. Platinum-Tin Reforming Catalysts: I. The Oxidation
State of Tin and the Interaction Between Platinum and Tin. J. Catal.
1981,71, 348−359.
(18) Barias, O.; Holmen, A.; Blekkan, E.. In Catalyst Deactivation
1994; Delmon, B., Froment, G., Eds.; Studies in Surface Science and
Catalysis; Elsevier, 1994; Vol. 88; pp 519−524.
(19) Barias, O. A.; Holmen, A.; Blekkan, E. A. Propane
Dehydrogenation Over Supported Platinum Catalysts: Effect of Tin
as a Promoter. Catal. Today 1995,24, 361−364.
(20) Vladimir, H. Conversion of Hydrocarbons with Platinum
Composite Catalyst. U.S. Patent 2,602,772 A, 1952.
(21) Zhang, Y.; Yao, W.; Fang, H.; Hu, A.; Huang, Z. Catalytic
Alkane Dehydrogenations. Sci. Bull. 2015,60, 1316−1331.
(22) Bhasin, M.; McCain, J.; Vora, B.; Imai, T.; Pujadó, P.
Dehydrogenation and Oxydehydrogenation of Paraffins to Olefins.
Appl. Catal. 2001,221, 397−419.
(23) Nawaz, Z. Light Alkane Dehydrogenation to Light Olefin
Technologies: A Comprehensive Review. Rev. Chem. Eng. 2015,31,
413.
(24) Zhou, N.; Liu, W.; Jan, F.; Han, Z.; Li, B. Efficient Screening of
Metal Promoters of Pt Catalysts for C−H Bond Activation in Propane
Dehydrogenation from a Combined First-Principles Calculations and
Machine-Learning Study. ACS Omega 2023,8, 23982−23990.
(25) Ingale, P.; Knemeyer, K.; Preikschas, P.; Ye, M.; Geske, M.;
Naumann d’Alnoncourt, R.; Thomas, A.; Rosowski, F. Design of PtZn
Nanoalloy Catalysts for Propane Dehydrogenation Through Interface
Tailoring via Atomic Layer Deposition. Catal. Sci. Technol. 2021,11,
484−493.
(26) Docherty, S. R.; Rochlitz, L.; Payard, P.-A.; Copéret, C.
Heterogeneous Alkane Dehydrogenation Catalysts Investigated via a
Surface Organometallic Chemistry Approach. Chem. Soc. Rev. 2021,
50, 5806−5822.
(27) Van Assche, A.; Especel, C.; Le Valant, A.; Epron, F. Effect of
Potassium and Platinum Contents on Catalytic Performance of Pt/
Al2O3Monometallic Catalysts for Propane Dehydrogenation. Mol.
Catal. 2022,517, 112059.
(28) Lee, M.-H.; Nagaraja, B. M.; Natarajan, P.; Truong, N. T.; Lee,
K. Y.; Yoon, S.; Jung, K.-D. Effect of Potassium Addition on Bimetallic
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9016
PtSn/θ-Al2O3Catalyst for Dehydrogenation of Propane to Propylene.
Res. Chem. Intermed. 2016,42, 123−140.
(29) Gao, X.-Q.; Yao, Z.-H.; Li, W.-C.; Deng, G.-M.; He, L.; Si, R.;
Wang, J.-G.; Lu, A.-H. Calcium-Modified PtSn/Al2O3Catalyst for
Propane Dehydrogenation with High Activity and Stability.
ChemCatChem 2023,15, No. e202201691.
(30) Rimaz, S.; Chen, L.; Monzón, A.; Kawi, S.; Borgna, A.
Enhanced Selectivity and Stability of Pt-Ge/Al2O3Catalysts by Ca
Promotion in Propane Dehydrogenation. J. Chem. Eng. 2021,405,
126656.
(31) Cai, W.; Mu, R.; Zha, S.; Sun, G.; Chen, S.; Zhao, Z.-J.; Li, H.;
Tian, H.; Tang, Y.; Tao, F. F.; Zeng, L.; Gong, J. Subsurface Catalysis-
Mediated Selectivity of Dehydrogenation Reaction. Sci. Adv. 2018,4,
No. eaar5418.
(32) Liu, D.; Hu, H.; Yang, Y.; Cui, J.; Fan, X.; Zhao, Z.; Kong, L.;
Xiao, X.; Xie, Z. Restructuring Effects of Pt and Fe in Pt/Fe-DMSN
Catalysts and their Enhancement of Propane Dehydrogenation. Catal.
Today 2022,402, 161−171.
(33) Kwak, J. H.; Hu, J.; Mei, D.; Yi, C.-W.; Kim, D. H.; Peden, C.
H. F.; Allard, L. F.; Szanyi, J. Coordinatively Unsaturated Al3+ Centers
as Binding Sites for Active Catalyst Phases of Platinum on γ-Al2O3.
Science 2009,325, 1670−1673.
(34) Shi, L.; Deng, G.-M.; Li, W.-C.; Miao, S.; Wang, Q.-N.; Zhang,
W.-P.; Lu, A.-H. Al2O3Nanosheets Rich in Pentacoordinate Al3+ Ions
Stabilize Pt-Sn Clusters for Propane Dehydrogenation. Angew. Chem.,
Int. Ed. 2015,54, 13994−13998.
(35) Cesar, L. G.; Yang, C.; Lu, Z.; Ren, Y.; Zhang, G.; Miller, J. T.
Identification of a Pt3Co Surface Intermetallic Alloy in Pt−Co
Propane Dehydrogenation Catalysts. ACS Catal. 2019,9, 5231−5244.
(36) Rochlitz, L.; Pessemesse, Q.; Fischer, J. W. A.; Klose, D.; Clark,
A. H.; Plodinec, M.; Jeschke, G.; Payard, P.-A.; Copéret, C. A Robust
and Efficient Propane Dehydrogenation Catalyst from Unexpectedly
Segregated Pt2Mn Nanoparticles. J. Am. Chem. Soc. 2022,144,
13384−13393.
(37) Fan, X.; Liu, D.; Sun, X.; Yu, X.; Li, D.; Yang, Y.; Liu, H.; Diao,
J.; Xie, Z.; Kong, L.; Xiao, X.; Zhao, Z. Mn-Doping Induced Changes
in Pt Dispersion and PtxMnyAlloying Extent on Pt/Mn-DMSN
Catalyst with Enhanced Propane Dehydrogenation Stability. J. Catal.
2020,389, 450−460.
(38) Redekop, E. A.; Galvita, V. V.; Poelman, H.; Bliznuk, V.;
Detavernier, C.; Marin, G. B. Delivering a Modifying Element to
Metal Nanoparticles via Support: Pt−Ga Alloying during the
Reduction of Pt/Mg(Al,Ga)OxCatalysts and its Effects on Propane
Dehydrogenation. ACS Catal. 2014,4, 1812−1824.
(39) Sattler, J. J. H. B.; Gonzalez-Jimenez, I. D.; Luo, L.; Stears, B.
A.; Malek, A.; Barton, D. G.; Kilos, B. A.; Kaminsky, M. P.;
Verhoeven, T. W. G. M.; Koers, E. J.; Baldus, M.; Weckhuysen, B. M.
Platinum-Promoted Ga/Al2O3as Highly Active, Selective, and Stable
Catalyst for the Dehydrogenation of Propane. Angew. Chem., Int. Ed.
2014,53, 9251−9256.
(40) Searles, K.; Chan, K. W.; Mendes Burak, J. A.; Zemlyanov, D.;
Safonova, O.; Copéret, C. Highly Productive Propane Dehydrogen-
ation Catalyst Using Silica-Supported Ga−Pt Nanoparticles Gen-
erated from Single-Sites. J. Am. Chem. Soc. 2018,140, 11674−11679.
(41) Siddiqi, G.; Sun, P.; Galvita, V.; Bell, A. T. Catalyst
Performance of Novel Pt/Mg(Ga)(Al)O Catalysts for Alkane
Dehydrogenation. J. Catal. 2010,274, 200−206.
(42) Cornell, J.. In Experiments with Mixtures: Designs, Models, and
the Analysis of Mixture Data;Wiley Series in Probability and Statistics;
Wiley, 2002.
(43) Smith, W. Experimental Design for Formulation; ASA-SIAM
Series on Statistics and Applied Probability; Society for Industrial and
Applied Mathematics, 2005.
(44) Scheffé, H. Experiments with Mixtures. J. R. Stat. Soc., B: Stat.
Methodol. 1958,20, 344−360.
(45) Li, C.; Wang, G. Dehydrogenation of Light Alkanes to Mono-
Olefins. Chem. Soc. Rev. 2021,50, 4359−4381.
(46) Monai, M.; Gambino, M.; Wannakao, S.; Weckhuysen, B. M.
Propane to Olefins Tandem Catalysis: A Selective Route Towards
Light Olefins Production. Chem. Soc. Rev. 2021,50, 11503−11529.
(47) Cox, D. R. A Note on Polynomial Response Functions for
Mixtures. Biometrika 1971,58, 155−159.
(48) Maier, W. F.; Stöwe, K.; Sieg, S. Combinatorial and High-
Throughput Materials Science. Angew. Chem., Int. Ed. 2007,46,
6016−6067.
(49) Potyrailo, R.; Rajan, K.; Stoewe, K.; Takeuchi, I.; Chisholm, B.;
Lam, H. Combinatorial and High-Throughput Screening of Materials
Libraries: Review of State of the Art. ACS Comb. Sci. 2011,13, 579−
633.
(50) Toyao, T.; Maeno, Z.; Takakusagi, S.; Kamachi, T.; Takigawa,
I.; Shimizu, K.-i. Machine Learning for Catalysis Informatics: Recent
Applications and Prospects. ACS Catal. 2020,10, 2260−2297.
(51) Peng, J.; Schwalbe-Koda, D.; Akkiraju, K.; Xie, T.; Giordano,
L.; Yu, Y.; Eom, C. J.; Lunger, J. R.; Zheng, D. J.; Rao, R. R.; Muy, S.;
Grossman, J. C.; Reuter, K.; Gómez-Bombarelli, R.; Shao-Horn, Y.
Human- and Machine-Centred Designs of Molecules and Materials
for Sustainability and Decarbonization. Nat. Rev. Mater. 2022,7,
991−1009.
(52) Margraf, J. T.; Jung, H.; Scheurer, C.; Reuter, K. Exploring
Catalytic Reaction Networks with Machine Learning. Nat. Catal.
2023,6, 112−121.
(53) Taniike, T.; Fujiwara, A.; Nakanowatari, S.; García-Escobar, F.;
Takahashi, K. Automatic Feature Engineering for Catalyst Design
Using Small Data without Prior Knowledge of Target Catalysis.
Commun. Chem. 2024,7, 11.
(54) Ramirez, A.; Lam, E.; Gutierrez, D. P.; Hou, Y.; Tribukait, H.;
Roch, L. M.; Copéret, C.; Laveille, P. Accelerated Exploration of
Heterogeneous CO2Hydrogenation Catalysts by Bayesian-Optimized
High-Throughput and Automated Experimentation. Chem Catal.
2024,4, 100888.
(55) Barias, O. A.; Holmen, A.; Blekkan, E. A. Propane
Dehydrogenation over Supported Pt and Pt−Sn Catalysts: Catalyst
Preparation, Characterization, and Activity Measurements. J. Catal.
1996,158, 1−12.
(56) Strempel, V. E.; Knemeyer, K.; Naumann d’Alnoncourt, R.;
Driess, M.; Rosowski, F. Investigating the Trimethylaluminium/Water
ALD Process on Mesoporous Silica by In Situ Gravimetric
Monitoring. Nanomaterials 2018,8, 365.
ACS Catalysis pubs.acs.org/acscatalysis Research Article
https://doi.org/10.1021/acscatal.4c01740
ACS Catal. 2024, 14, 9008−9017
9017
