1
Dion Vlachos
Department of Chemical Engineering and Center for Catalytic Science and Technology
University of DelawareNewark, DE 19716
www.che.udel.edu/vlachos, [email protected]
Multiscale Modeling and its Application to Catalyst
Design and Portable Power Generation
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
2
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
Down-scaling for future energy needs
• Distributed energy– On-board H2 production– Electric reliability– Local solutions, e.g., farms
based on biomass
• Portable energy (electronics)
Smart CarCourtesy: Ballard Power Systems
3
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
Large scale H2 production is industrially mature• Steady state operation• Reforming:
endothermic, heat transfer controlled
- Fixed bed catalytic reactors with Ni catalyst for syngas
ReformingWGS, PurificationFuel
+ Air
Hydrocarbon
Steam
Pure H2
• Large scale flames supply the heat- Half of NG is burned to CO2 and
H2O• Complex downstream processing WGS
and PROX or membrane separation/PSA• Slow (τ~1s); bulky SR: CH4 + H2O = CO + 3H2 + 206 kJ/mol
WGS: CO + H2O = CO2 + H2 - 41 kJ/mol
4
Steam reforming is a bulky process
Sehested, Cat. Today 111 (2006) 103
First example of on-board reforming
• GM unveiled the world's first gasoline fuel processor for fuel cell propulsion at the annual automotive management conference in Traverse City, Mich.
• The Gen III processor, packaged in a Chevrolet S-10 pickup, reforms 'clean' gasoline onboard, extracting a stream of hydrogen to send to the fuel cell stack.
5
Microsystems for transportation and portable applications
• Advantages– Process intensification
• High heat and mass transfer coefficients
• Multifunctionality
– Compactness– Inherently safe
• Scale out is feasible for portable (small scale) devices
10 0
10 1
10 2
10 3
10 4
10 -2 10 -1 10 0 10 1
Reactor diameter [mm]
.01 atm
1 atm
u=1 m/s
Mas
s tra
nsfe
r coe
ffici
ent,
k
Reactor diameter [mm]
101
102
103
104
10010-1 10110-2
1 atm
0.01 atm
v = 1 m/s
100
Smart CarCourtesy: Ballard Power Systems
Microscales impose challenges and offer opportunities1
Laminar flows - mixing?Small systems - enough catalyst?Reactors shake - no moveable partsNeed small pressure dropsTransient operation very common
Fast startup and shutdown require active catalyst and fast heat transferCatalyst deactivation can become a major issue
Fast chemistryDifferent systems’ engineering2
?Monolithic type reactors
Dynamics
New chemistry and catalysts
1 Norton et al., "Downsizing chemical processes for portable hydrogen production", in Microreactor Techn. Process Intensification, ACS Symp. Series 914, 179 (2005)
2 Mitsos et al., IECR 43, 74 (2004)
Flowsheets/Optimization
6
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
The multiscale simulation paradigm: A bottom-up ladder
CFDMesoscopic Theory/CGMC
KMC, DSMCLB, DPD, BD
MD, extended trajectory calc.
10-2
LengthScale (m)
Time Scale (s)
Quantum Mechanics
TST
Lab Scale
10-10
10-9
10-7
10-6
10-12 10-9 10-3 103
Downscaling or Top-down info traffic
Upscaling or Bottom-up info traffic
Reviews: Raimondeau and Vlachos, Chem. Eng. J. 90, 3 (2002);Chatterjee et al., Chem. Eng. Sci., ISCRE Issue (2004);Vlachos, Adv. Chem. Eng. . 30, 1 (2005)
• Previous work focused usually on a single scale and one way of information passingdeveloped structure-properties relations (molecular descriptors) without attention to processing
7
The Multiscale Simulation Paradigm: Predict macroscopic performance from first principles
CFDMesoscopic Theory
KM, DSMC, LB, DPD, BD
MD, atomistic MC
10-2
LengthScale (m)
Time Scale (s)
Quantum Mechanics
TST
Lab Scale
10-10
10-9
10-7
10-6
10-12 10-9 10-3 103
Downscaling or Top-down info traffic
Upscaling or Bottom-up info traffic
• Challenges- Phenomena and models are
strongly coupled- Develop bridges between
models of various scales to enable accurate, robust, efficient, seamless coupling*
Reviews: Raimondeau and Vlachos, Chem. Eng. J. 90, 3 (2002);Chatterjee et al., Chem. Eng. Sci. 59, 5559 (2004);Vlachos, Adv. Chem. Eng. 30, 1 (2005)
DFT/MD Coupling
Ludwig and Vlachos, Mol. Simul. (2004)* For noise control in hybrid siml, see work by
groups of Braatz, Christofides, Vlachos
The multiscale simulation paradigm: A bottom-up ladder
CFDMesoscopic Theory/CGMC
KMC, DSMCLB, DPD,
MD, extended trajectory calc.
10-2
LengthScale (m)
Time Scale (s)
Quantum Mechanics
TST
Lab Scale
10-10
10-9
10-7
10-6
10-12 10-9 10-3 103
Downscaling or Top-down info traffic
Upscaling or Bottom-up info traffic
Reviews: Raimondeau and Vlachos, Chem. Eng. J. 90, 3 (2002);Chatterjee et al., Chem. Eng. Sci., ISCRE Issue (2004);Vlachos, Adv. Chem. Eng. . 30, 1 (2005)
• Direct multiscale simulation (hybrid, coarse graining) is possible for systems of moderate complexity
• It is plagued by computational cost for complex systems, such as chemical reactors
• Are all scales and phenomena important?
8
Hierarchical, multiscale model development
Aghalayam et al., AIChE J. 46, 2017 (2000)
Deshmukh et al., Int. J. Multiscale Comp. Eng. 2, 221-238 (2004)
Lower Level TheoryMicrokinetic model chemistry parameters
Semi-empirical techniques (BOC), TSTDensity functional theory, DFT-MD
Catalyst modelMean field approximationKinetic Monte Carlo
Fluid flow/TransportSimple reactor models (PFR, CSTR, transp. correlations)Computational Fluid Dynamics (CFD)
Hierarchical, multiscale model development
Aghalayam et al., AIChE J. 46, 2017 (2000)
Deshmukh et al., Int. J. Multiscale Comp. Eng. 2, 221-238 (2004)
Higher Level TheoryMicrokinetic model chemistry parameters
Semi-empirical techniques (BOC), TSTDensity functional theory, DFT-MD
Catalyst modelMean field approximationKinetic Monte Carlo
Fluid flow/TransportSimple reactor models (PFR, CSTR, transp. correlations)Computational Fluid Dynamics (CFD)
Feature identification toolbox enables hierarchical model development and reduction
Last theoretical level: Engineering models are needed for reactor optimization and control and for model-based catalyst design
9
Hierarchy enables rapid screening of chemistry, fuels, and catalysts
• Hierarchy adds a new dimension to multiscaling: at each scale, more than one model can be run
Semiempirical: UBI-QEP, TST
ab initio:DFT, TST, DFT-MD
Continuum/Mean-field: ODEs
Theor. parameter estimation
Catalyst & adsorbed phase description
Discrete: Kinetic Monte Carlo
Ideal: Fixed bed, CSTR, etc.
Reactor scaleComputational Fluid Dynamics
(CFD)
Hierarchy, accuracy, cost
Scale
Length
Time
Hierarchy
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
10
• Good hydrogen carrierHigh energy density Stored as a liquid (at 25 oC, 8 atm)One of the most widely produced chemicals (>100 metric tones/yr)- Haber-Bosch Process- Infrastructure is already set up
NH3 cracking for H2 production
3 2 22 NH N +3 H 46 /→ + kJ mol1 Deshmukh et al., Ind. Eng. Chem. Res. (2004)2 Ganley et al., AIChE J. (2004)
Meth
anol
Etha
nol
Meth
ane
Prop
ane
• Catalytic decomposition of pure NH3 on Ru1,2
– Slightly endothermic– Minimal downstream processing
0
5
10
15
20
25
30
% w
t
17.6%
Amm
onia
Octan
e
NH3 decomposition on Ru: 2NH3 =N2+3H2
• NH3 as a storage medium• ‘Pure’ H2 – No COx• A microkinetic model is
build using BOC and TST• Our microkinetic model
captures the trend• High N* coverages
196 µm x 84 µm x 1078 µm
Mhadeshwar et al., Cat. Letters 96, 13-22 (2004)
0
0.2
0.4
0.6
0.8
1
650 850 1050 1250T [K]
N*
Empty sites (*)
0
20
40
60
80
100
Expts. [Ganley et al.]
PFR model
Ganley et al., AIChE J.2003
*NH*NH 33 ⇔+*H*NH**NH 23 +⇔+
*H*NH**NH 2 +⇔+*H*N**NH +⇔+*2N*2N 2 +⇔*2H*2H 2 +⇔
11
DFT is used to estimate lateral interactions
Deshmukh et al., Int. J. Multiscale Comp. Eng. 2, 221-238 (2004)
• DACAPO (solid-state electronic structure package by Hammer and coworkers*)
• 3-Layer slab of Ru(0001)
• 2 × 2 unit cell
• All layers are relaxed
• Plane wave cutoff = 350eV
• 18 k-points for surface Brillouin zone
• Generalized gradient approximation (PW-91)
* Hammer et al., DACAPO version 2.7 (CAMP, Technical University, Denmark)
90
100
110
120
130
0 0.25 0.5 0.75 1(N*+H*) coverage [ML]
DFT Calculations (this work)Linear fit Q
N (at H*=0)=128.2-33.3N*
N on Ru(0001) 3 layer slab
Linear fit QN (at N*=0.25)=120.1-6.2H*
N-N interactions
N-H interactions
Exps: Ganley et al., AIChE J. (2004)
DFT-retrained microkinetic model describes the experimental data well
• H-H and N-H interactions are small
• N-N interactions completely change the chemistry
• Extensive validation against UHV and high P data has been done
0
20
40
60
80
100
650 850 1050 1250
Expts. [Ganley et al.]
PFR modelwithout interactions
T [K]
PFR modelwith interactions
0
0.2
0.4
0.6
0.8
1
650 850 1050 1250T [K]
*
N* without interactions* without interactions
H*
N*
NH3*
Deshmukh et al., Int. J. Multiscale Comp. Eng. 2, 221-238 (2004)
12
The multiscale simulation paradigm: A bottom-up ladder
CFDMesoscopic Theory/CGMC
KMC, DSMCLB, DPD,
MD, extended trajectory calc.
10-2
LengthScale (m)
Time Scale (s)
Quantum Mechanics
TST
Lab Scale
10-10
10-9
10-7
10-6
10-12 10-9 10-3 103
Downscaling or Top-down info traffic
Upscaling or Bottom-up info traffic
Reviews: Raimondeau and Vlachos, Chem. Eng. J. 90, 3 (2002);Chatterjee et al., Chem. Eng. Sci., ISCRE Issue (2004);Vlachos, Adv. Chem. Eng. . 30, 1 (2005)
• Typical objective- Mechanistic understanding- Reconcile large differences
in published data- Process optimization:
Process engineering
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
13
Computer-aided chemistry reduction• Sensitivity and Principal
Component Analyses – No a priori assumptions– Identification of important
reactions and species
• Small parameter asymptotics on species balances and site conservation– Simple algebra to derive a
rate expression2*N3
2N4N θPkθk 2*2 −=σ
2322 NNHNH -2 and 3 σ=σσ=σ0.5
HNH12104
1197
1
2N
4
3H
2
1NH
12
11
*
23223PP
kk2kkkk
kk
Pkk
Pkk
Pkk
1
1θ−++++
=ReducedModel
0
20
40
60
80
100
650 700 750 800 850 900 950
% N
H3 c
onve
rsio
n
Temperature [K]
ReducedModel
FullModel
FlowDirection
Mhadeshwar et al., Cat. Letters 96, 13 (2004)
µReactor is close to axial dispersion modelDeff vs. Geometric characteristics
• The method of homogenization is used that is based on separation of length scales
ξ1
ξ2
Ωf Ω1-Ωf
Γ
X0.50
0.25
0.00
Y
Post
PostUnit Cell
~ ( , )i ii o
u u xε ξ∞
=∑
, 1
( )m
iji j i j
u uDt x x=
∂ ∂ ∂=
∂ ∂ ∂∑ , 10
m
i iji j j
un Dx=∂
=∂∑
, 11
1 ( )[ ]f
meff k
ij is skk s s
D D dχξ δ ξξΩ=
∂= ∫ +Ω ∂∑
( ), 1 1
( ) ( ),m m
kij jk f
i j ji j j
D Dχξ ξ ξξ ξ ξ= =
⎛ ⎞∂ ∂ ∂= − ∈Ω⎜ ⎟⎜ ⎟∂ ∂ ∂⎝ ⎠
∑ ∑
, 1 1
m mk
i ij j jki j jj
n D n Dχξ= =∂
=∂∑ ∑
1L
ε =
14
Design and fab of microchemical systemsvia multiscale modeling
Contours of Velocity Magnitude
Norton et al., IECR 43, 4833 (2004)
catalyst
posts
Micromixer/reactor Hydrodynamically driven mixing
Top ViewCross Sectional View
Top ViewCross Sectional ViewCatalyst design
Microrxtr design
0
20
40
60
80
100
120
0 1 2 3 4 5 6
0 0.2 0.4 0.6 0.8 1
Pow
er o
utle
t (W
)
Inlet fuel velocity (m/s)
Fuel flow rate (SLPM)
Max. poweroutput
Materials limit
Breakthroughlimit
Attainable regions in multifunctional microdevices
• Multifunctional devices provide millisecond operation• Adjustment of flow rates can provide variable power• Scaling out can supply transportation power levels
NH3 flow rate increases
3 8 2 2 2C H +5O 3CO +4H O→
3 2 22 N H N + 3 H→Wall
50 W
H2 maker:
18 kW
Kaisare et al., IECR (submitted)
15
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
Water-gas shift reaction on Pt: CO+H2O=CO2+H2
Thermodynamics1 is important but not sufficient:kinetics ‘corrects’ the WGS speed
0
20
40
60
80
100
473 573 673 773 873
CO
con
vers
ion
[%]
Full mech.
Equil.Experiments(Xue et al.)
A/V=150 cm-1
No coupling and literature mechs
Temperature [K]
1Mhadeshwar et al., J. Phys. Chem. B (2003)
16
Reactor Superstructure Optimization
• Modeled as n PFRs in series• Side feed and side draw from each reactor• Each PFR: same length and different temperature
• Gradient-based optimizer
Yk,f, mf
ms1
Side Feed (steam)
Feedl1, T1 l2, T2 ln, Tn
ms2 msnYk,s
md1 md2 mdn
Side Draw Output
Process optimization: Optimum temperature profile in the WGS reaction
• Total length:
• Inlet: 40 sccm feed (dry basis), with 18% CO
• Steam:
• Temperature constraints:
• All cases:– No split feed or side draw– All steam utilized– T constraints were inactive
2cmii
l =∑
, 40sccms ii
m ≤∑
373K 873KiT≤ ≤
450
500
550
600
650
700
750
2 4 6 8 10
Tem
pera
ture
, T
(K)
# of PFR
218 ppm
356 ppm
COout
= 1174 ppm
Vlachos et al., Compt. Chem. Eng. (2006)
17
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
The multiscale simulation paradigm: A bottom-up ladder
CFDMesoscopic Theory/CGMC
KMC, DSMCLB, DPD,
MD, extended trajectory calc.
10-2
LengthScale (m)
Time Scale (s)
Quantum Mechanics
TST
Lab Scale
10-10
10-9
10-7
10-6
10-12 10-9 10-3 103
Downscaling or Top-down info traffic
Upscaling or Bottom-up info traffic
Reviews: Raimondeau and Vlachos, Chem. Eng. J. 90, 3 (2002);Chatterjee et al., Chem. Eng. Sci., ISCRE Issue (2004);Vlachos, Adv. Chem. Eng. . 30, 1 (2005)
• Typical objective- Mechanistic understanding- Reconcile large differences
in published data- Process optimization:
Process engineering
The multiscale simulation paradigm: A bottom up and top-down ladder
• Opportunity- Given a macroscopic
behavior, design materials and/or control nanoscale
- Product engineering
18
An example of catalyst optimization: NH3 decomposition
• Search is done on atomic descriptors while running the full chemistry model
• Libraries of computational information are created via DFT
• Models are built• Potential catalyst
candidates are identifiedHydrogen binding energy
Nitr
ogen
bin
ding
ene
rgy
Ammonia conversion (%) at 380 oC
Ulissi et al.
Outline
Decentralized, future energy productionMiniaturization differs from scaling upMultiscale modelingApplication of multiscale modeling to
Development of detailed reaction mechanismsMicroreactor designProcess optimizationCatalyst designExperiment design
19
Maximizing information content of a model
• Parameters are uncertain• Often refined using statistically based experiments• We need to bridge first-principles modeling with systems
approaches in designing experiments
SA
NoYes
Guidelinesfor reactor design
Sort
Initial model
Refine hierarchically
the model
Identify exptal
conditions
ExptInformatics database (model based)
Model globally valid?
Models database
No Mod
el
Global search
Vlachos et al., Comp. Chem. Eng. 30, 1712 (2006)
Model-based design of experiments: Ensuring global accuracy of models
• Global Monte Carlo search in exptl parameter space (τ, P, T, compos., A/V)
• Local sensitivity analysis– Only a few model parameters are important
and can be extracted, but change in manipulated parameter space
• Sort by max NSC, RDS, MARI,…
(c)
(a)
(b)
-8 -4 0 4 8
Rea
ctio
n
NH2*+* NH*+H*
N2+2* 2N*
NSC
NH3*+* NH
2*+H*
H2+2* 2H*
2H* H2+2*
2N* N2+2*
NH*+* N*+H*N*+H* NH*+*
NH*+H* NH2*+*
NH2*+H* NH
3*+*
NH3* NH
3+*
NH3+* NH
3*
-15 -10 -5 0 10 15
Rea
ctio
n
NH2*+* NH*+H*
N2+2* 2N*
NSC
NH3*+* NH 2*+H*
H2+2* 2H*2H* H 2+2*
2N* N 2+2*NH*+* N*+H*
N*+H* NH*+*
NH*+H* NH 2*+*
NH2*+H* NH 3*+*
NH3* NH 3+*NH3+* NH 3*
-0.5
0
0.5
1
1.5
2
0 20 40 60 80 100Ammonia conversion [%]
NH3*+*=NH2
*+H*
|NSC
|
731 0.3 2.0 1541 0.02 0.50 0.48
Vlachos et al., Comp. Chem. Eng., 30, 1712 (2006)
20
Normalized parameter sensitivity vs. conversion (CSTR)
NH2*+*=NH*+H* is the most sensitive reaction
H2
NH3
N2
RDS R2RDS R3 RDS R4
D Optimal
E Optimal
A OptimalRDS R1
RDS R5 RDS R6
500600
700800
900
0
1
2
3
4
50
5000
10000
15000
Sur
face
are
a / v
olum
e (1
/cm
)
Residence time (s) T (K)
RDS R5RDS R3
RDS R4E Optimal
A Optimal
RDS R1RDS R6D Optimal
RDS R2
Optimal statistical and physics-aided designsCompared D, A, E and physics-aided designs
Inlet composition space – no clear pattern
Optimum at relatively low temperature, intermediate cat. surface area/reactor volume
Prasad and Vlachos, Ind. Eng. Chem. Res. (submitted)
21
D optimal metric
Kinetic relevance –D Optimality and partial equilibrium (PE)
High values of the det. of the Fisher info matrix correlate with fewer reactions in PE and farther from equil.
Non PE
PE|PEI|: NH2 *+*=NH*+H* |PE
I|: N2+2
*=2N*
Prasad and Vlachos, Ind. Eng. Chem. Res. (submitted)
Partial equil. index:PEI=rf/(rf+rb)
0
0.2
0.4
0.6
0.8
1
3 4 5 6
Highest det(FIM)Mid-value det(FIM)Lowest det(FIM)
Prob
abili
ty
No. of reactions in partial equilibrium
Approach to overall equilibrium
Kinetic relevance –D Optimality and sensitivity coefficients
D optimal metric
High values of the Fisher information matrix correlate with larger normalized sensitivity coefficients of the sensitive reactions
Opt.
Equil.
|NSC|: NH2 *+*=NH*+H* |NS
C|: N 2+
2*=2N
*
Prasad and Vlachos, Ind. Eng. Chem. Res. (submitted)
22
Is a Single Optimal Point Good Enough?The D optimal response surface is highly nonlinear
You must be very close to the optimal point to ensure optimality
Experimental constraints may make this unachievable
Representative cross-section of D optimal response surface
D optimal metric vs. distance from optimal point
Prasad and Vlachos, Ind. Eng. Chem. Res. (submitted)
0
0.05
0.1
0.15
0.2
0.25
0.3
-20 -15 -10 -5 0 5
1%10%20%
Dis
tribu
tion
log(det(FIM))
Identifying regions (clusters) of D-optimal data using informatics tools
Clusters are identified using ‘partitioning among medoids’
Prasad and Vlachos, Ind. Eng. Chem. Res. (submitted)
23
Assessment of Informatics Approach
Sample anywhere within clusters (experimental flexibility)
Substantial improvement over single optimal points
Distribution of D optimal metric within optimal region and in entire parameter space
Prasad and Vlachos
0
0.05
0.1
0.15
0.2
0.25
-20 -15 -10 -5 0 5
Full ensembleWithin first D optimal region
Dis
tribu
tion
log(det(FIM))
Proof of concept via experiments
• 44 new experiments conducted in optimal region
• Varied– Temperature– Catalyst amount– Inlet composition
• Good agreement of model prediction and data
Karim, Prasad, and Vlachos (in preparation)
0
20
40
60
80
100
0 20 40 60 80 100
Con
vers
ion
% (e
xper
imen
t)
Conversion, % (model)
24
Summary and future directionsFuture energy generation will happen at much smaller scalesDownscaling is different even at the