Understanding and Predicting
Complex Gas Phase Kinetics
William H. Green
Hoyt C. Hottel Professor & Executive Officer
MIT Department of Chemical Engineering
Frontiers in Chemical Reaction Engineering
Lab. Chem. Tech., Univ. Gent
June 25, 2013
LCT has long been in the forefront of
understanding & predicting gas phase
reaction kinetics
• World leader in steam cracking kinetics, both
experimental and modeling
• Detailed reaction modeling for more than 25
years; accurate simulators used by industry.
• Significant contributions to methodology,
computation/estimation of rate coefficients…
But today I’ll tell you mostly about work done recently at MIT,
some in collaboration with LCT researchers
Automated Predictive Chemical Kinetics:
The Concept & Motivation
• Rapid, Easy Predictions are valuable & needed
– Assess alternative fuels & feedstocks
– Design new processes and engines
• Hydrocarbon/Fuel chemistry is complicated
– many reactions, species): need to automate
What an Engineer Really Wants
Operating
Conditions
Chemistry Knowledge
(clearly documented)
Reactor or
Engine Design
Feed/Fuel
Composition
We need this
Black Box
Predicted Performance
using this Feed or Fuel
(with error bars)
Computer Builds
& Solves the
Reaction Model
(automate modelbuilding just like
model-solving has
already been
automated.)
Commercial software can solve
detailed kinetic simulations…
Simulation
Predictions
Numerical
Diff. Eq. Solver
Very
Long
Simulation
List
Eqns
of
dY/dt = …
Interpreter
Reactions
with
Rate
Parameters Popular software:
CHEMKIN, Cantera, KIVA, GTPower
Commercial software can solve
detailed kinetic simulations…
…if someone can supply the
reaction mechanism.
Simulation
Predictions
Numerical
Diff. Eq. Solver
Very
Long
Simulation
List
Eqns
of
dY/dt = …
Interpreter
Reactions
with
Rate
Parameters Popular software:
CHEMKIN, Cantera, KIVA, GTPower
A Systematic Way to Construct the Models
Simulation
Predictions
Chemistry Knowledge
(some data, mostly
generalizations)
Unambiguous
Documentation
of Assumptions
about how
molecules react
Rxn Mechanism
Generator
Numerical
Diff. Eq. Solver
Very
Long
List
of
Reactions
with
Rate
Parameters
Interpreter
Simulation
Eqns
dY/dt = …
Current paradigm: CHEMKIN
Predictive Chemical Kinetics Challenges
• Identify all important reactions & species
– But not unimportant species & reactions: how to distinguish?
• Estimate all reaction rate coefficients (and molecular
properties, e.g. thermochemistry) to sufficient
accuracy.
– So many reactions!
– We use Benson-type extrapolations & quantum chemistry
• Large models pose numerical and computer problems
– Very challenging for humans to interpret or debug
• Identify source of discrepancy between predictions of
a large model & experimental data
Comprehensive Chemical Kinetics 35 (1997)
Advances in Chemical Engineering 32 (2007)
RMG algorithm: Faster pathways are
explored further, growing the model
Before:
“Current Model” inside.
RMG decides whether
or not to add species to
this model.
Final model typically
~500 species, 5000 rxns
Open-Source RMG software.
Download from
rmg.sourceforge.net
After:
Our Model-Construction Procedure
• Computer assembles large kinetic model for
particular conditions using rough estimates of rate
coefficients k to decide which species to include.
– Open-source software package RMG: rmg.sourceforge.net
• If sensitive to k derived from rough estimate,
recompute that k using quantum chemistry.
– Unfortunately, quantum calcs for rates not fully automated.
– Generalize from quantum to improve rate estimation rules.
• Iterate until not sensitive to any rough estimates.
• Repeat for different conditions (Co,T,P).
• Compute prediction & compare with experiment.
What do we Expect from Model vs. Data
Comparisons?
• At present, Thermo rarely known better than 1
kcal/mole, Ea’s uncertain by ~ 2 kcal/mole, and A’s
often uncertain by factor of 2. So….
– we don’t expect perfect agreement!
– Precise agreement means model parameters were
fitted to match experiment, not predictions. Or lucky.
• However, we think our estimates are reasonable,
and our software is pretty good.
• So… we expect discrepancies to be less than an
order of magnitude for both overall reaction
timescale and product distribution.
– Often good enough for predicting trends, and
understanding origin of strange engine phenomena
Test Case #1: Proposed Alternative Fuel
Butanol: can we predict its pyrolysis/combustion?
Engine makers would love a predictive model…
The New York Times
October 23, 2012
“Corn Ethanol Makers
Weigh Switch to Butanol”
Gevo has built 22M gallon/yr
butanol plant in Minnesota…but
is losing money….
Butanol beats ethanol “blending
wall” & satisfies RFS2 standard
Gevo & Butamax (BP/DuPont j.v.)
now litigating key patents
So RMG built us a mechanism for
butanol pyrolysis & combustion
Four isomers, very different octane numbers.
- isobutanol is most promising
Considered about 30,000 possible species,
CH3
selected as important:
HO
• 372 chemical species
CH3 H C OH CH
• 8,723 reactions
H C
OH
CH
3
3
3
Shamel S. Merchant, Everton Fernando Zanoelo, Raymond L. Speth,
H3C
Michael R. Harper, Kevin M. Van Geem and William H. Green,
Combust. Flame (2013, accepted)
3
CH3
OH
Triumph: Butenes yield from Butanols
pyrolysis predicted accurately (within 50%)
Predicted butene yield [=] wt%
15
1-Butanol (1-Butene)
iso-Butanol (iso-Butene)
2-Butanol (1-Butene)
2-Butanol (2-Butene)
tert-Butanol (iso-Butene)
10
Pyrolysis,
T ~1000 K
P ~ 2 bar
t ~ seconds
5
0
Measured by
Van Geem et al.
Univ. Ghent
0
5
10
Experimental butene yield [=] wt%
15
Similar accuracy for many other species’ yields.
Computer-Generated model accurately predicts
benzene & toluene, but not cyclopentadiene
Cyclopentadiene
Data from K. Van Geem, Ghent
pyrolysis of iso-butanol
Discrepany: Predicted Benzene Yield from
tert-Butanol Pyrolysis Off by Order of Magnitude
2
Predicted benzene yield [=] wt%
1.8
1.6
Pyrolysis,
T ~1000 K
P ~ 2 bar
t ~ seconds
1-Butanol
iso-Butanol
2-Butanol
tert-Butanol
1.4
1.2
Factor of 3
Discrepancy
1
20% discrepancies
Measured by
Van Geem et al.
Univ. Ghent
0.8
0.6
0.4
0.2
0
0
RMG algorithm
expected to be less
accurate for minor
products.
Order of magnitude
discrepancy
0.5
1
1.5
Experimental benzene yield [=] wt%
2
~1000 K model not great for ~1400 K
Stanford fast-pyrolysis experiments
Initial Model Predictions (no quantum calcs)
H2O concentration
OH concentration
Data: “Multi-Species Laser Measurement of n-Butanol Pyrolysis behind
Reflected Shock Waves”, R. Cook et al., Int.J.Chem.Kinet. (2012).
Microsecond H2O formation at 1400 K sensitive to
different reactions than long-time organic product
formation at 1000 K.
Bond scissions
dehydration
Stanford pyrolysis
of n-butanol
Early Times
t=3 µs~ 10-9 sec
So… computed improved estimates of bond scissions & dehydrations
based on quantum chemistry….
Triumph: After quantum calcs for sensitive k(T,P),
predictions satisfactorily close to experiment
Data: “Multi-Species Laser Measurement of n-Butanol Pyrolysis behind Reflected Shock
Waves”, R. Cook et al., Int.J.Chem.Kinet. (2012).
NIST (Rosado-Reyes & Tsang) measurements coming soon
Expand Model to Include Low P Flame Conditions
Starting Mechanism
Validated under high pressure combustion
and pyrolysis
334 species,
7113 reactions:
4288 k(T,P) kinetics
Flames 1 to 4
T = 300 – 1800 K
P = 30 torr
Merchant et al., Combust and Flame.
2013, accepted
372 species,
8723 reactions:
5398 k(T,P) kinetics
Hansen et al., Combust and Flame.
2013, accepted
Computer
considered:
>10,000 species and
>100,000 reactions
increasing pressure dependence
Often only a few small-molecule reactions are P-dependent. But
Many Rate Coefficients Strongly P-dependent in MBMS flames
B. M. Wong, D. M. Matheu, and W. H.
Green. J. Phys. Chem. A 2003, 107, p.
6206-6211.
“Normal”
Chemistry
MBMS Flame
increasing pressure dependence
Chemical Activation is a major complication!
Reactions faster than thermalization: Energized Initial product immediately reacts
Thermalization rate increases with pressure: k(T,P)
A+B can lead to many possible products due to “well skipping”
RMG automatically tracks down all the channels.
Example: H-catalyzed keto-enol tautomerism of propenol
Instead of 2 possible products, 10 are formed.
Instead of 2 Transition States, must compute 16 TS’s.
Lots of Unstable Enols formed in Flames
improved prediction for enols and their corresponding
aldehydes
kinetic model includes H-catalyzed keto-enol
isomerization
Enols sensitive to H-atom-catalyzed chemicallyactivated keto-enol tautomerization
+
+
“Chemically-activated” = “product reacts faster than thermalized”
Chemical activation is a major complication in automated reaction
generation: keep track of “well-skipping” reactions, compute k(T,P)
Instead of 2 possible products, 10 are formed.
Instead of 2 Transition States, must compute 16 TS’s.
H atom is one of the major radicals produced in hydrocarbon flames
Very little acid present to catalyze keto - enol tautomerization
Water catalysis (in gas phase flame) is slow, not significant
Flame Speed predictions adequate (?)
What accuracy do we need? How accurate
are experiments?
55
Laminar burning velocity [=] cm/s
iso-Butanol
50
45
40
35
Model very sensitive to
HCO + H2O = H + CO + H2O
30
Model
Prediction
25
Exptl data analysis issues
identified..
20
15
0.8
1
1.2
Equivalence ratio
1.4
Data measured at USC, by Veloo and Egolfopoulos
Models for iso-Butanol Flame Significantly Differ
MIT Model (Green et al.)
KAUSTModel
Model(Frassoldati
(Sarathy et et
al.)al.)
Milano
Many parameters in detailed kinetic models:
just because it matches expertiment does not
mean it is the truth!
Isobutanol ignition delay: predictions consistent with
recent high-T experiments (and also for low-T φ=1 expts
of Weber et al. & Hanson group)
Model Predictions quite close to experiment, better than factor of 2:
lucky? Sensitive to H2/O2 rate coefficients known very accurately.
Stranic et al., Combust. Flame, 2012, 159 (2), 516-527.
Important unresolved discrepancy: all existing
models completely miss [O2] effect on ignition
delay of butanols in engine-relevant conditions!
3.38% n-butanol, P = 15 bar
Exptl Data (Solid symbols):
Bryan Weber &
C.J. Sung, U.Conn.
2
Ignition delay / ms
10
MIT model (open symbols, dotted
lines)
φ = 0.5
1
φ = 1.0
10
φ = 2.0
1.1
1.2
1.3
1.4
1000 K / T
1.5
1.6
Mechanism construction algorithm
missing some minor intermediates .
Correct thermo, rates for ROO,
QOOH, HO2, ROOH?
Investigating O2-dependence of low-T ignition
led to discovery of new peroxy reactions
1. γ-ketohydroperoxides can isomerize to cyclic peroxides,
which unimolecularly decay to non-radical products
(rather than radical chain branching pathway)
-C(O)CH2CH(OOH)-
cyclic peroxide
-C(O)OH + CH3C(O)-
2. γ–hydroxyperoxy radicals eliminate H2O
(rather than radical chain branching pathway)
-CH(OH)CH2CH(OO•)-
H2O + -C(O)CH2CH(O•)-
Welz et al., J. Phys. Chem. Lett. (2013)
While we can predict a lot, still more chemistry to be discovered!
PES for the Cyclic Peroxide Pathway
Jalan et al., J. Am. Chem. Soc. (2013, accepted)
Overall:
More Triumphs than Challenges in C4
range. Still some issues in small molecule
combustion, but making good progress.
How about bigger molecules?
Nick Vandewiele of LCT has shown RMG
can do a good job for pyrolysis of JP-10, but
starting to hit memory limitations.
Moving beyond hydrocarbon pyrolysis
+ combustion chemistry
• Recent advances in organosulfur chemistry,
see for example work of Vandeputte et al. at
LCT, Class et al. at MIT
• Organonitrogen chemistry is being introduced
into RMG now by Buesser
• Solution phase kinetic models can also be
built by RMG, but RMG only knows a few
types of solvent effects…
Automated Predictive Chemical Kinetics:
Status as of June 2013
• Method seems sound. Where we have executed the
plan (e.g. butanols), predictions are usually accurate.
• Many Challenges, including:
– Ensuring reliable databases.
– Larger molecules greatly increase complexity,
computational cost, opportunities for mistakes and
misinterpretations.
– Automation will not solve all our problems: if reactions
are unknown, or if we have an incorrect number in our
databases, or if quantum chemistry gives a bad number,
the predictions are going to be wrong. But automated
Predictive Chemical Kinetics can help us quickly identify
where work is needed.
We are on our way to being able to
automatically predict pyrolysis &
combustion of alternative fuels/feeds,
even for technologically interesting (i.e.
really complicated) reacting mixtures.
The Goal is in sight.
Lots of the most interesting and
challenging work remains to be done.
The People who Did the Work (and the
agencies that paid for it)
I. Alecu, Michael Harper, Gregory
Magoon, Shamel Merchant, Nathan Yee,
G. Marin, K. Van Geem, E. Zanoelo, B.
Weber & C.J. Sung, P. Veloo & F.
Egolfopoulos, R. Cook, I. Stranic, D.
Davidson & R. Hanson, S.J. Klippenstein
MIT, U. Gent, UFPR (Brazil), U. Conn., USC,
Stanford, Argonne
Funded by: US DOE (CEFRC), Aerodyne, FWO (Belgium), CAPES
(Brazil). Heavily relies on RMG software developed with DOE
funding, and databases developed with NSF funding.