Computer-Aided Construction of Chemical Kinetic Models and predictive analytics chemistry
Computer-Aided Construction of Chemical
Towards Predictive Chemistry
William H. Green
MIT Dept. of Chemical Engineering
DOE Combustion Meeting
Potomac, MD, May 2014
1 Why do we want predictive fuel chemistry
• Historically, Fuels have changed pretty slowly… Engines evolved faster,
using more-or-less static fuels. Fuel was a constant, not a variable…
• But now, fuels are changing
– To reduce greenhouse gas, laws require renewable fuels
– Conventional light sweet petroleum is getting more & more expensive
– Possible to make many future fuels with synthetic biology & chemistry
• If fuel is a variable, we need to be able to predict how engine+fuel system
depends on fuel composition, not just on engine & operating conditions.
– Much faster to explore new concepts on computer than by building prototype
engines and synthesizing hundreds of barrels of a new fuel for engine tests
• Scientifically, we don’t really understand something until we can make
quantitative predictions…and discrepancies between those predictions and
experiments reveal opportunities to learn something new
2 Challenge: Lots of alternative fuels to evaluate
Not practical to synthesize & test every possible future fuel in every possible
…rapid, easy predictions are valuable and needed
3 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
4 Fuel Chemistry models are very big
• Many chemical intermediates between fuel and CO .
• Multiple competing pathways
• Nonlinear, sometimes non-monotonic kinetics
• Real fuels (gasoline, jet, diesel) have many chemical species
– So even initial conditions are complicated
• Models built by hand: 1,000 species, 5,000 reactions
– Human tries to include just the more important reactions.
• Models built by computer: 500 species, 10,000 reactions
– RMG software typically considers 30,000 species and 100,000 reactions
• most of them are found to be numerically negligible
5 We need this black box
and solves the
Similar issues for fuel-formation chemistry e.g. in refinery
6 Commercial software can solve
detailed kinetic simulations…
…if one can supply the fuel’s
Diff. Eq. solver
dY/dt = …
7 How we construct fuel chemistry models
Diff. Eq. solver
dY/dt = …
8 Chemical Kinetic Modeling Challenges
• Identify all important reactions & species
– But not unimportant species & reactions: how to distinguish?
• Estimate all reaction rate coefficients (and properties,
e.g. thermochemistry) to sufficient accuracy.
– We use Functional Group extrapolations & Quantum Chemistry
• Large models pose numerical and computer problems
– Very challenging for humans to handle, interpret, debug…
…SO WE TRY TO AUTOMATE EVERYTHING
We build on prior efforts by combustion research community, e.g.
Comprehensive Chemical Kinetics 35 (1997)
Advances in Chemical Engineering 32 (2007)
9 RMG algorithm: Faster pathways
explored further, growing the model
“Current Model” inside.
RMG decides whether
or not to add species to
Final model typically
500 species, 8000 rxns
Open-Source RMG software.
10 Kinetic Model Predictions Rely on
Quantum Chemistry for Thermo, Rates
• Functional Group approximation
– Compute a few examples of each reaction type with quantum, then use same
barrier, A factor for analogous reactions.
• Most of our calculations at CBS-QB3 level
– Geometries, Vibrational Frequencies from DFT
– Single point energies at stationary points at higher level
– Extrapolation to Basis Set Limit
• Recent calculations use F12 methods
– Explicit dependence on distance between every pair of electrons
– Much faster basis set convergence
• Most calculations rely on common approximations
– Rigid-Rotor Harmonic-Oscillator approximation
– Conventional Transition State Theory (dividing surface at saddle point)
– Simple corrections for internal rotors and tunneling
– Modified Strong Collision approx. for k(T,P)
Are computed thermo, rates accurate enough??
Testing Accuracy of Model Predictions vs.
Experiment: CEFRC data on Butanols
With collaborations from
other institutes like
Univ. Ghent, NIST.
Rare Situation where detailed data available at many different conditions
12 We used RMG to build a mechanism for
butanol pyrolysis and combustion.
Four isomers, very different octane numbers.
Octane number = 86 Octane number = 98 Octane number = 100
More reactive Less reactive
RMG considered about 30,000 possible species, selected
• 372 chemical species
• 8,723 reactions
Important k’s with quantum chemistry
Shamel S. Merchant, E.F. Zanoelo, R.L. Speth,
M.R. Harper, K.M. Van Geem and William H. Green,
Combustion & Flame (2013)
13 Model predicted butenes yield from
butanols pyrolysis accurately
T 1000 K
P 2 bar
Van Geem et al.
0 5 10 15
Experimental butene yield = wt%
Predicted butene yield = wt%The kinetic model also quantitatively predicts
formation of aromatics from butanols
(via rather complicated reaction sequences)
Data from K. Van Geem, Ghent
pyrolysis of iso-butanol 1000 K, 2 atm, 2 seconds
15 When you change reaction conditions, need
to add reactions, compute rate coefficients
Predictions of 1000 K pyrolysis model for 1467 K
H O concentration
16 Microsecond H O formation at 1400 K
sensitive to different reactions than
long-time product formation at 1000 K.
Early Times 10 sec
So… computed improved estimates of bond scissions
& dehydrations based on quantum chemistry…. Quantum calcs for k(T,P) sensitive at
1467 K significantly changes predictions
H O concentration
18 Using quantum calcs (rather than rough estimates by
analogy) for most sensitive high T reaction k(T,P)
improves predictions of shock tube pyrolysis:
It is always a good idea to do the quantum calcs for the
Sensitive numbers in the model
H O concentration
Experimental data: Stranic et al., Combust. Flame (2012).
See also Rosado-Reyes et al., J. Phys. Chem. A (2013).
19 Advanced Light Source allows quantification of
dozens of species including key radicals in flames
• Flames are analyzed with molecular
beam time-of-flight mass
• Photoionization with tunable
photons allows identification of
• by mass
• by ionization energy
Data measured by Nils Hansen (Sandia)
at Advanced Light Source (LBNL).
Hansen et al. PCCP (2012)
Hansen et al. Combust. Flame (2013)