1. ECN 4 – topic 4 Chemistry effects on ignition
Evatt Hawkes
University of New South Wales, Australia
Fourth Workshop of the Engine Combustion Network,
Kyoto, Japan, 6 September 2015

2. Background: ECN3 recap All models over-predicting ID in spray flames.
Spray flame ignition delay results from ECN3
Shock tube ignition delay results from ECN3
All models over-predicting ID in spray flames.
Closest available shock-tube data suggested the problem could be the chemistry mechanisms which appear to over-predict ID at low T.

3. Session Objectives Prior to ECN4:
Try to improve our predictions of ignition delay, focussing on spray A
Try to pin down whether the problems are with the chemistry or elsewhere (TCI and other sub-models, numerics)
Working with chemical kinetics experts improve and validate chemical kinetic sub-models having manageable size
At ECN4:
Comparisons of different models (kinetic and otherwise) in prediction of experimental ignition delay
Shock tube data as well as spray A
Quantitative comparisons of species relevant to ignition
Understand the reasons behind any differences ignition behaviour, and any interesting features of these ignitions

4. Talk outline Motivation and objectives Discussion of chemistry models
Comparison to new shock tube data
Features of the modelled ignitions
Comparisons of model and experiment for ID and LOL for Spray-A parametrics
Spatial comparison at early times
Conclusions & recommendations

5. Chemistry mechanisms
Mechanisms available at ECN3 (manageable sized ones)
Luo: 106 species (topic 5 baseline), reduction of detailed LLNL mechanism (Sarathy et al) by UConn and ANL
Luo, Z., Som, S., Sarathy, S.M., Plomer, M., Pitz, W.J., Longman, D.E., and Lu, T., Development and validation of an n-dodecane skeletal mechanism for spray combustion applications, Combustion Theory and Modelling, 18(2), 2014:
Pei, 88 species, further reduction of Luo mechanism by UConn for UNSW; mechanism gives almost identical results to Luo and can be considered the same.
Pei, Y., Hawkes, E.R., Kook, S., Goldin, G.M, Lu, T., Modelling n-dodecane spray and combustion with the transported probability density function method, Combust. Flame 162(5), 2015:
Narayanaswamy: 255 species, started with LLNL, updated kinetics for small hydrocarbons, lumping+reduction and revised some rates to improve high T performance
Narayanaswamy, K., Pepiot, P., Pitsch, H., A chemical mechanism for low to high temperature oxidation of n-dodecane as a component of transportation fuel surrogates, Combust. Flame 161, 2014: 866–884.

6. Chemistry mechanisms New mechanisms since ECN3
Wang: 100 species, ERC Wisconsin, based on earlier Ra and Reitz mechanism; note that rates were optimised to shock-tube data for n-decane
Wang, H., Ra, Y., Jia M., Reitz, R.D., Development of a reduced n-dodecane-PAH mechanism, and its application for n-dodecane soot predictions, Fuel 136 (2014) 25–36.
Yao: 54 species, reduction and re-optimisation of LLNL by Uconn and ANL; note that spray A was included in the optimisation targets.
Yao, T., Pei, Y., Zhong, B.-J., Som, S., Lu, T., A hybrid mechanism for n-dodecane combustion with optimized low-temperature chemistry, 9th U. S. National Combustion Meeting, May 17-20, 2015, Cincinnati, Ohio.
Polimi: 96 species; reduced mechanism based on comprehensive Polimi detailed mechanism
A. Frassoldati, G. D’Errico, T. Lucchini, A. Stagni, A. Cuoci, T. Faravelli, A. Onorati, E. Ranzi, Reduced kinetic mechanisms of diesel fuel surrogate for engine CFD simulations, Combustion and Flame, in press.
Cai: 57 species from Aachen – reduction and re-optimisation of Narayanaswamy mechanism
Cai. L., Kroger, L., Pitsch, H., Reduced and Optimized mechanism for n-Dodecane Oxidation, 15th International Conference on Numerical Combustion.

7. Shock-tube data
φ=1
φ=2
Prior to ECN4, only data for n-C12 was for φ=1 and 20 bar, closest available conditions were n-C10 at 50 bar
New shock tube data obtained from Stanford (Davidson, Hanson) and Rensselaer Poly. (Oehlschlaeger)
Data reasonably consistent, though no one-to-one comparisons possible.
Rensselaer n-C12 & 60 bar and Aachen n-C10 & 50 bar for φ=1 and 2 to be used for comparison to models

8. Performance of mechanisms
φ=1
φ=2
Luo & Pei (UConn/LLNL) based on big LLNL over-predict almost everywhere and show way too much NTC.
Narayanaswamy (Aachen) improved high T results considerably compared with Luo (and LLNL not shown), but still low T under-prediction.
New Yao (UConn/LLNL) mechanism under-predicts & too much NTC. Tuned to spray cases, not shock-tube cases.
New Cai (Aachen) under-predicts.
New Polimi under-predicts.
New Wang (ERC) excellent agreement. Tuned to match the n-decane data at 50bar.

9. Behaviours in mixing space
Ignition delays from homogeneous reactor calculations with different mixture-fractions and initial T according to adiabatic mixing.
Most reactive
900K
1100K
For spray A conditions, model predicts two-stage ignition for φ up to about 2, single-stage low-T ignition for richer conditions.
For higher T, lean mixtures undergo single-stage, high T ignition, while moderately rich mixtures undergo two-stage ignition and very rich mixtures undergo single-stage low-T limit ignition.
Spray A conditions are very hard because high-T, low-T and NTC all involved!

10. Behaviours in mixing space
Simpler to order the mechanisms when most reactive mixture-fraction taken into account.
Generally: Yao < Cai < Polimi < Wang < Narayanaswamy < Luo/Pei
Mixture-fraction of ignition mostly consistent between Luo, Narayanaswamy, Yao; Polimi, Cai and Wang ignite richer.

11. Quenching behaviour
Flamelet calculations with different scalar dissipation rates, spray A conditions.
Ordering of ignition delays at zero dissipation rate is not the same as the quenching limit.

12. Spray experiments and models
no change since ECN3; used ECN3 data as presented in Section 2.1 Combustion Indicators
ignition delay (ID) based on broadband chemiluminescence, lift-off length (LOL) based on OH*
averaged data from Sandia, CMT, IFPEN (acknowledgements please see ECN3)
Models:
ANL: Argonne National Laboratory (ANL): Yuanjiang Pei, Muhsin Ameen, Sibendu Som
CMT-Motores Térmicos. Universitat Politècnica de València: Eduardo Pérez, Alberto Viera, J Peraza
ETH-Zurich: Sushant S. Pandurangi, Yuri M. Wright, Konstantinos Boulouchos
PoliMi: Politecnico di Milano: Gianluca d’Errico, Tommaso Lucchini
TUE: Technische Universiteit Eindhoven: L.M.T. Somers
UNSW: University of New South Wales, Australia: Aqib Chishty, Michele Bolla, Evatt Hawkes; input from Y Pei (now ANL)
USyd: University of Sydney: Fatemeh Salehi, Matt Cleary

13. Model contributions Group TCI model(s)* Chemistry model(s) ANL
Large-eddy simulation with well-mixed (LES)
Luo et al.
Yao et al.
CMT
Unsteady flamelet progress-variable (UFPV)
Narayanaswamy et al.
ETH
Conditional moment closure (CMC)
Polimi
Multiple representative interactive flamelets (MRIF)
Yao et al., Luo et al., Wang et al., Narayanaswamy et al., Polimi
TUE
Flamelet-generated manifold (FGM)
- Manifold from homogeneous reactors or flamelet
UNSW
Transported probability density function (TPDF) & well mixed model (WM)
All but Narayanaswamy
USyd
Multiple mapping conditioning LES
(MMC-LES)
*All models are RANS-based except ANL and USyd

14. Legends
TCI
model
Kinetic
mechanism
Institution

15. Ignition delay – T variation
Trends OK but considerable scatter
We are still talking about factors of two differences between model and between model and experiment
This is presumably not acceptable?

16. Ignition delay – T variation
ECN3
ECN4
Compared with the situation at ECN2 & ECN3 there are some more accurate results

17. Ignition delay (log scale)
Relative errors are similar across the T range, actually slightly higher for higher T.

18. Ignition delay – O2 variation
Again trends OK but lots of scatter
O2 trends are quite similar to T trends and will not be further discussed in detail

19. Luo Mechanism
For ID, scatter is greatly reduced when mechanism is held fixed. Not much sensitivity to the TCI model and numerical approach.
Similar to ECN2 and ECN3, models over-predict ID with Luo chemistry. (The reason for having this session.)
LOL still has a lot of sensitivity even when kinetics are fixed. LOL is sensitive to TCI and/or numerical approach.

20. Luo Mechanism
ECN3
ECN4
At ECN4, improved convergence of different modelling results at low T if chemistry held fixed
Also true at higher T if ANL’s LES result not taken into account

21. Wang, Polimi Mechanisms
Similar to the case for Luo mechanism, when the chemistry is constrained the scatter is lower.
Wang mechanism over-predicts ID but slightly better than Luo. Note that Wang mechanism gave the best results for the shock-tube data.
Polimi mechanism is great for UNSW well-mixed model, slightly worse for Polimi’s MRIF. Polimi mechanism generally better than Luo.

22. Yao mechanism
Very good results for this mechanism with 4 different models.
This mechanism was tuned based on spray flame results.

23. Comparison to homogeneous reactors
UNSW well-mixed results compared with corresponding most-reactive homogeneous reactors (HRs)
HRs generally show the same trends as the spray flame model.
HRs are mostly shorter than the spray flame model. This is expected because of the finite time to form flammable mixture, strain effects, etc.
However, sometimes the differences are not.
Strain-rate effects (not shown, but looked at) did not seem to explain why some mechanisms seem to show more differences. E.g. Wang mechanism has the highest quenching dissipation rate but shows large differences between HR and spray flame models.

24. Ignition delay – T variation
Cai mechanism under-predicts.
Narayanaswamy mechanism over-predicts

25. TCI Effect
With the model constrained to one group (with two models done in the same code), ID is all about the chemistry sub-model.
Similar to ECN2 and ECN3, well-mixed model (which ignores turbulent fluctuations) gives slightly longer ID that TPDF (which models full statistics of fluctuations). For mechanism that under-predict ID, this causes TPDF results to be worse than WM. The effect is pretty small though.
Similar to ECN2 and ECN3, LOL is much more affected by the inclusion of TCI. LOL is less sensitive to chemistry than ID, with fixed TCI model.

26. RANS versus LES
Similar to the comparison of TPDF and WM, comparison of RANS and LES shows ID and LOL are shorter for LES.
Inclusion of fluctuations shortens ID and LOL. RANS versus LES seems to have a bigger effect than TPDF versus WM. Unclear why. Might just be 3 different codes. Could also be statistical convergence?

27. Error analysis ID
LOL
ECN2,3
ECN4
New mechanisms developed since ECN3 significantly improve ID. ID is strongly affected by chemistry.
Improvement in LOL is also there but less prominent. LOL is affected by chemistry, but less strongly.

28. Why is it so? Analysis of Y_OH transport equation
from TPDF method – steady flame period
Centreline profiles of mixture-fraction
and Y_OH at ignition time
Ignition occurs in a region of low gradients behind the head of the jet
i.e. region of low macro-mixing between reacting fluid and cold ambient
Turbulent diffusion is a significant influence
Pei, Hawkes, et al. Combust. Flame, submitted

29. Yao mechanism Good ID results and 4 models
=> Opportunity to drill down on spatial data

30. Spatial structure @ 0.4 ms Formaldehyde IFPEN PLIF ANL Polimi UNSW WM
TPDF

31. Spatial 0.4 ms OH
IFPEN
PLIF
ANL
Polimi
UNSW WM
UNSW
TPDF

32. ECN4 Kyoto 2015/Flame Structure
Early stages clear differences (AR) OH
CH2O
OH*
ECN4 Kyoto 2015/Flame Structure

33. Summary & conclusions ECN2/3 Between ECN3 and ECN4
All groups under-predicted ignition delay, sometimes considerably
Between ECN3 and ECN4
Mechanism performance at low T, high P conditions seemed to be problematic based on available shock tube data for n-decane.
Problem identified not to be in chemistry reduction but to also be present in large detailed mechanisms; not just n-dodecane but also other alkanes.
A number of reduced chemistry mechanisms addressing this proliferated, some of which adjusted/optimised the low temperature rates.

34. Summary & conclusions ECN4 findings
New shock tube data requested and obtained. Mostly consistent with earlier n-C10 data.
Still a lot of scatter in ID results – factors of two
However, better convergence of results between different groups when chemistry held fixed (compared with earlier ECNs).
ID seems to be mainly about chemistry and not much affected by TCI/numerical approach (consistent with ECN3 and some but not all ECN2 results). TCI effect results in slightly shorter ID.
All the new mechanisms improved results for the spray flames. Yao et al. mechanism particularly good.
BUT new mechanisms are now mostly under-predicting the shock-tube data.
Underlying problems in detailed chemistry not fixed – are we getting the right answer for the wrong reason?
Still insufficient canonical data in spray-A relevant conditions.
ECN4 topic 4 “achieved objectives”, but did it do so for the right reasons?

35. Recommendations Observations Recommendations
ECN started on spray A fuel and conditions where there was not a lot of of fundamental data useful for validating kinetic models.
Even large detailed kinetic models turned out to be inadequate, probably because they had not been tested in spray A conditions.
Current approach to improve with demonstrated success is basically empirical adjustment against certain targets.
If the targets emphasize spray A then of course we can get a good result, but this approach cannot be projected into new situations.
Recommendations
Abandon Luo mechanism as baseline. Require that all groups submit the new (TBD) baseline. Additional mechanisms optional.
Seek greater engagement with chemical specialists – experiments and modelling
Need more fundamental data from shock tubes, flow reactors, etc., in relevant conditions (some progress since ECN3 but not “there” yet – e.g. no reduced O2 data)
Need improved detailed models which can then be reduced – probably a better approach that reducing a non-performing detailed model and “optimising”
Fuels and conditions roadmap… chemists need time to generate new fundamental data and validated kinetic models – can we specify a roadmap in advance?
Further drill down looking at space- and time-resolved data required to understand structural differences between models despite similar ID

36. Recommendations Observations Recommendations
Spray A – no fundamental data for kinetics, no validated models in conditions
Detailed kinetics failing
Current approach to fix = tuning
Recommendations
Abandon Luo mechanism as baseline. Require that all groups submit the new (TBD) baseline. Additional mechanisms optional.
Seek greater engagement with chemical specialists – experiments and modelling
Fuels and conditions roadmap… chemists need time to generate new fundamental data and validated kinetic models – can we specify a roadmap in advance?
Further drill down looking at space- and time-resolved data required to understand structural differences between models despite similar ID

37. n-Dodecane Auto-Ignition at Low Temperature
Spray A: Simulation and Experiments
Shock Tube: Simulation and Experiments
Ignition delay was defined as the time from the start of injection to the time where the maximum rate of rise of the maximum Favre-averaged temperature occurs, which has been suggested at the Engine Combustion Network (ECN) workshops
Comparison of ignition delay between experiments and computations from various numerical models[1]
Comparison of ignition delay between experiment[2] and model[3]
Inhibited ignition behaviors of mechanisms in spray simulation
Over-prediction of ignition delay times at low temperatures
[1] Y. Pei et al., Combust. Flame, 162 (2015)
[2] S. Vasu et al., Proc. Combust. Inst. 32 (2009)
[3] K. Narayanaswamy et al., Combust. Flame 161 (2014)

38. Optimized reduced mechanism with 55 species
n-Dodecane Mechanism Reduction and Optimization
Short and Accurate Mechanism for N-Dodecane
Optimized reduced mechanism with 55 species
Automatic model reduction + optimization
Example: Short n-Dodecane mechanism based on Narayanaswamy et al.[3]
Strong reduction
55 species
Potential inaccuracies
Optimization for various targets
Excellent agreement with experiments
Potential further reduction
Exclusion of certain targets
Quasi steady state assumption
Mechanism accurate, but
Accuracy through optimization
Root cause of inaccuracies not fixed
Comparison of ignition delay between reduced models[4,5]
Compact size, optimized model performance, non-stiff
[1] Z. Luo et al., Combustion Theory and Modelling 18 (2014)
[2] H. Wang et al., Fuel 136 (2014) 25-36
[3] K. Narayanaswamy et al., Combust. Flame 161 (2014)

39. New detailed mechanism based on optimized reaction rate rules
n-Dodecane Mechanism Development
Errors NOT specific for n-dodecane models
Systematic errors in chemical mechanisms in literature[2]
Thermodynamic data
Rate rules
Error compensation
Mechanism improvement in three steps
Revision of rate rules and thermo data by Curran[3]
Optimization of rate rules for series of n-alkanes
Uncertainty quantification with Bayesian framework
Calibrated for C7-C11 normal alkanes  Prediction improvement for C7-C11
New n-dodecane mechanism based on optimized rate rules
New detailed mechanism based on optimized reaction rate rules
Comparison of ignition delay between experiment[2] and model[3]
Improved prediction accuracy  Shorter ignition delay at low T  Development of short version ongoing
[1] S. Vasu et al., Proc. Combust. Inst. 32 (2009)
[2] K. Narayanaswamy et al., Combust. Flame 161 (2014)
[3] J. Bugler et al., J. Phys. Chem. A 119 (2015)

40. Performance of mechanisms
φ=1
φ=2
Blind test…
Greatly improved results for stoichiometric mixtures
Similar results to small Aachen mechanism for rich mixtures, but at least there is better confidence here the reasons are right

41. Acknowledgements
Thanks to all contributors of experimental and modelling data
Special thanks to those who provided the new shock-tube data at quite short notice
Matt Oehlschlaeger of Rensselaer Polytechnic Institute
David Davidson and Ron Hanson of Stanford University
Thanks to Yuanjiang Pei (ANL), Prithwish Kundu (NCSU), and Chao Yu (UConn) for assistance with analysis of the chemical mechanisms
Evatt Hawkes acknowledges funding from the Australian Research Council, computing time from the National Computational Infrastructure (Australia), Intersect, and the UNSW Faculty of Engineering cluster.

42. Questions and discussion