Environ. Sci. Technol. 1998, 32, 2018-2024
Condensing Complex Atmospheric
Chemistry Mechanisms. 1. The
Direct Constrained Approximate
Lumping (DCAL) Method Applied to
Alkane Photochemistry
S . W . W AN G AN D P . G . G E O R G O P O U LO S *
En viron m en tal an d Occu pation al Health Scien ces In stitu te,
170 Frelin gh u ysen Road, Piscataw ay, New Jersey 08855
G . LI AN D H . R AB I T Z
Departm en t of Ch em istry, Prin ceton Un iversity,
Prin ceton , New Jersey 08544-1009
Atmospheric chemistry mechanisms are the most computationally intensive components of photochemical air quality
simulation models (PAQSM s). The development of a
photochemical mechanism, that accurately describes
atmospheric chemistry w hile being computationally efficient
for use in PAQSM s, is a difficult undertaking that has
traditionally been pursued through semiempirical (“ diagnostic” ) lumping approaches. The limitations of these
diagnostic approaches are often associated w ith inaccuracies
due to the fact that the lumped mechanisms have typically
been optimized to fit the concentration profile of a
specific species. Formal mathematical methods for model
reduction have the potential (demonstrated through past
applications in other areas) to provide very effective solutions
to the need for computational efficiency combined w ith
accuracy. Such methods, that can be used to “ condense”
a chemical mechanism, include “ kinetic lumping” and
“ domain separation” . An application of the kinetic lumping
method, using the direct constrained approximate lumping
(DCAL) approach, to the atmospheric photochemistry of
alkanes is presented in this w ork. It is show n that the lumped
mechanism generated through the application of the DCAL
method has the potential to overcome the limitations of
existing semiempirical approaches, especially in relation to
the consistent and accurate calculation of the timeconcentration profiles of multiple species.
gen eration s of PAQSMs in corp orate m ore detailed con sideration s of h eterogen eou s ch em istry cou p led with th e gasp h ase p rocesses. Th is com p u tation al bu rden is p artly du e
to th e fact th at atm osp h eric ch em ical kin etic system s are
very “stiff”, i.e., th ey in clu de reaction s ran gin g from very fast
to very slow an d so in volve ch an ges associated with a very
disp arate ran ge of tim e scales; th is requ ires th e u se of
elaborate n u m erical in tegration sch em es (“stiff in tegrators”)
(6, 7). It is tru e th at rap id advan ces in com p u ter tech n ology
an d in th e availability of cost-effective com p u tation al
p latform s (su ch as lin ked server “clu sters”) (8-10) as well as
im p rovem en ts in n u m erical m eth ods [e.g., th e sp arse m atrix
vectorized Gear m ethod (7)], con tin ue to reduce the sim ulation tim e for existin g ap p lication s. However, th e dem an ds
arisin g from th e con tin u ou sly evolvin g redefin ition of
m odelin g p roblem s (which, in the case of PAQSMs, are driven
by both scien tific an d regu latory requ irem en ts) easily ou tp ace th e above advan ces. In deed, in recen t years, th e
ap p lication of PAQSMs h as rap idly evolved from u rban scale
to region al scale, to n ested m u ltiscale (11). At th e sam e tim e,
con sideration s of u n certain ty an d risk in develop in g defen sible air qu ality m an agem en t strategies requ ire th e p erform an ce of m u ltip le sim u lation s in order to iden tify an d select
a robu st con trol ap p roach . Fu rth erm ore, tasks su ch as
ap p lyin g region al scale m odels to th e m axim u m resolu tion
of em ission in ven tories, u n dertakin g com p reh en sive sen sitivity an d u n certain ty assessm en ts for PAQSMs, etc.,
con stitu te p roblem s th at in p ractice exceed available com putation al resources. Thus, there is a stron g n eed for accurate
an d validated bu t com p u tation ally efficien t version s of
PAQSMs that can be used as acceptable approxim ation s (with
kn own levels of accu racy) of th e “stan dard” m odels. In deed,
efforts h ave focu sed recen tly on develop in g su ch “fast” an d
“en gin eerin g” version s of regu latory an d related PAQSMs.
Th ese efforts in clu de th e u se of th e “fitted” GRS m ech an ism
(12) an d of a p rop rietary fast n u m erical solu tion sch em e
(13).
Altern atively, “form al” m ath em atical m eth ods for m odel
redu ction h ave th e p oten tial (dem on strated th rou gh p ast
ap p lication s in a n u m ber of areas) to p rovide very effective
solu tion s to th e n eed for a com p u tation al ap p roxim ation of
a ch em ical m ech an ism th at com bin es efficien cy with accu racy. Th e “form al” m eth ods can in fact be u sed to
overcom e som e of th e lim itation s associated with sem iem p irical (“diagn ostic”) ap p roaches. For exam ple, in diagn ostic
lum pin g, in accuracies often arise from optim izin g the lum ped
m ech an ism to fit th e con cen tration p rofile of a sp ecific
sp ecies. So, if a m ech an ism h as been op tim ized sp ecifically
to p redict ozon e con cen tration s, its ability to accu rately
p redict p rofiles of oth er sp ecies m igh t h ave been com p rom ised.
Introduction
Alth ough th e ch em istry m echan ism s in corp orated in existin g
com p reh en sive p hotochem ical air quality sim ulation m odels
[e.g., CBM-IV(1), th e SAPRC93 m ech an ism s (2), an d RADMCM (3), th at is th e ch em ical m ech an ism u sed in th e region al
acid dep osition m odel (4) an d SAQM (5)] are already
sim plified represen tation s of the actual atm ospheric chem ical
p rocesses, th ey are still th e m ost com p u tation ally in ten sive
com p on en ts of th ese m odels. In deed, ch em ical kin etics
calcu lation s typ ically con su m e over 90% of th e CPU tim e in
sim u lation s th at em p loy th e ch em istry-tran sp ort m odu les
of com p reh en sive p h otoch em ical m odelin g system s. Th is
situ ation is on ly goin g to becom e m ore severe as n ew
* To wh om all corresp on den ce sh ou ld be addressed.
2018
9
ENVIRONM ENTAL SCIENCE & TECHNOLOGY / VOL. 32, NO. 13, 1998
Background
Gas-phase reaction m echan ism s for the atm ospheric organ ic/
NO x / O 3 system h ave reach ed an advan ced state of develop m en t. Available com p reh en sive ch em ical m ech an ism s can
be classified as Explicit (or Detailed or “Referen ce”) an d as
Redu ced (or Lu m ped). Exp licit m ech an ism s aim to accou n t
for th e detailed ch em istry of all sp ecies an d in term ediates
in volved in th e atm osp h eric p ollu tion system u n der con sideration . Typ ically, th ey con sist of h u n dreds of reaction
step s an d are too lon g to be in corp orated in atm osp h eric
m odels in ten ded for rou tin e u se. For th is reason , redu ced
or lum p ed m echan ism s, gen erally in volvin g fewer than about
a h u n dred reaction s, h ave been develop ed as system atic
ap p roxim ation s of th e detailed ch em istry th at is described
S0013-936X(97)00967-X CCC: $15.00
1998 Am erican Chem ical Society
Published on Web 05/27/1998
FIGURE 1. Semiempirical (diagnostic) and mathematical (formal) approaches for “condensing” chemical kinetic systems.
by th e exp licit m ech an ism s. Com m on ly, th e in organ ic
ch em istry is retain ed in exp licit form , wh ereas th e ch em istry
of organ ics is sim p lified by “lu m p in g” togeth er a n u m ber of
reaction s an d/ or chem ical sp ecies. This m echan ism reduction h as been p u rsu ed in th e air qu ality m odelin g field
th rough sem iem p irical (“diagn ostic”) ap p roaches that defin e
a n u m ber of “sp ecies su bstitu tion s” for th e fu ll set of organ ics
of con cern . In th ese ap p roach es, p h en om en ological rate
con stan ts for th e reduced set are determ in ed through a fittin g
p rocess th at adju sts calcu lation s p erform ed with th e lu m p ed
m ech an ism so as to m atch observation s from sm og-cham ber
exp erim en ts. Two m ajor ap p roach es for p erform in g diagn ostic lu m p in g can be iden tified:
(1) Lum ped Molecule. The organ ics are grouped together
in classes accordin g to th eir ch em ical ch aracter (i.e., as
alkan es, olefin s, arom atics, aldeh ydes, etc.). Th en , eith er a
gen eralized (“h yp oth etical”) sp ecies or a su rrogate (“actu al”)
sp ecies is u sed to rep resen t th e ch em istry of each lu m p ed
class. Exam p les of m ech an ism s derived th rou gh th is ap p roach are th e RADM (3) an d SAPRC (14) an d SAPRC93 (2)
sch em es. (2) Lum ped Structure. Th e organ ics are grou p ed
n ot accordin g to classical typ es of com p ou n ds, as in th e
lu m p ed m olecu le m eth od, bu t accordin g to stru ctu re an d
reactivity ch aracteristics. For exam p le, carbon atom s m ay
be grou p ed, based on th eir bon din g, as sin gle bon ded, fast
dou bly bon ded, slow dou bly bon ded, an d carbon yl atom s.
Th e carbon bon d m ech an ism s (1) are th e p rim ary exam p le
of em p loyin g th is ap p roach of lu m p in g.
A gen eral altern ative to diagn ostic lu m p in g ap p roach es
is offered by form al m ath em atical m eth odologies of system
(or m odel) redu ction (see Figu re 1), th at aim to derive lower
dim en sion al form u lation s th at m atch th e resp on se of th e
origin al m odel (see, e.g., ref 15). In gen eral, su ch m eth ods
can be classified in to (i) aggregation or lu m p in g m eth ods,
wh ich are based on th e aggregation of state variables of th e
origin al system ; th ese m eth ods will be fu rth er discu ssed in
th e followin g section s; (ii) dom ain sep aration m eth ods (th e
m ost im p ortan t bein g th e sin gu lar p ertu rbation m eth od),
wh ich , for a kin etic system th at evolves in tim e, con sider
tem p oral dom ain s an d are based u p on sp littin g th e system
in to “fast” an d “slow” m odes (15, 16); variou s com p u tation al
im p lem en tation s of su ch m eth ods h ave been p resen ted in
th e literatu re (e.g.; 17-21); (iii) p artial realization m eth ods,
wh ich are based u p on m atch in g certain m om en ts of th e
origin al an d redu ced system s (22, 23); an d (iv) fittin g or
op tim u m ap p roxim ation or resp on se m eth ods, wh ich are
based upon m in im izin g the differen ce between the respon ses
of th e origin al system an d an ap p rop riately selected redu ced
rep resen tation , su ch as a sm aller system of differen tial
equ ation s or a m u ltivariate su rface p rovided by an algebraic
exp ression (24-26).
In p rin cip le, th e selection of a ch em ical m ech an ism
redu ction m eth od sh ou ld be based on th ree criteria: (a) th e
reduced m odel m ust satisfy m ass con servation , (b) the steadystate resp on se of th e redu ced m odel sh ou ld be iden tical to
(or very closely ap p roxim ate) th at of th e origin al m odel, an d
(c) it sh ou ld be p ossible to m ath em atically relate th e
p aram eters of th e redu ced m odel to th ose of th e origin al
m odel. Form al m eth ods th at actu ally satisfy all th ese criteria
are the m athem atical kin etic lum pin g an d dom ain separation
m eth ods. However, m eth ods from categories iii an d iv above
h ave th e p oten tial of p rovidin g, th ou gh at som e loss of
gen erality, th e m ost “h igh ly con den sed” rep resen tation s of
th e origin al m odel. Th is cou ld be of p articu lar im p ortan ce
in som e ap p lication s, su ch as u n certain ty p rop agation
calcu lation s for en viron m en tal im p act assessm en t an d risk
ch aracterization . In su ch ap p lication s, an d if th e lim itation s
of th e restricted rep resen tation are taken in to accou n t, th e
loss of accu racy in in dividu al sim u lation s m ay be su fficien tly
com p en sated by th e fact th at it wou ld be p ossible to p erform
h u n dreds or th ou san ds of sim u lation s in a Mon te Carlo or
sim ilar fram ework.
Th e detailed descrip tion of th e kin etic lu m p in g m eth od
an d its ap p lication to alkan e p h otoch em istry is given in th e
n ext section . Th e gen eral backgrou n d of th e m eth ods
com p risin g categories ii-iv above will be p resen ted, togeth er
with dem on stration s of th e ap p licability of th ese m eth ods
to atm osp h eric ch em istry p roblem s, in u p com in g articles.
VOL. 32, NO. 13, 1998 / ENVIRONM ENTAL SCIENCE & TECHNOLOGY
9
2019
Kinetic-Lumping Approach: The DCAL Method
Th e gen eral aggregation or lu m p in g m eth od for m odel
redu ction was form alized by Aoki (27) in th e late 1960s. (Th e
term “aggregation ” is used m ostly in the electrical en gin eerin g
literatu re, wh ereas lu m p in g is u sed m ore exten sively in
ch em ical an d com bu stion en gin eerin g ap p lication s; h ere,
we u se th e term lu m p in g sp ecifically for ch em ical reaction
system s: th is im p lies certain gen eral con strain ts, su ch as
m ass con servation .) It h as also been sh own (28) th at several
earlier m odel redu ction m eth ods (29-31) are sp ecial cases
of lu m p in g. In th e late 1960s, Wei an d Kuo (32, 33) develop ed
th e basic th eory for ch em ical kin etic lu m p in g of a relatively
sim p le n com p on en t, reversible first-order (lin ear) reaction
system ; Li an d Rabitz (34, 35) h ave sin ce exten ded th e th eory
to n on lin ear reaction system s wh ile oth er research h as
focused on the lum pin g of con tin uous reaction m ixtures (3639). Th e work p resen ted in th is p ap er focu ses on th e
ap p lication of th e lin ear kin etic-lu m p in g ap p roach . Th is
ap p roach is based on a form al m ath em atical an alysis of th e
en tire reaction system , to obtain lu m p ed sp ecies th at can
ap p roxim ate th e p rediction s for th e origin al or exp licit
sp ecies. Th is ap p roach has the p oten tial to reveal sim ilarities
in ch em ical sp ecies th at m ay n ot be ap p aren t in an an alysis
solely based on con sideration s of ch em ical attribu tes an d
can be u sed to ap p ly sim ilar weigh t factors on sim ilar sp ecies.
A brief descrip tion of th e th eoretical basis of lin ear kin etic
lu m p in g is given below.
Let th e kin etics of an n com p on en t reaction system be
described by
dy
) f(y)
dt
(1)
wh ere y is an n -dim en sion al com p osition vector, an d f(y) is
an n -dim en sion al vector fu n ction th at describes th e reaction
system . In lin ear kin etic lu m p in g, an n̂ -dim en sion al lu m p ed
com p osition vector ŷ corresp on din g to y is obtain ed by
ŷ ) My
(2)
wh ere M is a con stan t, n̂ × n dim en sion al lu m p in g m atrix.
Th e n -dim en sion al system is con sidered to be exactly
lu m p able if th ere exists a vector fu n ction f̂(ŷ) su ch th at
dŷ
) f̂(ŷ)
dt
(3)
Li an d Rabitz (34) p roved th at an exactly lu m p ed kin etic
system is
f̂(ŷ) ) Mf(M
h ŷ)
(4)
wh ere M
h is a gen eralized in verse of M. Th e m atrix M
h is
defin ed by
MM
h ) In̂
(5)
wh ere In̂ is th e n̂ -dim en sion al iden tity m atrix. Fu rth erm ore,
Li an d Rabitz (34) sh owed th at wh en th e lu m p in g is exact,
th e su bsp ace sp an n ed by th e rows of M m u st be JT(y)
in varian t, wh ere JT(y) is th e tran sp ose at Jacobian of f(y). A
m eth od of determ in in g M was described based on exp an din g
JT(y) accordin g to
m
JT(y) )
∑ a (y)A
k
k
(6)
k )1
wh ere con stan t m atrices Ak are viewed as a set of basis
m atrices of JT(y), an d fin din g th e n̂ -dim en sion al su bsp ace
th at is sim u ltan eou sly in varian t with resp ect to all Ak. It was
2020
9
ENVIRONM ENTAL SCIENCE & TECHNOLOGY / VOL. 32, NO. 13, 1998
also sh own th at alth ou gh M
h is n ot u n iqu e, a u n iqu e lu m p ed
system is obtain ed u sin g an y M
h , p rovided th e lu m p in g is
exact.
In p ractice, the lum p ed m odel is usually required to satisfy
som e restriction s. For exam p le, it m ay be requ ired to keep
certain ch em ical sp ecies of in terest u n lu m p ed. To accom m odate su ch restriction s, “con strain ed” lu m p in g m eth ods
h ave been develop ed (35, 40): th e lu m p in g is con strain ed by
a priori sp ecification of a p art of th e lu m p in g m atrix. In
con strain ed lu m p in g, M is rep resen ted as
M)
( )
MG
MD
(7)
wh ere MG is given , an d MD is to be determ in ed. Un der th ese
con strain ts, exact lum pin g schem es m ay n ot exist. Therefore,
m eth ods for th e determ in ation of con strain ed ap p roxim ate
lu m p in g sch em es are n ecessary.
In ap p roxim ate lu m p in g, th e rows of th e M m atrix are n ot
in varian t with resp ect to JT(y). Li an d Rabitz (35) develop ed
a m eth od of determ in in g a globally op tim al M for ap p roxim ate lu m p in g, su ch th at th e sp ace sp an n ed by th e rows
of M is “m ost n early” JT(y) in varian t for all y, an d sh owed th at
th e op tim al ch oice of M
h for an orth on orm al lu m p in g m atrix
M is sim p ly MT. More recen tly, Li an d Rabitz (40, 41) h ave
p rop osed a m ore straigh tforward m eth od for ap p roxim ate
lum pin g, the direct con strain ed approxim ate lum pin g (DCAL)
m eth od. Th e DCAL m eth od is based on determ in in g basis
m atrices Ak’s in eq 6 for JT(y) n u m erically, by u sin g valu es
of y in a region of n -dim en sion al com p osition sp ace th at is
of in terest. It was sh own th at M can be determ in ed from
()
XTMT ) 0
(8)
wh ere X is “m ost n early” orth ogon al to
MG
MGAT1
...
MG(AT1 )s 1-1
Z ) ...
MG
(9)
MGATm
...
MG(ATm )s m -1
wh ere sm is th e ran k of Ak or set to n an d can be determ in ed
by th e followin g p rocedu re (35, 42): tran sform MG(ATk )i to an
orth on orm al m atrix Q(G)Tk i u sin g Gram -Sch m idt orth ogosk-1
n alization , con stru ct a sym m etric m atrix Y ) ∑km)1 ∑i)0
QT
(G)k iQ(G)k i an d determ in e th e eigen valu es an d eigen vectors
of Y.
Th e eigen vectors corresp on din g to th e sm allest n - n̂
eigen valu es of Ycom p rise th e Xm atrix, an d th e eigen vectors
corresp on din g to th e largest n̂ eigen valu es of Ycom p rise th e
rows of M. To force MG to corresp on d to th e eigen vectors
of Y with th e largest eigen valu es, MG in Z is m u ltip lied by a
large con stan t (40). Th is en su res th at th e u n lu m p ed sp ecies
sp ecified by MG are p art of th e lu m p ed system .
On ce th e eigen valu es an d eigen vectors of th e Y m atrix
h ave been determ in ed, it is relatively straigh tforward to
determ in e th e M m atrix requ ired to derive a lu m p ed system
of arbitrary dim en sion n̂ < n . Wh en th e basis m atrices Ak
of JT(y) are u sed, th e resu ltan t M is globally op tim al. In
p ractice, it is often desired to ap p ly th e lu m p ed m odel to
on ly a lim ited region of th e com p osition sp ace. In th is case,
TABLE 1. Initial Conditions (ppm) for Chemical Species in the Explicit Model Mechanism
no.
name
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
O3
NO2
NO
NO3
N2O5
HNO3
HONO
RNO3
N2
O2
H2O
H2O2
CO
CO2
O3P
O1D
HO2
OH
RO2R
description
ozone
nitrogen dioxide
nitric oxide
nitrogen trioxide
dinitrogen pentoxide
nitric acid
nitrous acid
alkyl nitrate
nitrogen gas
oxygen gas
w ater
hydrogen peroxide
carbon m onoxide
carbon dioxide
oxygen atom (triplet)
oxygen atom (singlet)
hydroperoxy radical
hydroxyl radical
chem ical operator for
NO to NO2 conversions
w ith generation of HO2
conc
no.
name
7.08 × 10-2
1.09 × 10-3
0
0
0
0
0
0
7.81 × 10+5
2.09 × 10+5
1.04 × 10+4
0
0
0
0
0
0
0
0
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
NC4
ISC4
NC5
ISC5
NEC5
CLC5
NC6
2M C5
3M C5
22M B
23M B
CLC6
M CL5
NC7
3M C6
24M P
23M P
M CL6
NC8
butane
isobutane
pentane
isopentane
neo-pentane
cyclopentane
hexane
2-m ethylpentane
3-m ethylpentane
2,2-dim ethylbutane
2,3-dim ethylbutane
cyclohexane
m ethylcyclopentane
n -heptane
3-m ethylhexane
2,4-dim ethylpentane
2,3-dim ethylpentane
m ethylcyclohexane
n -octane
description
1.5 × 10-2
3 × 10-3
4.5 × 10-3
7.5 × 10-3
2.4 × 10-4
4.4 × 10-4
4.5 × 10-3
2.1 × 10-3
1.3 × 10-3
2.4 × 10-4
2.4 × 10-4
9.2 × 10-4
1.35 × 10-3
5.87 × 10-3
1.1 × 10-3
2.4 × 10-4
2.4 × 10-4
1.92 × 10-3
1.26 × 10-3
conc
4M C7
ISC8
ECL6
4-m ethylheptane
2,2,4-trim ethylpentane
ethylcyclohexane
4.87 × 10-4
2.4 × 10-4
2.1 × 10-4
20
RO2N
chem ical operator for
NO to Nitrate conversions
0
44
45
46
21
R2O2
chem ical operator for
NO to NO2 conversions
0
47
48
NC9
4EC7
n -nonane
4-ethylheptane
5.3 × 10-4
2 × 10-4
22
23
24
NC1
NC2
NC3
m ethane
ethane
propane
1.8
6 × 10-3
1.5 × 10-3
49
50
51
52
NC10
4PC7
NC11
NC12
n -decane
4-propylheptane
n -undecane
n -dodecane
3.92 × 10-4
2 × 10-4
1.06 × 10-3
5.1 × 10-4
Aks can be rep laced by JT(yk )s calcu lated at som e com p osition
p oin ts yk (k ) 1, 2, 3, ...) in th e con sidered regim e.
Application to Alkane Photochemistry
An ap p lication of th e DCAL m eth od was p erform ed startin g
from a “m odel” exp licit m ech an ism , wh ich con tain s 68
reaction s in volvin g 52 sp ecies. Th is “m odel” exp licit m ech an ism con siders detailed alkan e p h otoch em istry an d in organ ic ch em istry in th e trop osp h ere. Th e lu m p in g p rocess
is restricted to th e 30 n on m eth an e alkan e sp ecies sh own in
Table 1, wh ile th e 21 in organ ic com p ou n ds an d m eth an e
sh own in Table 1 are con strain ed to rem ain as in dividu al
sp ecies. Th e effects of tem p oral variation in tem p eratu re
an d ligh t in ten sity on th e rate con stan ts of th is reaction
m ech an ism were also con sidered. Dep en din g on th e n atu re
of tem p eratu re dep en den ce, rate con stan ts can be sp ecified
as differen t fun ction s of tem perature. In this “m odel” explicit
m echan ism , three differen t groups of fun ction s are em ployed
for the rate con stan ts. The first group of fun ction s is followin g
th e Arrh en iu s form u la to sp ecify tem p eratu re dep en den ce
for the rate con stan ts of in organ ic reaction s (except photolysis
reaction s). Th e rate con stan ts of p h otolysis reaction s are
in flu en ced by ligh t in ten sity an d are sp ecified by th e secon d
grou p of fu n ction s obtain ed via a sp lin e in terp olation
tech n iqu e, wh ich can gen erate th e fitted cu rves from valu es
of the photolysis rate con stan ts at differen t observation tim es.
Th e th ird grou p of fu n ction s is u sed to sp ecify tem p eratu re
effects on rate con stan ts for alkan e reaction s. Sin ce th ere
are n o Arrh en iu s kin etic data available for every sin gle
n on m eth an e alkan e sp ecies, we also u se th e sp lin e in terp olation tech n iqu e to obtain exp licit fu n ction s by in terp olatin g m easu rem en t data of rate con stan ts at differen t
tem p eratu res from literatu re. Th erefore, a total of 68
fu n ction s were develop ed alon g with th is “m odel” exp licit
m echan ism to accoun t for the tem perature an d light in ten sity
effects on th e 68 rate con stan ts in th is reaction m ech an ism .
Th ese 68 rate con stan ts are all fu n ction s of tem p eratu re
or tim e. Th u s, on ce th e tem p eratu re p rofile is sp ecified as
a fu n ction of tim e, th e dep en den t variable of th ese 68
fu n ction s is ju st tim e. Th erefore, th e rate con stan ts can be
u p dated at each tim e step du rin g th e in tegration p rocess for
solvin g th e differen tial equ ation s for ch em ical kin etics.
Methodology. Th e p erform an ce of th e DCAL-derived
m ech an ism s was evalu ated by sim u latin g a 24 h sm ogcham ber experim en t usin g these m echan ism s an d com parin g
th e resu lts with sim u lation s of th e sam e exp erim en t u sin g
the “m odel” explicit m echan ism , as well as the correspon din g
alkan e p ortion of CB4 m ech an ism an d SAPRC93 m ech an ism .
In itial con cen tration s sh own in Table 1 for th e sim u lated
sm og-ch am ber exp erim en ts were typ ical of p ollu ted u rban
air for th e sp ecies in clu ded in th e exp licit m ech an ism (43).
Sim u lation s with th e “m odel” exp licit, DCAL-derived, th e
alkan e p ortion s of CB4 an d SAPRC93 m ech an ism s were
perform ed usin g a box m odel (44) with the LSODE (45) routin e
to solve th e stiff ch em ical kin etic system . To ob tain th e
lu m p in g m atrix M for derivin g th e con den sed m ech an ism s,
th e sim u lation with th e “m odel” exp licit m ech an ism is
perform ed first to evaluate the Jacobian m atrices of the kin etic
equ ation s at differen t tim es. For th is “m odel” exp licit
m ech an ism , 52 differen tial equ ation s are gen erated to
describe th e kin etic beh avior of 52 exp licit ch em ical sp ecies.
Th e evalu ation s of 52 × 52 Jacobian m atrices are recorded
at 13 equ ally sp aced tim e p oin ts of th e sim u lation to serve
as th e basis m atrices for fin din g th e sim u ltan eou s “m ostly
in varian t” su bsp ace as th e lu m p in g m atrix. Sin ce 22 sp ecies
in clu din g 21 in organ ic sp ecies an d m eth an e are kep t
u n lu m p ed, MG is a 22 × 52 m atrix, th e first 22 colu m n s of
wh ich are com p rised of orth ogon al u n it vectors, an d th e rest
of th e colu m n s are n u ll vectors. With th e in form ation for
the 13 basis m atrices an d MG, on e can calculate the sym m etric
m atrix Y from eq 9, by settin g all sk ) n . Wh en th e
eigen vectors of th e sym m etric m atrix Y are arran ged acVOL. 32, NO. 13, 1998 / ENVIRONM ENTAL SCIENCE & TECHNOLOGY
9
2021
FIGURE 2. Comparisons of concentration-time profiles for (a) O3, (b) NO, (c) NO2, and (d) OH predicted by the “model” explicit mechanism,
the CB4 mechanism, the SAPRC93 mechanism and three DCAL-derived mechanisms for alkane photochemistry.
cordin g to th e decreasin g order of th e m agn itu des of th eir
eigen valu es, th e first n̂ eigen vectors are th e best con strain ed
lu m p in g m atrix MT with dim en sion n̂ . Th erefore, th e
eigen vector m atrix of Ysu p p lies all th e best lum p in g m atrices
with differen t n̂ . Th e eigen valu es can be u sed as a relative
m easu re of th e error of th e lu m p in g m atrices. If on e ch ooses
th e first n̂ eigen vectors to be th e lu m p in g m atrix, th e su m
of th e eigen valu es for oth er eigen vectors is th e relative
m easu re of its error. Th erefore, wh en th e dim en sion of th e
lu m p in g m atrix becom es h igh er, th e error becom es sm aller.
Wh en th e eigen valu es of oth er eigen vectors are all zero, th e
lu m p in g m atrix is exact. After determ in in g th e lu m p in g
m atrix M, th e n̂ dim en sion al lu m p ed system can be con structed by applyin g eq 4, which was presen ted in the previous
section . By settin g th e n̂ dim en sion al lu m p ed com p osition
vector ŷ as th e dep en den t set of variables, th e origin al n
dim en sion al com position vector y can be expressed as a lin ear
com bin ation s of n̂ com p on en ts of th e vector ŷ th rou gh th e
tran sform ation of M
h ŷ. Th u s, th e n dim en sion al vector
fu n ction f(M
h ŷ) h as on ly n̂ dep en den t variables. Th erefore,
th e n̂ dim en sion al vector fu n ction f̂(ŷ), wh ich describes th e
2022
9
ENVIRONM ENTAL SCIENCE & TECHNOLOGY / VOL. 32, NO. 13, 1998
ch em ical kin etics of th e lu m p ed system , can be obtain ed by
Mf(M
h ŷ). Th e m atrix m an ip u lation s requ ired to obtain th e
DCAL derived m ech an ism were p erform ed u sin g a com bin ation of FORTRAN code an d MAPLE (46). Th e evalu ation
of th e m eth od was con du cted by com p arin g th e ozon e, NO,
NO 2, an d OH p rofiles p redicted by this m echan ism with those
p redicted by th e “m odel” exp licit m ech an ism an d th e
corresp on din g CB4 an d SAPRC93 m ech an ism s.
Results and Discussion
Accordin g to th e m agn itu des of th e eigen valu es, th e first 25
colu m n s of th e eigen vector m atrix of Y can com p ose an
“alm ost exact” lu m p in g m atrix, sin ce th e first 25 eigen vectors
h ave eigen valu es larger th an 0.1 an d th e rest of th e 27
eigen vectors corresp on d to eigen valu es sm aller th an 1.0
×10-4. Th erefore, th e lu m p in g m atrix is obtain ed by takin g
th e tran sp ose of th e first 25 colu m n s of th e resu ltan t
eigen vector m atrix of Y. Sin ce 22 sp ecies h ave been kep t
u n lu m p ed, th is in dicates th at th ree DCAL m ech an ism s can
be form u lated with on e, two, or th ree lu m p ed sp ecies.
Th rou gh th e tran sform ation of th e lu m p in g m atrix, th e 30
n on m eth an e alkan e sp ecies can be lu m p ed in to on e, two,
or th ree lu m p ed sp ecies. Th e con cen tration -tim e p rofiles
for ozon e, NO, NO 2, an d OH obtain ed from th e sim u lated
sm og-cham ber experim en ts with the “m odel” explicit m echan ism , th ree DCAL-derived m ech an ism s, an d th e alkan e
p ortion s of th e CB4 an d SAPRC93 m ech an ism s are sh own in
Figu re 2. All th ree DCAL m ech an ism s sh ow sign ifican tly
better agreem en t with th e p rediction s of th e fu ll m odel th an
th e alkan e p ortion of th e CB4 m ech an ism . Th e DCAL
m ech an ism with th ree lu m p ed sp ecies ap p ears to p rodu ce
p rediction s alm ost iden tical to th at of th e fu ll m odel. For
reaction tim es exceedin g 600 m in , th e alkan e p ortion of
SAPRC93 m ech an ism sh ows better agreem en t with th e fu ll
m odel in p redictin g ozon e p rofile th an th e corresp on din g
CB4 m ech an ism as well as th e two DCAL m ech an ism s with
on e an d two lu m p ed sp ecies resu lts. However, wh en
p redictin g NO an d NO 2 p rofiles, th e SAPRC93 m ech an ism
exh ibits larger deviation s from th e fu ll m odel calcu lation s
th an th e CB4 m ech an ism an d th e th ree DCAL m ech an ism s.
It sh ou ld be n oticed th at two differen t lu m p in g m atrices
were gen erated by u sin g two differen t “stan dard” available
software p rogram s (e.g. IMSL rou tin e EIGRS an d NETLIB
rou tin e DSYEV) for fin din g eigen valu es an d eigen vectors of
th e sam e sym m etric m atrix Y. Th e two differen t lu m p ed
system s con stru cted from th ese two differen t lu m p in g
m atrices h ave th e sam e con cen tration -tim e p rofiles for all
th e sp ecies. Th is in dicates th at th e two su bsp aces sp an n ed
by th e rows of th ese two lu m p in g m atrices m igh t be
overlappin g. On e way to check this poin t here is to determ in e
th e degree of coin ciden ce between two su bsp aces. We u se
d c to rep resen t a qu an titative descrip tion of th e degree of
coin ciden ce between two su bsp aces. Accordin g to th is
geom etric con cep t, wh en on e of th e two subsp aces lies in side
th e oth er on e, th e basis vectors of on e su bsp ace are lin ear
com bin ation s of th ose in th e oth er su bsp ace. In th is case,
d c is u n ity. Wh en th e two su bsp aces are orth ogon al to each
oth er, d c is equ al to zero. In oth er cases, 0 < d c < 1. Let th e
n × r an d n × r′ m atrices P(r) an d P(r′) be th e m atrix
rep resen tation s of th e two su bsp aces with r′ e r. Th e degree
of coin ciden ce d c of th e two su bsp aces is defin ed as follows
to satisfy th e above requ irem en ts:
1
d c ) tr[P(r′)TP(r)P(r)TP(r′)]
r′
We calcu lated th e d c between th e two differen t lu m p in g
m atrices in th is stu dy, an d it is equ al to 0.999. Th u s, th e two
subspaces span n ed by these two lum pin g m atrices are alm ost
overlap p in g each oth er. Th erefore, th e two differen t lu m p ed
system s h ave th e sam e resp on ses as th e fu ll m odel.
Testing for Different Initial Conditions. Th e above
resu lts were obtain ed by sim u latin g on ly on e set of in itial
con dition s for ch em ical sp ecies con cen tration s. To test th e
robu stn ess of th e lu m p ed m odel, variou s sets of in itial
con dition s were also sim u lated for th e com p arison of
resp on ses between th e lu m p ed m odel an d th e fu ll m odel.
Th e in itial con dition s for fou r ch em ical sp ecies, N 2, O 2, H 2O,
an d CH 4, were kep t th e sam e, corresp on din g to valu es of
typ ical u rban air con cen tration s am on g variou s sets of in itial
con dition s, sin ce th e con cen tration s of th ese fou r sp ecies
are m u ch h igh er th an th ose of oth er sp ecies an d, th u s, cou ld
be treated as con stan ts. Th e in itial con cen tration s of th e
oth er sp ecies were taken from ran dom sam p les in th e ran ge
between p ossible m in im u m an d m axim u m con cen tration
valu es for each sp ecies. On e th ou san d sim u lation s were
p erform ed by takin g 1000 ran dom sam p les of in itial con dition s to stu dy th e agreem en t between th e lu m p ed m odel
an d th e fu ll m odel again st differen t in itial con dition s. For
each of th ese 1000 sim u lation s, th e ran ges of th e absolu te
errors an d relative errors for th e resp on ses (e.g., ozon e
con cen tration s) between th e lu m p ed m odel an d th e fu ll
m odel were also calcu lated. On th e basis of th e in form ation
of th e absolu te an d relative error ran ges from th ese sim u lation s, th e m axim u m an d m in im u m valu es of th e absolu te
an d relative errors were determ in ed as in dicators of th e
robu stn ess of th e lu m p ed m odel. It was fou n d th at th e ran ge
of th e absolu te error for p redicted ozon e con cen tration s
between two m odels ran ges from -0.6 to 0.1 p p b, an d th e
ran ge of th e relative error is from 0 to 0.6%. Th u s, it ap p ears
th at th e lu m p ed m odel con stru cted from th e set of in itial
con dition s typ ical of u rban air is fairly robu st with resp ect
to variation in th ese con dition s. Typ ically, th e m ost robu st
lu m p ed m odel sh ou ld be con stru cted based u p on con sideration of th e wh ole com p osition regim e of th e in itial sp ecies
con cen tration s. Th e reason for requ irin g th e lu m p ed m odel
to be so robu st is th at th e lu m p in g m atrix is con stru cted
based on a sim ulation scen ario where the com position regim e
is covered p rop erly by th e ch osen p oin ts, in wh ich th e
Jacobian m atrices are servin g as th e basis m atrices. It m ay
n ot be th e case th at a robu st m odel can be con stru cted based
u p on on ly on e set of in itial con dition s wh en m ore com p licated reaction system th at in clu de oth er organ ic grou p s su ch
as alken es, arom atics, an d carbon yls h ave to be con sidered.
Testing for Com putational Efficiency. Th e CPU tim es
for p erform in g 100 sim u lation s, corresp on din g to variou s
in itial con dition s, on a SUN ULTRA SPARC-1 170 MHz
workstation , for th e fu ll alkan e m ech an ism , for th e alkan e
p ortion s of CB4 an d SAPRC93 m ech an ism s, an d for th e th ree
DCAL m ech an ism s associated with on e, two, an d th ree
lum ped species, are 246.4, 125.6, 134.3, 117.5, 122.7, an d 130.5
s, resp ectively. So, th e CB4 an d SAPRC93 m ech an ism s an d
all th ree DCAL m ech an ism s ach ieve abou t 50% redu ction in
com p u tin g tim e com p ared with th e fu ll m ech an ism ; fu rth erm ore, th e DCAL m ech an ism with th ree lu m p s also
ach ieves excellen t accu racy. Th ese resu lts su ggest th at th e
DCAL m eth od can be u sed to derive th eoretically op tim al
lum ped representation m echanism s and to determ ine achievable error bou n ds for oth er lu m p ed m ech an ism s. In sigh ts
obtain ed by applyin g this m ethod can also be used to im prove
th e lu m p in g strategies em p loyed in existin g p h otoch em ical
m ech an ism s.
Acknow ledgments
Th is work was con du cted u n der th e au sp ices of th e Ozon e
Research Cen ter (ORC); base fu n din g for th e ORC is p rovided
by th e New Jersey Dep artm en t of En viron m en tal Protection
(NJDEP); addition al fu n din g was p rovided by U.S.EPA’s
Nation al Exp osu re Research Laboratory (NERL) u n der Coop erative Agreem en t CR 823717, an d by a Du p on t Edu cation al Gran t.
Supporting Information Available
Fou r tables [th e first 25 eigen valu es of m atrix Y(Table 2), list
of reaction s in th e exp licit m odel m ech an ism (Tables 3 an d
4), list of reaction s of th e alkan e p ortion of CB4 m ech an ism
(Table 5), an d list of reaction s of th e alkan e p ortion of
SAPRC93 m ech an ism (Table 6)] (4 p ages) will ap p ear
followin g th ese p ages in th e m icrofilm edition of th is volu m e
of th e jou rn al. Ph otocop ies of th e Su p p ortin g In form ation
from th is p ap er or m icrofich e (105 × 148 m m , 24 × reduction ,
n egatives) m ay be obtain ed from Microform s Office, Am erican Ch em ical Society, 1155 16th St. NW, Wash in gton , DC
20036. Fu ll bibliograp h ic citation (jou rn al, title of article,
n am es of au th ors, in clu sive p agin ation , volu m e n u m ber, an d
issu e n u m ber) an d p rep aym en t, ch eck or m on ey order for
$13.50 for p hotocop y ($15.50 foreign ) or $12.00 for m icrofiche
($13.00 foreign ), are requ ired. Can adian residen ts sh ou ld
add 7% GST. Su p p ortin g In form ation is also available via
th e World Wide Web at URL h ttp :/ / www.ch em cen ter.org.
VOL. 32, NO. 13, 1998 / ENVIRONM ENTAL SCIENCE & TECHNOLOGY
9
2023
Users sh ou ld select Electron ic Pu blication s an d th en En viron m en tal Scien ce an d Techn ology un der Electron ic Edition s.
Detailed in stru ction s for u sin g th is service, alon g with a
descrip tion of th e file form ats, are available at th is site. To
down load th e Su p p ortin g In form ation , en ter th e jou rn al
su bscrip tion n u m ber from you r m ailin g label. For addition al
in form ation on electron ic access, sen d electron ic m ail to
si-h elp @acs.org or p h on e (202)872-6333.
Literature Cited
(1) Gery, M. W.; Wh itten , G. Z.; Killu s, J. P.; Dodge, M. C. J. Geoph ys.
Res. 1989, 94, 12925.
(2) Carter, W. P. L. Atm os. En viron . 1995, 29, 2513.
(3) Stockwell, W. R.; Middleton , P.; Ch an g, J. S.; Tan g, X. J. Geoph ys.
Res. 1990, 95, 16343.
(4) Ch an g, J. S.; Brost, R. A.; Isaksen , I. S. A.; Madron ich , S.;
Middleton , P.; Stockwell, W. R.; Walcek, C. J. J. Geoph ys. Res.
1987, 92, 14681.
(5) Ch an g, J. S.; Li, Y.; Beau h arn ois, M.; Hu an g, H.; Lu , C.; Wojcik,
G. SAQM User’s Gu ide. Tech n ical rep ort, Atm osp h eric Scien ces
Research Cen ter, Un iversity at Alban y, State Un iversity of New
York, Alban y, NY, 1996.
(6) Aiken , R. C. Stiff Com pu tation . Oxford Un iversity Press, 1985.
(7) Jacobson , M. Z.; Tu rco, R. P. Atm os. En viron . 1994, 28A, 273.
(8) Sabot, G. High Perform an ce Com pu tin g: Problem Solvin g w ith
Parallel an d Vector Arch itectu res; Addison -Wesley Pu blish in g
Com p an y, In c., 1995.
(9) Bru egge, B.; Ried el, E.; Ru ssell, A.; Segall, E.; Steen kiste, P.
New sjou rn al of th e Society for In du strial an d Applied Math em atics, 1995; Vol. 28 (9).
(10) Pfister, G. F. Search of Clu sters; Pren tice Hall, 1995.
(11) Georgop ou los, P. G. En viron . Health Perspect. 1995, 103 (Su p p l.
2), 107.
(12) Ven katram , A.; Karam ch an dan i, P.; Pai, P.; Goldstein , R. Resu lts
an d Statu s of Modelin g, Proc. of th e AWMA In tl. Sp ecialty
Con feren ce on Region al Ph otoch em ical Measu rem en t an d
Modelin g Stu dies; San Diego, CA, 1995; Vol. II, p p 677-705.
(13) Gu th rie, P.; Yarwood, G.; Sh ep ard, S.; Ligocki, M.; Myers, T.;
Yocke, M. Resu lts an d Statu s of Modelin g, Proc. of th e AWMA
In tl. Sp ecialty Con feren ce on Region al Ph otoch em ical Measu rem en t an d Modelin g Stu dies; San Diego, CA, 1995; Vol. II,
p p 758-761.
(14) Carter, W. P. L. Developm en t an d Im pem en tation of an Upto-Date Ph otoch em ical Mech an ism for Use in Airsh ed Modelin g;
Tech n ical rep ort, Statewide Air Pollu tion Research Cen ter;
Un iversity of Californ ia: Riverside, CA, 1988.
(15) Sin h a, N. K.; Ku szta, B. Modelin g an d Iden tification of Dyn am ic
System s; Van Nostran d Rein h old Com p an y: New York, 1983.
(16) Brackbill, J. U.; Coh en , B. I. Mu ltiple Tim e Scales; Academ ic
Press, 1985.
(17) Howes, F. A. Arch . Rat. Mech . An al. 1986, 91 (2), 153.
2024
9
ENVIRONM ENTAL SCIENCE & TECHNOLOGY / VOL. 32, NO. 13, 1998
(18) Lam , S. H. Com bu st. Sci. Tech n ol. 1993, 89, 375.
(19) Lam , S. H.; Gou ssis, D. A. In Proc. of th e 22n d In tl. Sym posiu m
on Com bu stion ; C., Brezin ski, Ed.; Un iversity of Wash in gton :
Seattle, WA, 1988; p p 931-941.
(20) Lam , S. H.; Gou ssis, D. A. In t. J. Ch em . Kin et. 1994, 26, 461.
(21) Ru sso, M. F. Au tom atic Gen eration of Parallel Program s Usin g
Non lin ear Sin gu lar Pertu rbation Th eory; Ph .D. Th esis; Ru tgers,
Th e State Un iversity of New Jersey: New Bru n swick, NJ, 1989.
(22) Rossen , R. H.; Lap idu s, L. AICh E 1972, 18 (4), 673.
(23) Sein feld, J. H.; Lap idu s, L. Math em atical Meth ods in Ch em ical
En gin eerin g: Process Modelin g, Estim ation , an d Iden tification ;
Pren tice-Hall: En glewood Cliffs, NJ, 1973.
(24) Kh u ri, A. I.; Corn ell, J. A. Respon se Su rfaces. Marcel Dekker,
1987.
(25) Tu ryan i, T. Com pu t. Ch em . 1994, 18 (1), 45.
(26) Tu ryan i, T. In Tw en ty-Fifth Sym posiu m on Com bu stion ; 1994;
p p 949-955.
(27) Aoki, M. Au tom . Con trol 1969, AC-13, 246.
(28) Hicken , J. A Un ified Th eory of Model Redu ction for Lin ear Tim e
In varian t Dyn am ical System s; Ph .D. Th esis; McMaster Un iversity, 1978.
(29) Davison , E. J. IEEE Tran s. Au tom . Con trol 1966, AC-11, 93.
(30) Marsh all, S. A. Con trol En g. 1966, 10, 642.
(31) Mitra, D. On th e Redu ction of Com plexity of Lin ear Dyn am ic
Models; Tech n ical rep ort, U.K. Atom ic En ergy Au th ority, 1967;
Rep ort AEEW-R520.
(32) Wei, J.; Ku o, J. C. W. In d. En g. Ch em . Fu n dam . 1969, 8, 114.
(33) Ku o, J. C. W.; Wei, J. In d. En g. Ch em . 1969, 8 (1), 124.
(34) Li, G.; Rabitz, H. Ch em . En g. Sci. 1989, 44 (6), 1413.
(35) Li, G.; Rabitz, H. Ch em . En g. Sci. 1990, 45 (4), 977.
(36) Aris, R. AICh E, 1989, 35 (4), 539.
(37) Astarita, G.; Ocon e, R. AICh E 1988, 34 (8), 1299.
(38) Bailey, J. E. Ch em . En g. J. 1972, 3, 52.
(39) Ch ou , M. Y.; Ho, T. C. AICh E 1988, 34 (9), 1519.
(40) Li, G.; Rabitz, H. Ch em . En g. Sci. 1991, 46 (1), 95.
(41) Li, G.; Rabitz, H. Ch em . En g. Sci. 1991, 46 (2), 583.
(42) Bellm an , R. In trodu ction to Matrix An alysis; McGraw-Hill: New
York, 1970.
(43) Stockwell, W. R. Th e RADM ph otoch em ical box m odel; Tech n ical
rep ort, Woodward-Clyde Con su ltan ts, San ta An a, CA, 1992.
(44) System s Ap p lication s In tern ation al, Research Trian gle Park, NC.
User’s gu ide to version IV of th e Reactive Plu m e Model (RPM-IV),
Ap ril 1992. SYSAPP-92/ 037.
(45) Hin dm arsh , A. C. Liverm ore solver for ordin ary differen tial
equ ation s (lsode); GAMS, 1983.
(46) Map le, version v, release 3, 1994; Waterloo Map le Software,
Un iversity of Waterloo.
Received for review Novem ber 3, 1997. Revised m an u script
received March 24, 1998. Accepted March 31, 1998.
ES970967B