Professional Documents
Culture Documents
冕
共capturing only photons hitting it from within a small cone of
solid angles兲 monitoring the radiation from a large furnace filled s 共 r兲
⫹ I j 共 r,ŝ⬘ 兲 ⌽ 共 r,ŝ⬘ ,ŝ兲 d⍀ ⬘ , j⫽1,2,
with an absorbing, emitting and scattering medium. In a standard 4 4
Monte Carlo simulation one would emit many photon bundles
within the furnace, and would trace the path of each of these (1)
photons, even though only a very small fraction will hit the de- subject to the boundary condition
tector. It may take many billion bundles before a statistically
meaningful result is achieved—at the same time calculating the I j 共 rw ,ŝ兲 ⫽I w j 共 rw ,ŝ兲 , j⫽1,2, (2)
intensity field everywhere 共and without need兲: clearly a very where r is a vector pointing to a location within the medium, ŝ is
wasteful procedure. Obviously, it would be much more desirable a unit direction vector at that point, S is the local radiative source,
if one could just trace those photon bundles that eventually hit the  is the extinction coefficient, s the scattering coefficient, ⌽ is
detector. the scattering phase function, and ⍀ denotes solid angle. The prin-
This idea of a backward tracing solution, sometimes also ciple of reciprocity states that these two solutions are related by
known as reverse Monte Carlo has been applied by several inves- the following identity:
冕冕
tigators 关4 –10兴. All of these investigations have been somewhat
limited in scope, looking at light penetration through nonemitting 关 I w2 共 rw ,ŝ 兲 I 1 共 rw ,⫺ŝ兲
oceans and atmospheres 关4 – 6兴, computer graphics 关7,8兴, reflecting A n̂•ŝ⬎0
boundaries 关9兴, and emitting media 关10兴. All the aforementioned
papers have dealt with large light sources 共in volume and/or solid ⫺I w1 共 rw ,ŝ兲 I 2 共 rw ,⫺ŝ兲兴共 n̂•ŝ兲 ␦ ⍀dA
angle range兲, making a backward simulation straightforward. A
number of other methods to overcome the inefficiency of standard
Monte Carlo implementations in problems with small sources
⫽ 冕冕V 4
关 I 2 共 r,⫺ŝ兲 S 1 共 r,ŝ兲 ⫺I 1 共 r,ŝ兲 S 2 共 r,⫺ŝ兲兴 d⍀dV,
Journal of Heat Transfer Copyright © 2003 by ASME FEBRUARY 2003, Vol. 125 Õ 57
⫹ 冕 0
l
S 1 共 r⬘ ,⫺ŝ⬘ 共 r⬘ 兲兲 exp ⫺ 冋冕 册 l⬘
0
共 r⬙ 兲 dl ⬙ dl ⬘ ,
(9)
再
function. And, if the internal source of radiation is due to isotropic
0, r⫽ri , emission, then, comparing the standard RTE 关1兴 with Eq. 共1兲 we
␦ 共 r⫺ri 兲 ⫽ , (5a) find S 1 (r⬘ ,⫺ŝ⬘ )⫽ (r⬘ )I b (r⬘ ). Thus,
⬁, r⫽ri
冕V
␦ 共 r⫺ri 兲 dV⫽1, (5b)
I n 共 ri ,⫺ŝi 兲 ⫽ ⑀ 共 rw 兲 I b 共 rw 兲 exp ⫺ 冋冕 册 0
l
共 r⬘ 兲 dl ⬘
冕 冋冕 册
and similarly for solid angle. If the infinitesimal cross-section of l l⬘
the source, normal to ŝi , is dA i , then this results in an I 2 inten- ⫹ 共 r⬘ 兲 I b 共 r⬘ 兲 exp ⫺ 共 r⬙ 兲 dl ⬙ dl ⬘ ,
sity at ri of 0 0
␦ 共 ŝ⫺ŝi 兲 (10)
I 2 共 ri ,ŝ兲 ⫽ . (6)
dA i
As the I 2 light beam travels through the absorbing and/or scat- where the subscript ‘‘l’’ has been dropped since it is no longer
tering medium, it will be attenuated accordingly. needed. Equation 共10兲 may be solved via a standard Monte Carlo
Substituting Eqs. 共4兲 into Eq. 共3兲 yields the desired intensity as simulation or using the energy partitioning scheme described by
Modest 关1兴 and Walters and Buckius 关3兴. For the standard method
I 1 共 ri ,⫺ŝi 兲 ⫽ 冕冕
A n̂•ŝ⬎0
I w1 共 rw ,ŝ兲 I 2 共 rw ,⫺ŝ兲共 n̂•ŝ兲 d⍀dA
scattering lengths l are chosen as well as an absorption length
l . The bundle is then traced backward from ri unattenuated 关i.e.,
the exponential decay terms in Eq. 共10兲 are dropped兴, until the
⫹冕冕 V 4
S 1 共 r,ŝ兲 I 2 共 r,⫺ŝ兲 d⍀dV. (7)
total path length equals l or until the emission location rw is
reached 共whichever comes first兲. Thus,
冕
While the I 2 problem is much simpler to solve than the I 1
冦
l
problem, it remains quite difficult if the medium scatters radiation, 共 r⬘ 兲 I b 共 r⬘ 兲 dl ⬘ , l ⬍l,
making a Monte Carlo solution desirable. Therefore, we will ap- 0
冕
proximate I 1 as the statistical average over N distinct paths that a I n 共 ri ,⫺ŝi 兲 ⫽ l
.
photon bundle emitted at ri into direction ŝi traverses, as sche- ⑀ 共 rw 兲 I b 共 rw 兲 ⫹ 共 r⬘ 兲 I b 共 r⬘ 兲 dl ⬘ , l ⭓l
matically shown in Fig. 1, or 0
N (11)
1
I 1 共 ri ,⫺ŝi 兲 ⫽ 兺
I 共 r ⫺ŝ 兲 ,
N n⫽1 1n i i
(8)
If energy partitioning is used only scattering lengths are chosen
and I n is found directly from Eq. 共10兲.
where the solution for each I 1n is found for its distinct statistical
path 共with absorption and scattering occurrences chosen exactly as Radiative Fluxes. If radiative flux onto a surface at location
in the forward Monte Carlo method兲. Along such a zig-zag path of ri over a finite range of solid angles is desired, the flux needs to be
total length l from ri to rw , consisting of several straight segments computed using the statistical data obtained for I n (ri, ⫺ŝi ). For
pointing along a local direction ŝ⬘ (r⬘ ), I 2 is nonzero only over example, for a detector located at ri with opening angle max one
an infinitesimal volume along the path, dV⫽dA i l, and an infini- obtains
0
max
0
⑀ ⬘ 共 , 兲 I in共 , 兲 cos sin d d
⫽
1
2冕 冕 0
2 1
cos2 max
⑀ ⬘ 共 , 兲 I ,in共 , 兲 d 共 cos2 兲 d
⯝ 共 1⫺cos2 max兲 兺 ⑀ ⬘ 共 ŝ 兲 I
n⫽1
n n 共 ⫺ŝn 兲 , (12)
冕 1
cos2 n
d
1⫺cos2 n sin2 n
tions兲, and a multiply-scattered and reflected part. Again, we let I d
satisfy the radiative transfer equation without the inscattering
term, or,
冕
R ⫽ ⫽ ⫽ 2 ,
1 1⫺cos max sin max
2
ŝ•ⵜI d 共 r,ŝ兲 ⫽S d 共 r,ŝ兲 ⫺  共 r兲 I d 共 r,ŝ兲 , (17)
d
cos2 max
which has the simple solution
or n ⫽sin⫺1 共 冑R sin max兲 ,
where R and R are random numbers picked uniformly from 0
⭐R⭐1. If the detector is of finite dimension, points distributed
(13)
I d 共 r,ŝ兲 ⫽ 冕 冋冕
S d 共 r⬘ ,ŝ兲 exp ⫺
r→r⬘
册
共 ⫹ s 兲 ds ⬘ ds, (18)
across the surface are chosen like in a forward Monte Carlo simu- where the main integral is along a straight path from the boundary
lation. of the medium to point r in the direction of ŝ. For example, if
there is only a simple point source at r0 with total strength Q 0 ,
Collimated Irradiation. Backward Monte Carlo is extremely emitting isotropically across a tiny volume ␦ V, Eq. 共18兲 becomes
冋冕 册
efficient if radiative fluxes onto a small surface and/or over a
small solid angle range are needed. Conversely, forward Monte Q0
I d 共 r,ŝ兲 ⫽ exp ⫺ 共 ⫹ s 兲 ds ⬘ ␦ 共 ŝ⫺ŝ0 兲 ,
Carlo is most efficient if the radiation source is confined to a small 4 兩 r0 ⫺r兩 2 r0 →r
volume and/or solid angle range. Both methods become extremely (19)
inefficient, or fail, if radiation from a small source intercepted by
a small detector is needed. For collimated irradiation 共and similar where ŝ is a unit vector pointing from r0 toward r, and use has
problems兲 backward Monte Carlo can be made efficient by sepa- been made of the fact that
rating intensity into a direct 共collimated兲 and a scattered part, as ␦ ⍀ 0␦ s
outlined in Chapter 16 of 关1兴. Thus, letting I(r,ŝ)⫽I d (r,ŝ) ␦ V⫽ ␦ A ␦ s⫽ , (20)
兩 r0 ⫺r兩 2
⫹I s (r,ŝ), results in a direct component, attenuated by absorption
and scattering, where ␦ ⍀ 0 is the solid angle, with which ␦ V is seen from r.
冋冕 册
Equation 共19兲 can be used to calculate the direct contribution of
I d 共 r,ŝ兲 ⫽q coll共 rw 兲 ␦ 共 ŝ⫺ŝ0 兲 exp ⫺ 共 ⫹ s 兲 ds ⬘ , (14) Q 0 hitting a detector, and it can be used to determine the source
r→r⬘ term for the RTE of the scattered radiation as
which satisfies the RTE without the inscattering term. This leads
to a source term in the RTE for the scattered part of the intensity,
S 1 共 r,ŝ兲 ⫽
s 共 r兲
4 冕
4
I d 共 r,ŝ⬘ 兲 ⌽ 共 r,ŝ⬘ ,ŝ兲 d⍀ ⬘
due to 共first兲 scattering of the collimated beam, of
S 1 共 r,ŝ兲 ⫽ s 共 r兲
q coll共 rw 兲
4
exp ⫺ 冋冕 lc
0
册
共 ⫹ s 兲 dl c⬘ ⌽ 共 r,ŝ0 ,ŝ兲 ,
⫽
s 共 r兲 Q 0
16 2 兩 r0 ⫺r兩 2
exp ⫺ 冋冕 r0 →r
册
共 ⫹ s 兲 ds ⬘ ⌽ 共 r,ŝ0 ,ŝ兲 .
(15) (21)
where q coll is the collimated flux entering the medium at rw , The rest of the solution proceeds as before, with I n (ri ,⫺ŝi ) found
traveling a distance of l c toward r in the direction of ŝ0 , and the from Eq. 共16兲.
scattering phase function ⌽(r,ŝ0 ,ŝ) indicates the amount of col-
limated flux arriving at r from ŝ0 , being scattered into the direc- Sample Calculations
tion of ŝ. Therefore, the diffuse component of the intensity at ri is Isotropically Scattering, Nonabsorbing Medium With Colli-
found immediately from Eq. 共9兲 as mated Irradiation. As a first example we will consider a one-
I n 共 ri ,⫺ŝi 兲 ⫽ 冕 l
0
S 1 共 r⬘ ,⫺ŝ⬘ 兲 exp ⫺ 冋冕 l⬘
0
册
dl ⬙ dl ⬘ , (16)
dimensional slab 0⭐z⭐L⫽1 m of a gray, purely isotropically
scattering medium ( s ⫽1 m⫺1 ⫽const; ⌽⫽1), bounded at the
top (z⫽0) by vacuum and at the bottom (z⫽L) by a cold, black
with S 1 from Eq. 共15兲. As before, Eq. 共16兲 may be solved using surface. Collimated irradiation of strength Q⫽100 W is normally
standard tracing 关picking absorption length l , and dropping the incident on this nonreflecting layer, equally distributed over the
exponential attenuation term in Eq. 共16兲兴 or energy partitioning disk 0⭐r⭐R⫽0.1 m, as shown in Fig. 2. A small detector 2 cm
关using Eq. 共16兲 as given兴. ⫻2 cm in size, with an acceptance angle of max is located on the
black surface at x⫽x 0 ⫽0.2 m, y⫽0. The object is to determine
Point and Line Source. Backward Monte Carlo also be- the flux incident on the detector for varying acceptance angles,
comes inefficient if the radiation source comes from a very small comparing forward and backward Monte Carlo implementations.
surface or volume and/or if the source is unidirectional. The trick In all simulations the number of photon bundles was doubled
is again to break up intensity into a direct component 共intensity again and again until a relative variance of less than 2 percent of
coming directly from the source without scattering or wall reflec- the desired quantity was achieved.
z1 j
e ⫺共 ⫹s 兲z
dz
szj
acceptance angles, rising to 512⫻106 for max⫽10 deg. Results Q e ⫺z1 j
⫺e ⫺  z 2 j
are difficult to obtain for max⬍10 deg. Similar remarks can be
made for detector area: as the detector area decreases, the neces-
⫽
4 2R 2 兺 j szj
, (25)
sary number of bundles increases. Modeling a more typical detec- where and  are scattering albedo and extinction coefficient,
tor 1 mm⫻1 mm in size would almost be impossible. respectively. And again, an absorption length l is chosen, and the
In a Backward Monte Carlo simulation, since no direct radia- addition in Eq. 共25兲 is stopped as soon as the total path reaches l
tion hits the detector (x 0 ⬎R), the scattered irradiation is calcu- or the bundle leaves the layer 共which ever comes first兲.
lated from Eqs. 共16兲 and 共15兲 with q coll⫽Q/ R 2 as
冕
Backward Monte Carlo—Energy Partitioning. Again, the
l sQ ⫺ z scattering source must be attenuated as in Eq. 共25兲, but the expo-
I n 共 ri ,⫺ŝi 兲 ⫽ s H 共 R⫺r 共 l ⬘ 兲兲 dl ⬘ ,
2 2e (23)
nential attenuation term in Eq. 共16兲 must also be retained. Thus,
0 4 R
where l consists of a number of straight-line segments, for which
dl ⬘ ⫽dz ⬘ /cos , and H is Heaviside’s unit step function. There-
I n 共 ri ,⫺ŝi 兲 ⫽
sQ
4 2R 2 冕
0
l
e ⫺  z 共 l ⬘ 兲 ⫺ l ⬘ H 共 R⫺r 共 l ⬘ 兲兲 dl ⬘ , (26)
fore,
冕
where the integrand contributes only where the source is active
sQ z2 j dz
兺 ⫺sz (r⭐R), but attenuation of the bundle takes place everywhere
I n 共 ri ,⫺ŝi 兲 ⫽ e
4 2R 2 j z1 j szj (l ⬘ ⫽total distance along path from ri to r⬘ ). The rest of the simu-
lation remains as in the previous case. Results are summarized in
Q e ⫺ s z 1 j ⫺e ⫺ s z 2 j Table 1. As expected, if standard ray tracing is employed, the
⫽
4 2R 2 兺j szj
, (24) number of required bundles grows exponentially if the absorption
coefficient becomes large, both for forward and backward Monte
where s z j ⫽cos j is the z-component of the direction vector for Carlo. While backward Monte Carlo retains its advantage 共indeed,
the j th segment, and z 1 j and z 2 j are the z-locations between which the forward Monte Carlo simulation for ⫽5 m⫺1 could only be
carried out to a variance of 5 percent兲, the relative growth of where the rm are M random locations chosen uniformly along path
required bundles appears to be worse for backward Monte Carlo. l , j . Results for detector flux as function of scattering coefficient
If energy partitioning is employed, the number of bundles remains are shown in Fig. 4.
unaffected by the absorption coefficient for both, forward and For small values of s the number of photon bundles required
backward Monte Carlo. to achieve a relative variance of 2 percent is much smaller for the
backward Monte Carlo method, as expected, since the volume
Isotropically Scattering, Nonabsorbing Medium With Inter- with secondary scattering 共i.e., the Source S 1 ) is relatively large,
nal Point Source. In a final example a point source of strength and the detector is small. However, as s increases, the size of the
Q 0 ⫽100 W, located at x 0 ⫽y 0 ⫽0, z 0 ⫽0.5L will be considered secondary scattering volume decreases, and backward Monte
for a purely scattering medium. Again, flux hitting the detector Carlo becomes less and less efficient. For both methods large s
will be compared using forward and backward Monte Carlo meth- mean smaller l , j , leading to increased tracing effort for each
ods. individual bundle. Numerical integration via Eq. 共28兲 was gener-
The forward Monte Carlo simulation is almost identical to that ally much more efficient than Newton-Cotes quadrature, with M
of the first example, except that all photon bundles are now emit- ⫽1 usually being sufficient 共since the integral is evaluated so
ted from a single point, but into random directions. In the back- many times兲. However, for large s this method became ineffi-
ward Monte Carlo simulation, the detector flux again consists of a cient, requiring many photon bundles to achieve a 2 percent rela-
direct and a scattered component and, again, the direct component tive variance. In addition, all methods became inefficient for s
is zero, this time because all direct radiation hits the detector at an ⬎10 m⫺1 .
angle larger than the acceptance angle. The I n are then found from
Eqs. 共21兲 and 共16兲 as Summary
I n 共 ri ,⫺ŝi 兲 ⫽
sQ
16 2 兺 j
冕 e ⫺ s 兩 r0 ⫺r兩
l , j 兩 r0 ⫺r兩
2 dl ⬘ , (27)
A comprehensive formulation for backward Monte Carlo simu-
lations, capable of treating emitting, absorbing and anisotropically
scattering media, media with diffuse or collimated irradiation
where the l , j are the straight paths the bundle travels between 共with large or small footprints兲, media with point or line sources,
scattering events. Equation 共27兲 must be integrated numerically, etc., has been given. The basic backward Monte Carlo simulation
and this can be done using a simple Newton-Cotes scheme; here of Walters and Buckius 关3兴 was reviewed and was extended to
no optimization of the quadrature was attempted, except that— allow for collimated irradiation, point sources, and other sources
away from the source—the number of integration points was of small volume/area and/or small solid angle range. In addition,
minimized for small l 共large s ). Alternatively, the integral can the method was extended to allow standard ray tracing 共bundles of
be obtained statistically from fixed energy兲 as well as energy partitioning 共bundles attenuated by
absorption兲. Sample results for radiation hitting a small detector
sQ l , j e ⫺ s 兩 r0 ⫺rm 兩
兺 兺
show that forward Monte Carlo methods degrade rapidly with
I n 共 ri ,⫺ŝi 兲 ⫽ , (28) shrinking detector size and acceptance angle. Backward Monte
16 2 j M m 兩 r0 ⫺rm 兩 2
Carlo, on the other hand, is unaffected by detector size, but re-
quires a relatively large radiation source, which—in the case of
collimated irradiation or point sources—needs to be created arti-
ficially by separating direct and scattered radiation. Even for rela-
tively large detectors/opening angles, using backward Monte
Carlo can result in several orders of magnitude lesser computer
effort, and becomes the only feasible method for very small de-
tectors. Similarly, using energy partitioning in strongly absorbing
media also reduces numerical effort by orders of magnitude for,
both, forward and backward Monte Carlo simulations.
Nomenclature
A ⫽ area, m2
I ⫽ radiative intensity, W/m2 sr
l ⫽ geometric length, m
n̂ ⫽ unit surface normal
N ⫽ number of photon bundles
q ⫽ radiative heat flux, W/m2
r ⫽ position vector
Fig. 4 Detector fluxes and required number of photon bundles R ⫽ random number
„to achieve relative variance less than 2 percent… for one- ŝ ⫽ unit direction vector
dimensional slab with internal point source S ⫽ radiative source, W/m3 sr