Page 1
Chemical Transport Modeling
Peter J. Adams Center for Atmospheric Particle Studies (CAPS)
Civil and Environmental Engineering
Engineering and Public Policy
Carnegie Mellon University
University of Pittsburgh
4 April 2011
Slide credits: Cliff Davidson, Neil Donahue, Yunha Lee, Spyros
Pandis, Jeff Pierce, Rob Pinder, Win Trivitayanurak, Knut von
Salzen
Page 2
Outline
• 1) Chemical transport models (CTMs) • Overview: formulation and solution
• 2) Processes:
• 3) Sample Applications
• 4) How models fail • How to be an intelligent user of models
• …and consumer of model output
Page 3
Chemical Transport Models: CTMs
• Overview
• Complex, 3D computer simulations of air quality
• Goal is to predict pollutant concentrations,
ci(x,y,z,t):
• For various species i: typically ~100
• At many locations x,y,z: typically ~104-105
• As a function of time t: day, week, year+
• Given
• Emissions
• Meteorological conditions
• Model description of relevant chemical & physical
processes
Page 4
Processes To Consider
• Emissions
• Atmospheric Transformations (“chemistry”)
• Chemical production / loss
• Physical transformations: phase changes, particle
growth, etc
• Transport
• Vertical dilution (stability, mixing height, etc)
• Horizontal transport
• Deposition
• Wet deposition
• Dry deposition
Page 5
East – West (x)
Ver
tica
l Hei
ght
(z)
Grid Set-up:
A + B C
Advection
.(u )iciRReaction
Emission, Ei Deposition, Di
CTMs: Overview
• Divide atmosphere into 3D grid of locations
• Each grid box has a mass balance (“continuity” equation)
• Use rate of change (dci/dt) to step forward in time
Page 6
Each Box: Atmospheric Processes
Emissions (E) Deposition (D)
Flow Out (Fout) Flow In (Fin) Production - Loss
(P – L)
Mass Balance
Equation:
Mi = Mass of a given chemical species in a given model box
(kg/s)
Page 7
How CTMs Work
• Inside each grid cell is a (familiar) mass balance for each species i:
• Solution procedure: • Start at t=0
• Calculate burdens at t= t (short time later)
• Repeat for many t’s until time period of interest is done
• Typical time steps, t, are 10 min to 1 hour
• For global model, t = 1 hour so mass balance equation solved:
(8760 box-1 yr-1)(30,000 boxes)(240 species) = 6 x 1010 yr-1
iioutiiniiii LPFFDE
dt
dM,,
tLPFFDEMM iioutiiniiioifi ,,,,
Page 8
Sample Output: Ozone
Page 9
Sample Output: Carbon Monoxide
Page 10
Emission Factors
• Emissions typically proportional to amount of “activity” that generates them:
E = (Emission Factor) x (Activity Level)
• All important details are buried in emission factor
• lots of work required to measure/estimate them
• Significant remaining challenges
• Volatility (and therefore evaporation) of primary organic particles
• Size distribution of primary particles, especially in ultrafine mode
• Widespread, highly variable “area” sources (e.g. ammonia, biogenic VOCs)
• Some emissions are “online” vs “offline”
• Offline: pre-calculated input data fed into model (e.g. vehicles)
• Online: emissions rate depends on meteorological conditions (e.g. isoprene, wind-driven sea spray and mineral dust)
• Emissions are nearly always a significant source of uncertainty in any CTM modeling study
Page 11
Emissions: Broad Categories
Natural Anthropogenic
Point Volcano (SO2) Power plant smokestack
(SO2)
Line VOCs from cars on
highway
Area VOCs from
vegetation
NH3 from livestock on
farm
more uncertain
less uncertain
Page 12
Meteorological Data
• Daily / seasonal variations in pollutant
concentrations largely a result of meteorological
variability
• Different CTMs (with same emissions,
chemistry, etc) can produce very different
predictions based on their meteorology
Page 13
Meteorological Data
• “Offline” met data (assimilated meteorology)
• CTM model is (pure) CTM-only
• Met fields are imported from a separate meteorological model (even “historical” meteorology is still a model simulation)
• Input data files are pretty large
• Precludes any chemistry-climate interactions (e.g. aerosol forcing affecting meteorology)
• No computer time spent on meteorological simulation (…often small compared to CTM calculations anyway)
• Often met fields have been evaluated more extensively
• “Online” met data (predicted meteorology)
• CTM model is part of a larger “host” model that includes meteorology
• Host model may be GCM (global climate model) or regional meteorology model (e.g. MM5, WRF)
• No need for meteorological input data files
• Allows chemistry-climate feedbacks (e.g. aerosol cloud lifetime effect)
• If your chemistry feeds back on climate, need to evaluate perturbed met fields as well as your chemistry simulation
Page 14
Meteorological Data
• Historical reanalysis data (most common offline met data) • Best attempt at met fields that match some specific time period in the past
• Useful for comparison against observations, especially short-term observations (e.g. field campaign data)
• Bear in mind that reanalysis data is part observations and part model
• Assimilation (aka “nudging”) • Process by which reanalysis data is generated
• Run climate or meteorological model for historical time period
• At each time step, available observations are imported to model by “nudging”
• Nudging = changing model values to be closer to observations
• Nudging is complex art, best to “outsource” this activity to meteorologists
• Not all reanalysis parameters are equally good • Geopotential fields (pressure distributions) are always nudged → resulting
large-scale wind patterns tend to be quite good
• Other variables (clouds, precipitation) are usually not nudged (problematic to do so)
• → still depend on meteorological model’s clouds and precipitation schemes (weak points in any met model)
• → reanalysis data often rains at the wrong time
• Pays to have meteorological expertise and well evaluated wind fields
Page 15
Transport Algorithms
• Advection algorithms are conceptually simple
but numerically complex
x (location)
x i-th grid cell
Ci Wind speed, u
Page 16
Advection Algorithms
x
x
i-th grid cell
Ci
x
x
Wind speed, u
Initial state
(t=0)
Later state,
exact (t= t)
Later state,
remapped
(t= t)
u t
Remapped to
conserve mass New Ci New Ci+1
Page 17
Advection Algorithms
x
Ci
x
x
Wind speed, u
Initial state
(t=0)
Later state,
exact (t=3 t)
Later state,
remapped
(t=3 t)
• Repeated remapping leads to “numerical diffusion”
• Consider 3 time steps at 3u t= x
Ci
Page 18
Piecewise Parabolic Algorithm
x
i
Initial state
(t=0)
i-1 i+1 For grid box i
1. Look at adjacent boxes
2. Infer probable concentration gradient within box i
3. Fit with parabola
4. Advect parabola
x
i
x
i i-1 i+1
Ci Ci+1
Ci-1
Page 19
Advection Algorithms
• Goal of advection algorithm is to minimize
numerical diffusion
• Results from
• Too much remapping because…
• Ci (species conc.) assumed constant within a grid cell
• Common schemes allow ci to vary within a grid
cell
• Piecewise parabolic method
• Lin and Rood, 1996
• Moment-based methods
• Prather, 1986
• Numerical method usually not significant source
of error/uncertainty
Page 20
Model For Deposition Velocity
A. Surface Layer
-turbulent eddy transport
B. Quasilaminar Sublayer
-Diffusion
-Inertial impaction
C. Interactions with the Surface
D. Gravitational Transport
-Settling velocity
Surface Layer
Quasi-laminar
Sublayer
Surface
Reference height
grcrbr(z)ar(z)tr
(z)d
v111
Page 21
Dry Deposition: General Observations
• Ultrafine mode
• Diffusion gives higher vd than accumulation mode
• But coagulation tends to dominate losses under most
conditions
• Accumulation mode • Minimal deposition velocity
(timescale is ~months)
• Wet deposition matters more
• Coarse mode • Dry deposition is dominant
removal (timescale ~1 day)
• Strongly size-dependent → useful to track several size categories to account for differing vd
Depositio
n V
elo
city
Particle Diameter
Acc. Ultra
-fine Coarse
Page 22
Wet Deposition
• Dominant removal for fine particulate matter
• Depends on meteorological conditions
• Predicting clouds and precipitation is a weak point in meteorological forecast and climate models • Different global models estimate
atmospheric lifetime of fine PM between ~4 and ~8 days
• Not uncommon for model to have rain at wrong time
Page 23
Chemistry
• Typical ozone photochemical mechanism has >50
species
• VOCs are lumped into categories based on reactivity
• Coupled differential equations, especially through
radical species (e.g. HOx)
• Equations are “stiff” (wide range of timescales from very
short-lived to long-lived species)
• Distinction between “species” and “tracers”
• Species: everything that participates in chemistry (e.g. OH, O3,
VOCs, CO, etc)
• Tracer: sufficiently long-lived to be transported from cell to cell
(OH excluded)
• Numerics typically done with “GEAR” solvers
Page 24
Aerosol Thermodynamic Modeling Approaches
• Bulk Equilibrium
CONS: Loses size information, timescales, mixing of particles
PROS: Fast, reliable; can be merged with size-resolved models given simple assumptions
• Size-resolved equilibrium (e.g. Pilinis and Seinfeld, 1987; Jacbson, 1999)
CONS: Timescales, infinite solutions, convergence problems
PROS: Average speed, size information
• Hybrid (Koo et al., 2003)
CONS: Some mixing of particles, more expensive than equilibrium
PROS: Relative fast, reliable, some size information
• Full dynamics (e.g. Pilinis et al., 2000)
CONS: Extremely stiff close to equilibrium, computationally intensive
PROS: Rigorous, size information
Page 25
Particle Size Distributions
• For particulate matter, size governs • Penetration into lungs
• Deposition from atmosphere
• Visibility
• Climate impacts (e.g. on clouds)
Particle
diameter (Dp)
Number of
particles
Page 26
Numerical Algorithms: Archetypes
Sectional (fixed or moving): n(Dpi)
Modal: (Ni, Dpgi, i) Moment-based: N, S, M
? (any distribution
consistent with number, surface area, mass)
• CTMs can treat particle size in different ways
Bulk: mass only
No representation of particle size; can be ok
for predicting PM2.5
Page 27
Tips: Learning a Model
• Codes are long, complex (50,000+ lines)
• Reading code not very useful
• Documentation better
• User manual (if it exists)
• Papers (info spread across many as model develops
over time)
• People (model how to as an “oral culture”)
• Balance “reading about” with “learning by
doing”
• Reproduce past results
• “play” with model by changing emissions,
parameters
Page 28
Modeling Tips: Budget Diagnostics
• One great benefit of CTMs: a world with perfect
information (no measurement limitations)
• In principle, should always be able to say why something
happened in the model
• In practice, storage constraints limit model predictions that are
output
• Minimum output (“diagnostics”) are c(x,y,z,t)
• Budget diagnostics are powerful, often overlooked
• Rates of sources and sinks (production and loss, Pi and Li)
• At steady state:
• Any change in model burden can be analyzed as
follows:
ii LPiL
M
iPM
ii PPM
(note it is changes in lifetime rather than changes in loss rates per se)
= lifetime
M = burden/mass
≈ source effect + sink effect
Page 29
Model Skill
• How good are models?
• Depends…
• Do we understand physics / chemistry?
• Do we know the relevant emissions?
• So model skill / reliability will depend a lot on
what species you are doing, where, etc
• Following slides are example (Karydis et al.
2007)
• PMCAMx CTM
• July 2001
• Various components of PM2.5
Page 30
Model Skill
Sulfate PM2.5
• Sulfate is best case scenario
• Chemistry (SO2→sulfate) is
understood
• Emissions are predominantly
from well monitored point
sources (power plants with
CEMS)
• Errors here largely result of
errors in model meteorology
• Bias: 5-10%; Error: 50-60%
Page 31
Model Skill
Nitrate PM2.5
• Nitrate is tough
• Chemistry (NOx→nitrate) is
understood
• But hypersensitive to other
errors
• Emissions are mix of point
sources (power plants) with
automobiles
• Measurements are problematic
• Bias: 50-60%; Error: ~90%
• (better in winter when nitrate is
more abundant)
Page 32
Model Skill
Organic PM2.5
• Organics also tough
• Chemistry is poorly understood
• Emissions are predominantly
area sources & wide variety
• Errors here largely result of
errors in model meteorology
• Bias: ~30%; Error: 50-60%
Page 33
Model Evaluation: CN10
•Captures variation in CN10 over two orders of
magnitude
•Average error (LMNE) is a factor of 2
Page 34
How Models Fail
• Poorly understood physics / chemistry
• Examples: new particle formation, secondary
organic PM
• A collective failure
• Science understood but system is complex
• Examples: clouds
• Mostly a problem of horizontal (or time) resolution of
model
• Bad input data
• Emissions (e.g. sea spray, mineral dust)
• Meteorology and deposition (e.g. everything)
Page 35
How Models Fail: Deposition
• COSAM: intercomparison of global sulfur models
• Same emissions, but models differ by factor of 2 mostly due to meteorology and deposition
Barrie et al., 2001
Page 36
Objectives of CTM Modeling
Most CTM modeling achieves one of 4 objectives:
1) Hypothesis testing
→does a newly discovered or hypothesized process make a difference in realistic atmospheric context?
2) Inferring unknown parameters (inverse modeling)
→what (uncertain) inputs are consistent with available observations (model outputs)?
3) Testing completeness of our knowledge
→can we predict observed behavior based on what we know?
4) Policy decision-making
→what will be effect of a specific emissions control?
Page 37
Primary vs Secondary Particles
Sensitivity of
CCN to secondary avg. = 11% increase
Sensitivity of
CCN to primary avg. = 27% increase
• Sensitivity of cloud-active particles (CCN) to uncertainties in primary (emitted) and secondary (nucleated) particles
• Indicates need to prioritize measurements of ultrafine particles from combustion sources
Ternary
Binary
PE
PE/3
Page 38
CTMs: Inverse Modeling
Arellano, A.F., P.S. Kasibhatla, L. Giglio, G.R. van der Werf, J.T. Randerson, and G.J. Collatz, Time-dependent
inversion estimates of global biomass-burning CO emissions using Measurement of Pollution in the Troposphere
(MOPITT) measurements, JGR, 111 (D9), 2006.
Chemical Transport
Model Find CO
emissions from
these regions
…consistent with
MOPITT CO
measurements
Careful: confounding or unknown factors
Page 39
CTMs: Testing Our Knowledge
Selin et al., JGR, 2007
Observations
Model Base Case
Model Sensitivity Cases (Blue/Green)
Arctic:
Obs: Large springtime depletion
Model: ~constant
→missing halogen chemistry?
NH Midlatitudes:
Obs: Modest late summer depletion
Model: reproduces well
→OH/O3 photooxidation working ok
Page 40
Organic PM: Testing Our Knowledge
“Urban” “Urban
Downwind”
“Rural
Remote”
•HOA ≈ fresh vehicle exhaust
•OOA: complex, but largely from atmospheric
oxidation of VOCs
•OOA fraction is high, even in urban areas
•Chemical production of organic material more
important than previously recognized
(Zhang et al.,
GRL 2007)
Page 41
Organic PM: Model Improvements
• Limitations of earlier models • Treated primary organic particles (e.g. vehicle
exhaust) as non-volatile
• Limited chemistry
• New processes • All organics PM is semi-volatile (condense &
evaporate between gas & particle phases)
• Secondary organic PM from intermediate VOCs
• New observations: volatility, oxygenation,
isotopic
Page 42
Organic PM: Testing Our Knowledge
• Updates to global OA
• semi-volatile POA
• IVOCs and their
chemistry…
• Predicted OA (left) is
not better than
traditional model
• Met fields / deposition
always going to give
factor of ~2 or so
• Probable tuning in
previous model
Page 43
Organic PM: Testing Our Knowledge
• Shantanu’s updates give obvious improvement in OOA/OA ratios (and volatility distribution)
• You need an indicator that is specific to / sensitive to your process of interest (usually not aerosol mass)
Page 44
Emissions Control Strategies: Ammonia and PM
Source: CMU Ammonia Inventory (2003)
Livestock, 54%
Fertilizer, 25%
Vehicles, 6%
Industrial, 2%
Domestics, 8%
Wild animals, 4%Wildfire, 1%
Misc., 1%
Nationally, agricultural
operations (green) are
the largest sources
Besides contribution to
PM2.5, ammonia is
problematic because:
•Odor
•Eutrophication
•Soil acidification
Page 45
Emissions Control Policy
• EPA has pursued SO2/NOx cuts from vehicles
and power plants
• Ammonia controls have been unattractive
because
• Uncertainties in emissions overall
• Area sources are more difficult to quantify / enforce
• PM2.5 only sensitive to ammonia under some
conditions (cold, free ammonia)
• Control technologies more diverse and untested
Page 46
SO4-NO3-NH3 (iPM2.5) Thermodynamics
SO42- (p) HNO3 (g)
NH3 (g) NH4+ (p) NH3 (free)
2 x SO42-
> Keqm(T,RH)?
NH4NO3 (p)
Sulfate-Nitrate “Substitution”
Sulfate aerosol ↓
Free ammonia ↑
Nitrate aerosol ↑
Net aerosol mass ~
Cold T
favors
nitrate
formation
Ammonia is an “enabler” for nitrate PM
formation
Page 47
NH3 and Fine Inorganic Particulate Matter (iPM2.5)
Source: West et al. 1999 Marginal PM2.5: Nonlinear
Aerosol Mass Response to Sulfate Reductions in the
Eastern United States JAWMA 49
Sulfate-
nitrate
substitution
regime
Page 48
National Inventory
Counties.shp0
< 4040 - 120120 - 250
250 - 500500 - 1500
kg NH3 km-2
Page 49
PMCAMx Chemical Transport Model
• 36 km x 36 km
• 97 cells (E-W) x 90 cells (N-S)
• CBM-IV gas phase chemistry
• CMU aerosol modules
• 14 layers (summer) 16 layers (winter)
Simulations
• January / July: two weeks each
• Every combination of
• SO2 / NOx / NH3 reductions
• 10%, 20%, 30%, 40%, 50%
• 125 scenarios total
Page 50
Effectiveness: iPM2.5 Sensitivity %
Reduction
in iP
M2.5
January July 50
40
10
30
20
50%
NH3
50%
NOx
50%
SO2
50%
NH3
50%
NOx
50%
SO2
Page 51
∆iPM2.5 for 10% NH3 Emission Reduction (Jan)
CTMs: Decision-Making Tools
Pinder, R.W., P.J. Adams, and S.N. Pandis, Ammonia emission controls as a cost-
effective strategy for reducing atmospheric particulate matter in the eastern
United States, Environmental Science & Technology, 41 (2), 380-386,
2007.
• Inorganic PM2.5 sensitive to ammonia
emissions in wintertime
• Relatively insensitive to SO2 controls
during winter (sulfate-nitrate
substitution)
• Ammonia emissions reductions
appear to be cost-effective strategy
for PM2.5
Page 52
Objectives of CTM Modeling
Most CTM modeling achieves one of 4 objectives:
1) Hypothesis testing
→does a newly discovered or hypothesized process make a difference in realistic atmospheric context?
2) Inferring unknown parameters (inverse modeling)
→what (uncertain) inputs are consistent with available observations (model outputs)?
3) Testing completeness of our knowledge
→can we predict observed behavior based on what we know?
4) Policy decision-making
→what will be effect of a specific emissions control?
“Meta-research”: real problems solved with experiments and measurements; models about setting priorities and testing completeness
Large uncertainties and
gaps in knowledge
Well understood system
Page 53
Summary
• Primary strength of CTMs: Our major attempt at being comprehensive • Most ambient and experimental work is necessarily focused on a given
location or process, mostly in isolation
• CTMs attempt to treat all processes (in theory) and most (in practice) in some detail
• Interactions/competitions between processes can be examined
• Integrate across different atmospheric conditions/regimes (in time and space)
• Building an aerosol model requires tradeoffs • Probably best to have a clear application: e.g. CCN or PM2.5
• Difficult to build a model that is all things to all people and all applications
• “Benchmark” models (comprehensive, accurate) versus “engineering” models (fast, used for production runs)
• Important to know model strengths/weaknesses • Inorganic chemistry and US SO2 emissions well understood → PM sulfate
well predicted
• SOA chemistry poorly understood (a problem not unique to modelers) → models more useful for “what if” rather than prediction