Development and applications of benchmarking aerosol models … · 2020. 1. 6. · regional scale using a stochastic particle-resolved approach Jeffrey H. Curtis, Nicole Riemer and

Development and applications ofbenchmarking aerosol models on the

regional scale using a stochasticparticle-resolved approach

Jeffrey H. Curtis, Nicole Riemer and Matthew West

University of Illinois at Urbana-Champaign

International Aerosol Modeling Algorithms ConferenceDecember 5, 2019

Curtis (University of Illinois) Benchmarking aerosol models on the regional scale December 5, 2019 1/15

Atmospheric modeling: A multiscale challenge

0.0·107

0.5·107

1.0·107

1.5·107

2.0·107

2.5·107

3.0·107

3.5·107

4.0·107

4.5·107

5.0·107

5.5·107

6.0·107

6.5·107

n(logd)p

0.01 0.10 1.00

d in mmp

ifjfic

sjc

total

µm

n(log d p)

1

Global scale

Regional scale

Mesoscale

Microscale

Par7cle scale

Molecular scale Li#et#al.,#Atmospheric#Environment,#45,#248892495,#2011#


Atmospheric modeling: A multiscale challenge

0.0·107

0.5·107

1.0·107

1.5·107

2.0·107

2.5·107

3.0·107

3.5·107

4.0·107

4.5·107

5.0·107

5.5·107

6.0·107

6.5·107

n(logd)p

0.01 0.10 1.00

d in mmp

ifjfic

sjc

total

µm

n(log d p)

1

Global scale

Regional scale

Mesoscale

Microscale

Par7cle scale


0.0·107

0.5·107

1.0·107

1.5·107

2.0·107

2.5·107

3.0·107

3.5·107

4.0·107

4.5·107

5.0·107

5.5·107

6.0·107

6.5·107

n(logd)p

0.01 0.10 1.00

d in mmp

ifjfic

sjc

total

µm

n(log d p)

1

Global scale

Regional scale

Mesoscale

Microscale

Par7cle scale



How do models represent aerosol composition?

amount

amount

radiusradius

Modalradiusradius

Sectional

I Simplifying assumptions regarding the aerosol compositionI Sectional model: aerosols in a bin are fully internally mixed.I Modal model: aerosols in a mode are fully internally mixed.


Alternative representation: Particle-resolved

Particle-resolved

I Use a discrete representation of particlesI Representation of processes are straight-forward to modelI No bins or modesI No assumption made regarding how particles are mixed


N. Riemer, M. West, R. A. Zaveri, and R. C. Easter, Journal of Geophysical Research, 2009

Model verification of aerosol representation

We need approximations at the regional and global scales.But approximations cause error and uncertainties.

0.0·107

0.5·107

1.0·107

1.5·107

2.0·107

2.5·107

3.0·107

3.5·107

4.0·107

4.5·107

5.0·107

5.5·107

6.0·107

6.5·107

n(logd)p

0.01 0.10 1.00

d in mmp

ifjfic

sjc

total

µm

n(log d p)

1

Global scale

Regional scale

Mesoscale

Microscale

Par7cle scale

Molecular scale

amount

amount

radiusradius radiusradius Howwell is this complexity captured?


Particle-resolved modeling technique

What is composition space? Each particle is uniquely represented as anA-dimensional vector with mass composition components {µi1, µi2, . . . , µiA}

Particle 1

Particle 2

Particle 3

BC 3 10 1 SO4 12 3 4 OC 5 8 2

BC

SO4

OC

Sulfate (SO4)

Black carbon (BC)

Organic carbon (OC)



Particle-resolved modeling technique

What is composition space? Each particle is uniquely represented as anA-dimensional vector with mass composition components {µi1, µi2, . . . , µiA}

Particle 1

Particle 2

Particle 3

BC 3 10 1 SO4 12 3 4 OC 5 8 2

BC

SO4

OC

Sulfate (SO4)

Black carbon (BC)

Organic carbon (OC)

Composition space A ≈ 20



Benefits of particle-resolved models

I No approximation need for representing mixing stateI Coarse graining tool: deriving parameters for more approximate modelsI Benchmark and error quantification for more approximate modelsI Detailed studies on the particle scale and experimental intercomparison.

I Scales efficiently for high-dimensional data (number of aerosol species)I Avoids curse of dimensionality

I Efficient algorithms make particle-resolved modeling feasibleI Accelerated binned coagulation (Riemer et al. 2009, Michelotti et al. 2013)I Particle weighting methods to reduce statistical error (DeVille et al. 2011, 2019)I Accelerated particle removal algorithms (Curtis et al. 2016)


Benchmarking approximate models

Simulation inputs and processes should be as similar as possible

I Same meteorological model

I Same chemical mechanisms

I Consistency in emissions

I Identical particle removal processes

I Identical transport algorithms

Only change the aerosol microphysics


Particle-resolved modeling on the regional scale

I PartMC coupled with WRF allows regionalsimulations with highly-detailed mixing state.

I Each grid cell simulates 10 000 computationalparticles - billions of particles for the domain.

I Many levels of detail from the large-scale topopulation level to single-particle details ofcomposition and emission source.

I Computational expense: 300 000 core hours for2 day simulation from the domain to right

106 107 108 109 1010 1011

number concentration # / m3

pt natgasonroad hddiesel

nonroad gas4strk

np other pt other

nonroad diesel

10�9 10�8 10�7 10�6

dry diameter (m)

0.0

0.5

1.0

OC

mas

sfr

action

101 102 103 104 105

number concentration #/cm3

SO4

NO 3

OIN O

C BC

0

1

2

3

mas

s(k

g)

⇥10�19

aircraftegu natgas

nonroad diesel

nonroad gas2strk

nonroad gas4strknp natgas

np other

onroad ldgaspt natgas pt other

Population level

Single-particle level

Source contributions in grid cell

Source contributions in single particle


Curtis, Riemer and West, Geoscientific Model Development, 2017

How do we move vectors of particle composition?

Transport PDE→ Discretize in space, time, and particles→ Determine probabilities→ Sample particle sets

qt+1i,j − qti,j = ∆tF ti+1/2,j

−F ti−1/2,j

∆x + ∆tF ti,j+1/2

−F ti,j−1/2

∆y

(a) (b) (c)

Replicates deterministic finite volume method to isolate importance of representation


Simulating stochastic aerosol transport

Testcase: 1D constant positive u advection (third order)


Results: Simulating stochastic aerosol transport

0.0 0.5 1.0x

0.00

0.25

0.50

0.75

1.00

1.25

q

spatial order = 1

0.0 0.5 1.0x

spatial order = 2

0.0 0.5 1.0x

spatial order = 3

0.0 0.5 1.0x

spatial order = 4

0.0 0.5 1.0x

spatial order = 5

0.0 0.5 1.0x

spatial order = 6

100 101

Frequency (1/m)

103

Pow

er

1st

2nd

3rd

4th

5th

6th

Odd orders perform better (implicit diffusion)

102 105 108 1011 1014

Number of computational particles npart

10−8

10−6

10−4

10−2

100

Mea

nre

lati

veer

rorē

Converges to FV method in particle number


Curtis, Riemer and West, Geoscientific Model Development (in prep)

Transport performance in real-world case

Complex terrain, complex and evolving wind field

t = 0 hr t = 6 hr t = 12 hr

0 10 20 30 40

Wind speed m s−110−1 101

q



Stochastic algorithm applied to third order monotonic advection scheme in WRF

Npart = 10 Npart = 100 Npart = 1000 FV

10−1 100 101 102

q



Stochastic algorithm applied to third order monotonic advection scheme in WRF

Npart = 10 Npart = 100 Npart = 1000 FV

10−1 100 101 102

q

101 102 103

Number of particles Npart

10−2

10−1

RM

SE

Par

t-F

V


First step: CCN error quantification for a sectional projection

0 20 40 60 80 100

Mixing state χccn (%)

−100

0

100

200

Per

cent

erro

rin

CC

Nco

nce

ntr

ati

on

(%)

100

101

102

103

104

105

grid

cell

cou

nt

0 2

0

1Externallymixed

0 2

0

1 Internallymixed

0 2

0

1

Underestimated

Overestimated

Complex mixing stateresults in larger error


Concluding thoughts

Future work: Model benchmarking

Use particle-resolved modeling andmixing state metrics to benchmarkaerosol models that use varying levelsof mixing state

amount

amount

radiusradius

Modalradiusradius

Sectional

Code availability

https://github.com/compdyn/partmc

Funding

DE-SC0011771DE-SC0019192


https://github.com/compdyn/partmc

TransportExampleConclusion

Documents

Development and applications of benchmarking aerosol models … · 2020. 1. 6. · regional scale using a stochastic particle-resolved approach Jeffrey H. Curtis, Nicole Riemer and