Parallel Stochastic Particle Methods for Atmospheric Aerosol Modeling
Sep 5, 2023
Jeffrey Curtis is a graduate student working in the Nicole Reimer group. His current research interests are in using the aerosol particle resolving model PartMC-MOSAIC and the Weather Research and Forecast model to examine spatial aging of black carbon aerosols and the associated impacts due to their optical and CNN activation properties. He also is interested in errors incurred due to the assumptions used in conventional aerosol representations in regional weather/climate models.
So we'll have three talks by CSE fellows.
Our first talk will be by Jeff Curtis from the Department
of Applied Sciences.
The title of the talk is Parallel Stochastic Formulas
of Development for a Mysteric Error Goal Model.
OK, so today I'm going to talk about the project that I'm
working on regarding the parallel stochastic particle
methods for atmospheric aerosol modeling.
I'm Jeff Curtis, and my advisors are Nicole Reamer from
the Department of Atmospheric Science and Matthew West
from Mechanical Science and Engineering.
First off, I wanted to define atmospheric aerosols.
So an aerosol is defined as a solid, liquid, or mixed-phase
particle suspended in the atmosphere.
So here I have some images of what these aerosols may look
like in the atmosphere.
And the important thing here is these particles consist of
various different chemical compounds due to coagulation
processes where two particles collide, and also
condensation of gases.
So we have these complex structures where we have
ammonium sulfate and dust and soot, and these can all
coexist within one aerosol particle.
So why are these aerosols of particular interest?
There's three major reasons why aerosols are relatively
important.
One of the reasons is the health impact.
Here in this image here we see this woman walking down the
street, and this haze is the aerosol particles.
And breathing this in has been linked to respiratory distress.
And also we have aerosols can scatter or absorb solar
radiation based on their composition size, and also
there's interactions with clouds.
Aerosols act as seed particles for brain formation.
So these two are of particular interest for global climate
change, so it's important to understand aerosols.
Okay, so a little bit about atmospheric modeling.
Atmospheric modeling ranges a whole wide span of sizes.
We have things that are occurring on the molecular scale,
things on the particle scale, which is like nanometers to
micrometers, which I'll be talking about today.
And then we have these scales that all fill in the mesoscale
and regional scale, which is on the order of like tens of
kilometers, and then finally to a global scale.
So there's real difficulty here.
In a global scale model, we can't represent all these small
scale processes.
But we hope to parameterize them in a way that we'll still get
the answer right.
So we have to look at these particles and understand them
in order to do that well.
So introduction to my project.
My project goal was to develop a particle resolved aerosol model
that also has full transport, so we have a spatial distribution
of these particles.
So on the left here, we have a model that I'm using,
PartMC.
It's a particle resolved Monte Carlo model.
It's a box model, and we have all this wonderful information
regarding size and composition.
However, since it's a box model, we have no spatial resolution.
On the right here, we have sort of the regional weather model
of choice these days.
It has a very large developing community.
It's the wharf model, and in the wharf model, we can resolve
the dynamics of the atmosphere.
So here in this image here, we have a bulk quantity that's just
like a mass concentration of a certain species.
And then on the left here, we have all this detail regarding
size and composition of particles.
So the real benefit of this project is I'm going to take the
two weaknesses of these models and sort of combine them
so we avoid those problems.
However, as you can imagine, particle resolved modeling
is very expensive.
So one, how do we accomplish this?
And more importantly, how can we actually do this
in an efficient manner?
Okay, so the concept is on the left here, we have basically
like a particle flow field, and then I've highlighted out
a certain little region here.
So on the left here, we can see all the particles look the same
from this distance.
When we zoom in, we can see the particles are very different.
They're going to interact in different ways,
and their composition is different.
So what we do is we are going to model coagulation by the
probability of two particles interacting, and then gas-to-particle
conversion is a deterministic process.
So what we're going to do is we're going to actually ignore
position within these boxes and just represent the aerosols
by this, where we have location at x with composition mu
and at time t.
So here is the model equation.
So the size distribution changes based on, obviously, transport.
So if particles are being transported to that spot,
we're going to change the distribution.
And then there's the coagulation gain and loss.
So when two particles coagulate, one larger particle is created
and two particles are destroyed.
Obviously, emissions.
And then we also have chemistry, where we have gases condensing
or evaporating off the particle.
And finally, air density changes.
So how do we go about representing those in weather models?
So as I've mentioned previously, one of the challenges is that
particle sizes span three orders of magnitude and can be composed
of a wide variety of chemical species.
And also, to complicate the problem, all these things
are constantly changing.
So on the left here, the two left images here are typical ways
that atmospheric models currently represent their aerosols.
One popular way is a modal approach, where you represent
the distribution of the solid modes.
This is particularly appealing because it's very computationally efficient.
Much more accurate but much more expensive is this idea
of sectional modeling, where it stores number and or mass
in each aerosol bin.
However, in each bin, every particle will have the same size
and the same composition.
So we're making an assumption there.
And on the right here, we have a particle result model
where we can easily represent very complex mixtures.
So particles can be the same size but have drastically different compositions.
So how do we do that?
Here's a schematic of how PartMC represents particles.
So over here on the left here, I have the particle field from before,
take three particles.
We've taken these three particles and we've put them in a vector space
where we have black carbon, sulfate, and organic carbon.
And basically, we can represent the constituent masses
in this vector space.
So this is very convenient in comparison to other ways of representing the data.
And this is why.
So here I have a true solution, a modal solution, a finite volume solution,
which is equivalent to the sectional solution,
and then a particle solution.
So the major advantage of this particle representation is what we like to call
is it breaks the curse of dimensionality.
In a particle result model, our calculation scales with the number of particles
rather than the number of species.
So say in the Whorf model, there's a sectional model
and it consists of eight size bins.
And if they were to represent the number of species that our model represents,
where A is the number of species, it would be 8 to the 20th.
Now obviously for memory reasons, that is not feasible.
So with this particle model, our storage is going to scale
based on just the number of particles.
You can use 1,000 or you can use 100,000,
but it's still significantly going to be less than 8 to the 20.
So what actual weather models do is they're going to project all these things
down to a series of one-dimensional distributions.
So the benefits of the stochastic processes in the model,
one such interaction is particle coagulation.
That's when two particles are going to collide.
So if we have n particles, there's n squared pairs of coagulation events.
So if we were to naively do this, we'd compute the probabilities
of each of those particles coagulating with others
and then test all those events.
However, in actuality, particle coagulation is heavily dependent on particle diameter,
so we can actually do a much more efficient way where we stochastically sample
based on the number of expected events and then accept-reject those coagulation events.
So we go from n squared calculations to just the number of coagulation events expected.
Then also a particle transport.
Naively, we could actually track all the particles.
All the particles in the atmosphere could have a position in space,
and we can update them each time step.
In actuality, we're going to ignore the particle location within a grid cell,
and then we'll stochastically sample the number of particles
that should be transported from one grid cell to another.
Just to briefly touch upon how we make this more computationally feasible,
on the left here I have...
Okay, so my model right now is just a vertical column model,
the vertical column being the most atmospherically interesting.
This is why we chose to do that.
So here I have four grid cells, and I've subdivided them across four processors in space.
However, with the particle populations,
we can actually distribute the particle population amongst many processors.
So we can actually apply many processors to a single grid cell,
which is really an interesting aspect
in comparison to what a weather model currently would be doing.
And finally, I have a brief example of the benefits of the result that we end up getting.
So on the left here...
Let's see here. I have it right here.
Okay, so on the left, vertically, we have what we call black carbon mass fraction,
and this is the mass of black carbon in a particle
divided by the total mass in the particle.
So this is basically showing how much black carbon is in a particle.
Black carbon is of particular interest because it absorbs solar radiation,
and it also dislikes water.
So the less black carbon in a particle, the more likely it can act as a CCN.
The more BC mass fraction in a particle,
the more likely it's going to absorb solar radiation and cause warming.
And then on the bottom here, we have just the particle dry diameter.
So on the left here, I have three different vertical levels in my model,
and we have lots of information.
So if we take a particle that's 10 to the minus 1 micrometers,
we can see that we have many different compositions.
However, what I can do with the particle resolve data
is actually I can first compositionally average it and then size average it,
and then that ends up being sort of the result
that a normal sectional model would give us.
So like right here, we have this point, which is sort of representative,
you know, it's representative of this diameter here.
If we go over here, we have all this information
regarding the particle composition at that given size.
And this has many ramifications, especially regarding radiative balance.
Okay, and just to conclude, I wanted to discuss the applications of the model
and some of the future work.
As a result of this model, we now have multidimensional aerosol size distribution
that's now spatially resolved.
So we've taken the box model and gone beyond the box.
And it may be applied to conducting detailed studies on particle scale
to study what we call particle aging.
Particle aging is how these black carbon particles, you know,
how they evolve in time due to processes such as coagulation and condensation,
which drastically changes their optical properties.
Another result is we can use this model to benchmark
against more approximate aerosol models.
We can compare it to current state-of-the-art sectional models.
And also the most interesting to me personally
is quantifying the error introduced in atmospheric models
due to the assumptions regarding the particle size and composition.
So we can take this particle resolve data
and compare it against model output from a sectional model.
And we can see how we can actually quantify that error
that's been in models and hasn't been previously quantified.
And then in terms of future model development,
like I mentioned earlier,
we can incorporate the use of a group of processors
to simulate a single grid cells population.
So, you know, we can instead of maybe 10,000 particles,
we could use 100,000 particles and distribute them on 10 processors
if we have the computing resources available.
And then we would like to extend this model to 3D,
which increases the computational burden
from approximately maybe 100 grid cells to the order of millions
to properly represent the domain of a field campaign, per se.
Okay, that's it.
All right, I think in the interest of time,
I realize there are really interesting questions,
so let's move on to the second speaker,
who will be Abhishek Singh
from the Department of Electrical and Computer Engineering.

