Machine Learning and Cosmology: New Approaches to Constraining the Dark Universe
Sep 5, 2023
Matias Carrasco Kind is a current astronomy graduate student and a CSE fellow at the University of Illinois at Urbana-Champaign. His current research interests lie in cosmology and extragalactic astronomy, especially in large scale structure, galaxy formation and evolution, computational and theoretical cosmology, environmental dependence of galaxy properties, photometric redshift estimation, machine learning techniques and data mining.
Yeah, my name is Matias Carrasco-King from the astronomy department. My advisors for this
fellowship are Robert Brenner from astronomy and John Taller from physics.
And I'll briefly give you an introduction to one of the biggest challenges in astrophysics and cosmology today,
which is basically constraining the dark universe.
So let me give you a little
introduction on what is the current picture of the universe today.
We got this recent image from Planck Collaboration, which is a satellite to measure in the cosmic background.
And so from them, this is one of the most precise measurements of the
of the component of the universe, as you can see here. And this
this was taken when the universe was 300,000 years old. So this is a very, very, very long in distance and in time.
So this is what we know of the universe today, that
roughly 70% of the universe is composed of dark energy and the other 30% is matter.
And from which we have a lot of
contribution from dark matter, almost 26%, and only
4 or 5% of
normal matter, which is the matter that we see in stars and galaxies and so on.
So even though Planck has measured this to very, very accurate numbers, there's still many open
questions about the universe. For example, what is dark matter? What is dark energy? What is the acceleration rate of the universe, which is a
measurement through the Hubble constant.
Well, the universe can be described as a set of cosmological parameters.
And there are several methods to
answer this question or to constrain this parameter. One of them is measuring this at the very early universe.
But the method we're using here is by studying the statistical analysis of special distribution of galaxies.
So seeing how the galaxies are clustering or how the galaxies are distributed in the sky
can tell you something about the underlying dark matter. Because since we can see dark matter, we want to see the galaxies and the stars.
This is the only
observable we have. And we have only this 5% of data to infer the interest of the universe.
So that's the challenge and we try to contribute a little.
So to study
the clustering of galaxies, the distribution of galaxies, we obviously need distances.
So you have the position of the galaxies in the sky, but you also need the distances in the line of sight.
So if we are sitting at the center of this plot and each point here is a galaxy, so you can see clustering.
This big cluster here and voids and all the other features.
And the way these galaxies are distributed in the sky depends strongly on the underlying dark matter that's dominating
this high mass region.
So basically, this dark matter forms these gravitational potentials and all these galaxies form these potentials. So we're studying how this density varies
especially. Then you can infer some information about this.
So how do we do this? We can use either the two-point correlation function or the power spectrum of the
densities
in the galaxy distribution. So we need distances. So this is the first
problem we have to address.
So to measure distance to galaxies, you can use this
full spectra, which is just the light of the galaxy going through a prism.
So you have all the wavelengths and you see all these features coming from the absorption lines or from different chemical elements on the galaxy.
Basically, the position of these lines or features here in white can tell you the shift
with respect to the
rest frame
spectra. And then you can infer the velocity and therefore the distance to the galaxy, all this idea of a distance.
But taking the spectra is very, very expensive in terms of time and money.
So even though you can do that, it's very hard to do it for millions of galaxies. So
in order to do a statistical analysis, you need
millions, hundreds of millions of data on galaxies, high statistics. And for that, the only option you have is
trying to measure distances through photometry. Photometry is just taking images of the galaxies at different filters, right?
So as you can see here, you have filters at different wavelengths. You take the images through these filters and then you get
some information about the underlying
spectra. So basically, the challenge is
trying to measure the distance to a galaxy using 8 or 10 points
instead of the 5,000 points that you will have if you have the full spectra.
So for now, when I say distance, I also say, or when I say photometric redshift or photo-z,
I also say it's the same as distances
because you measure the shift in the galaxies through photometry, not from the spectra.
So as I said, we need a lot of data and a lot of data can be done by taking spectra.
You have to do by photometry because it's very fast taking images of different filters
with large CCD cameras. And there's two big surveys that is very important for astrophysics and very important for Illinois.
One of them is the Dark Energy Survey, which is actually happening right now.
They'll start the official survey on September this year, but they already have data.
They expect to observe roughly 300 million galaxies up to redshift 1.5, which is very far from us.
It's 5,000 square degrees. All the data management is going to be done here at the NCSA.
One of the main goals of the Dark Energy Survey is actually to probe the origin of dark energy.
Obviously, photometric redshift is needed because you don't have all this spectral information. You don't have photometry.
And also, there's this other survey, which is even bigger than the DS,
which is the Large Synaptic Survey Telescope, the LSST, which has green light, so it's going to happen at some point.
They have the place already, and they're building the mirrors and the telescope. The first light is expected to be in 2020.
So it's eight years from now. They'll survey half of the sky.
And they expect this is a huge survey because they expect to have at least 30 kilobytes of data each night.
And also, the data management is going to be here. So these are very important surveys for Illinois because Illinois is actively involved in this.
And we're trying to address the distances for this data.
So basically, this is the main slide of my talk. This is our work.
We have developed an end-to-end cosmological understanding.
So basically, we compute the distances to galaxies, then we study the clustering of the galaxies, and then we fit cosmological parameters.
So from the galaxy images, we apply some statistical analysis to get the distances.
We developed recently TPC, which stands for Trees for Photosynthesis, which is a parallel photosynthesis algorithm that uses prediction trees and random forests,
and also extra information to predict this photometric predictive.
We release the code and we publish the paper along with it.
But also, we continue to develop this because this is a machine learning algorithm.
So it's always limited by the training sample you have.
Since you have some spectra, you can use that as a training data, a training set for your algorithm, but you're always limited by that.
So we also are currently developing a second technique to try to go beyond the training set under a Bayesian framework.
We already have some work done on this.
We're using a random Naive Bayes classifier prior for this combined technique.
And we also try to extend this to a self-organized map, which is unsupervised.
Sorry, I didn't mention, but prediction trees and random forests are supervised techniques.
While SOM are unsupervised techniques.
So we try to combine all this to leverage somehow the weakness and the strength of each technique to get a much powerful predictor.
So once we have the distances to galaxies, we can study the clustering.
For that, we apply a Karhunen-Laghe-Lev transformation, which is basically a PCA transformation or dimensionality reduction to the other densities or to densities of the galaxies.
From there, we compute the angular power spectrum, which is basically the composition of the densities in the spherical harmonics.
So basically, the power spectrum is telling you how strong is the clustering, which is related to how much dark matter is behind it.
And once we compute the angular power spectrum, we can fit a theoretical model using a Monte Carlo or a Markov chain Monte Carlo algorithm and get the best parameters or the best estimation of the parameters.
One of the advantages of using this angular power spectrum is since dark energy is basically driving the universe and is producing the origin of the universe.
So studying the universe at different snapshots in time can allow you to constrain dark energy.
So if you apply the same model to different redshift shells or beams, then you can put some constraint on the dark energy because that's related to the evolution in time, basically.
That's the point of all these techniques.
So I'll briefly go through some of the main features of our whole framework.
This is one of the tree examples that we use in our algorithm.
So you see, you start from the center and you start to split different dimensions.
So in each leaf, basically you have data and then you can make a prediction from this data.
You make hundreds of these trees in a forest way and you combine all these results together to get a strong prediction.
After that, you can even combine this with, as I said before, with a random Naive Bayes classifier prior or self-organized map that works as a prior.
So you can combine them under a Bayesian framework to get an even stronger prediction.
As an example, these data, they have some observations, but they still are in the calibration phase.
So we have some simulated data.
In this case, we have roughly 40 million galaxies that we wanted to predict the distances.
And we know the real distances, but we kept them blind.
So we see here how TPC, our code, works in prediction these distances.
So you see here in this axis, the vertical axis, the prediction.
And in the horizontal, you see the true distance.
So you see that it's very well correlated.
It's unbiased, so you don't see any tilt.
And that's a very good job.
So one of the ideas of developing TPC was to be an iCreate estimator and also able to deal with a lot of data.
So we're talking here about 40 million galaxies.
And we're not only producing the distances.
We're computing the PDF of the distances, or DPF of the redshift.
So basically, we have a probability distribution for each galaxy.
So we're talking about roughly 20 billion data points.
Because for each galaxy, we have roughly 500 points in the distribution.
So one important measurement that is needed in cosmology is the galaxy distribution.
So basically, how the galaxies are distributed.
Basically, it's a histogram.
It's a function of distance in the light of day.
So relative zero means today.
Relative 1.4 means very far in time and space.
So on gray here, you see the true distribution.
And on red here, you see what happens when you stack your prediction PDF, your prediction probability.
You stack them together to make one of these plots.
So basically, you are producing these 20 billion points to this curve here.
And you see they match very well, which is very important.
If you use the mean or the mode of the PDF, just one estimator instead of the 500 that I mentioned,
you see this fluctuation due to either shut noise or just biases.
So this galaxy distribution is very important in cosmology
because you can think in terms of basically the mass within this stretched range.
So since each point is a galaxy, and you can associate a galaxy to a mass,
even though they're not the same, this is a measurement of the mass of the universe.
So you can apply this for different cosmological studies.
So at the end, the point I tried to make here is that this galaxy distribution is very well reproduced when you stack this PDF.
So you have to keep all this information, which is a problem when you go to hundreds of millions of galaxies.
You have to store all this data.
So after you calculated these distances, you can calculate the angular power spectrum.
I won't discuss this in detail, but basically the shape of this angular power spectrum,
which is the composition of the over-density clustering in spherical harmonics, depends on cosmology.
So depending on the shape, you can see how dark matter or dark energy
or different other cosmological parameters affect these measurements.
And if you calculate this angular power spectrum at different relative beams,
like if you split this in five or six beams, and you compute the same,
you can't put constraints on dark energy because you can't see the evolution of this clustering.
And to finish our framework, so we computed the distances,
we computed the clustering analysis or how the galaxies are clustering in the sky,
and then we're using a Markov chain Monte Carlo, we can fit parameters.
We have to use a Markov chain because we normally fit five or six parameters.
And this is one example of a preliminary result.
For example, you can fit this omega here, basically the densities from the pie plot I showed at the beginning.
So we have dark matter and baryonic matter.
You can fit the baryonic matter just by itself.
The bias, which is basically, yeah, it's difficult to explain,
but basically it's the difference between the clustering galaxies and the clustering of the dark matter that is behind.
Some other parameters, like the Hubble parameter and the power spectrum.
So that's what I wanted to say.
As a conclusion, we have that photosynthesis is becoming very, very important for extragalactic study
because all the current surveys are just photometrics.
They have some training, but they're all photometrics.
We developed TPC and some other techniques to predict distances to this data.
We plan to extend this using a self-organized map.
We have a code that accurate and make a robust calculation of the angular power spectrum using this KL compression.
And we use this Markov chain Monte Carlo for parameter estimation.
Yeah, we plan to extend this to other statistical measurements.
Thanks.
Thank you.

