HPC To Orientate the Universe

HPC To Orientate the Universe

How I end up here

Hi, I’m Andres Vicente but everyone calls me Andreu, I’m an Astrophysicist (it’s official since yesterday in fact) based on the Canary Islands. I heard about the Summer of High performance computing program (SoHPC for short) in the course of programming techniques at my university and I couldn’t resist joining. I can tell that it was an awesome decision. Before I can talk to you about my project and how I planned to use Deep Neural Networks (DNN) in the field of galaxies let’s start from the beginning:

Let’s start by warming up our neurons: Training week

SoHPC starts with a training week where we were instructed with the basics of HPC and where our brains start to get used to thinking about performance and parallelization. The four days of the training fly by even quicker than the executions of our programs in the Supercomputers and then we are on our own, facing the project we dreamed about. But we are not worried, we don’t feel alone, our mentors and peers will make the journey easier.

In my case The project was ” Object Detection Using Deep Neural Networks – AI from HPC to the Edge ” and it was the perfect opportunity to unify both of my passions, the Astrophysics and the High-Performance Computing.

HPC from home to the Universe

The objective of the project is to use DNN to detect objects in images and we are lucky because, in the Astrophysics world, a very big portion of the data obtained from the Universe is in form of pictures. In the past, most of the classification of the galaxy morphologies was done by simple human inspection. Nowadays, we have better tools to classify the galaxies but almost all of them need to be applied using an analytical model and fit it to every one of the observations made, which is obviously time-consuming and computationally expensive. DNN and his object detection capabilities open a new world of possibilities in Astronomy and Astrophysics because we will not only be able to detect morphologies of galaxies (which is a rather easy task), but we also could go beyond that and infer physical properties of the galaxy just by looking at the raw image!!

In our case, we will try to detect the orientation of galaxies (which is closely related to the angular momentum vector if you are wondering). This is not an easy task since we don’t have “training data” to feed our network because we don’t know the ground truth of this magnitude in the observed galaxies but… Here comes the HPC again to rescue us again.

We can mock the observed data with high-resolution simulations of galaxies were we know all the parameters. These simulations are done for experts in the field in the most powerful supercomputers in the world as for example the Illustris simulation done at PRACE supercomputers: https://prace-ri.eu/universe-simulation-illustris-is-an-ongoing-success-story/

From those simulations, we can render images of galaxies in any orientation and point the angular momentum as a vector as we can see here:

These physical properties of the galaxies are relevant because they tell us the story of the galaxy evolution and how it has been formed and evolved. This helps us to understand our own galaxy and somehow why we the Universe is as beautiful as it looks.

With the Idea and the training the data, it’s time to get our hands dirty!!

We plan to use a convolutional deep neural network to do the job. The full architecture is not clear yet but the general scheme (simplifying a bit) will look something like figure 2.

Figure 2: Illustration of the DNN for computing the angular momentum of a galaxy.

I’m very excited to see how the project evolves and I hope you will join my journey and share some of my emotions to unveil the mysteries of the Universe through HPC!!

Keep an eye on the SoHPC blogs if you want to stay posted.

Please follow and like us:
Tagged with: , , ,

Leave a Reply

Your email address will not be published. Required fields are marked *


This site uses Akismet to reduce spam. Learn how your comment data is processed.