As the “About” section of this repo shows, I am an astrophysicist working on high-energy astronomical transients in radio wavelengths. My journey into ML and AI started when I realized that I have accumulated a large amount of data during my graduate studies. Most of this data consists of light curves, time series of flux (which tells us how much energy is emitted from these events), measured at different radio frequencies (in the ~GHz bands).
Since this is not intended to be an astro-focused blog, I will not go too deep into the physical details. Still, I want to give one example that motivated me to start thinking in a more data-driven way.
Type IIb supernovae as an example
A supernova is the explosion of a massive star (at least 8 times the mass of our Sun). Different kinds of stars result in different types of supernovae, and astronomers have this annoying habit of first dividing things into two types (Type I and Type II), and then realizing there are sub-types and giving them very creative names like Ia, Ib, Ic, IIb, IIp, and so on.
These events are typically discovered in optical wavelengths (the range we see with our eyes), and are then followed up in radio wavelengths (this is where I come in). As it turns out, the vast majority of supernovae ($\sim 80$%) are not detected in radio. These non-detections are actually very informative: they tell us about the energy of the explosion, the surrounding environment of the progenitor star, and its evolution.
If you are interested, I have a paper on the largest sample of supernovae observed in radio wavelengths.
Back to the data. At some point I started playing with this large sample of radio observations, trying to find a way to justify requesting more telescope time for these events. While exploring the data, I noticed something interesting: one specific sub-type stood out.
Type IIb supernovae have a significantly higher detection fraction in radio. More specifically, while they make up only $\sim 15\%$ of the full sample, they account for roughly $\sim 50\%$ (!!) of the radio-detected supernovae.
When you think about their physical origin, this actually makes sense. But to the best of my knowledge, no one had clearly demonstrated this connection observationally.
This was a turning point for me. It suggested that there is likely much more information hidden in the data, even without detailed physical modeling. In ML terms, this is essentially a pattern discovery problem in time-series data, where the structure might not be obvious from first principles.
That realization pushed me toward machine learning. I thought that using ML tools could help uncover new patterns or correlations that we might otherwise miss.
At the same time, I had a concern. I had heard the basic ML concepts many times and could repeat the standard jargon, but I was worried that I would end up using these methods as a black box, without really understanding what is happening under the hood.
So I decided to take a deeper dive into ML and AI.
The purpose of this blog
Most of the work I share in this blog has been done over a long period, and I only recently decided to write it up in this format. My goal is to use this blog in two ways:
-
First, as a way to explain concepts to myself. From my experience, the best way for me to learn something new is to read multiple explanations of the same idea, and then try to apply it to a problem I already understand.
-
Second, and more importantly, to document and share results from projects I have been working on. These will include both successful directions and things that did not work as expected.
I hope that some of these will be interesting or useful to others who are coming from a similar background.
Enjoy the ride!