Research – Baran Bodur

Super-Kamiokande

You might wonder how could my particle physics research result in me rowing in an underground tank located in an old zinc mine in Japan. That photo was taken during a brief period when we were doing hardware improvements to Super-Kamiokande, normally it looks much different. First of all, it is entirely filled with water (50,000 tons of it in fact). Second, it is very, very dark – except when a neutrino interaction occurs in the water. Those golden bulbs in the picture are light detectors that can see a single photon, to capture that tiny bit of light created by the products of that neutrino interaction. There are about 11,000 light detectors (photomultiplier tubes) looking inside, each of them half a meter across. Data collected from each of these are combined, properties such as incoming neutrinos energy, direction and type are reconstructed. There are also light detectors looking outside of the tank, just to make sure we are rejecting particles coming from outside that are not neutrinos. If you want to learn more about Super-Kamiokande and it’s Noble Prize worthy discoveries go to its homepage.

Why neutrinos?

Neutrinos are fundamental particles – They are made of no other smaller particles we know of. Their name literally means: Little neutral one. What separates them from other fundamental particles – such as an electron – is that they interact extremely rarely with any other particle. That is because they can only interact via the weak force which is, well, weak… They are also the most numerous particles in the universe after photons – little bits of light. To put these two properties into scale: About 100 billion neutrinos pass through your thumb every second, but you may expect about 1 neutrino to actually hit one of your atomic nuclei in your entire lifetime. A bunch of neutrinos can travel through a light year of lead -a very dense material commonly used to stop particle radiation- without stopping.

This unique property of neutrinos is a blessing and a curse at the same time. It is a curse because to study neutrinos we need massive detectors -remember that 50,000 tons of water- just to overcome their extremely low interaction rate. It is a blessing because they carry information to us from where no other particle can reach us. For example, all the light coming from the sun is emitted from its surface. How can we actually know what is happening at the core of the sun then? You guessed it, the answer is via neutrinos, because of their low interaction rate neutrinos can pass right through the sun and come straight from the fusion reactions happening at the solar core. Similary neutrinos can be considered as messangers from many extraordinary astronomical objects (blazars, neutron stars, supernovae…), often carrying information we could not learn via other means.

Apart from being messengers of the universe, neutrinos are also very interesting to study because of their properties. For example, different types of neutrinos can transform into each other as they travel (This is called neutrino oscillations). From this strange property, we recently learned that neutrinos have a very tiny bit of mass; whereas the Standard Model – our best model describing how all particles interact – predicts them to be massless. They can possibly be their own anti-particles, or we might have neutrinos and anti-neutrinos separately. They might be they key to understanding why our universe have more matter than anti-matter. This last one is a huge question in all physics, because if in the Big Bang equal amounts of matter and anti-matter are created, then those equal amounts would annihilate each other, and there would not be any matter to make up the stars, planets, or us. Therefore some asymmetry between properties of matter and anti-matter must be there for us to exist at all, and neutrios are considered one of the most likely candidates to have this asymmetry. You can read more about extraordinary properties of neutrinos in this Fermilab article.

What did I actually do

I am interested in measuring how often neutrino interactions with nuclei happen (Called neutrino-nucleus cross section). Specifically I, for the first time, measured how often a low energy electron neutrino interacts with an oxygen nuclei. Why? Well, if we do not know how often neutrinos interact in our detectors, then we would not know how many we have detected. Therefore, it would be much harder to study the properties and sources of neutrinos which we hope are hiding the mysteries of the universe. For example, when a supernova explodes within our galaxy, electron neutrinos from that supernova will travel to Earth and interact with oxygen nuclei in the water of Super-Kamiokande detector. If we know the electron neutrino – oxygen cross section, then we can use that to estimate how many electron neutrinos are coming from the supernova. By combining this with information from other interactions and detectors we can measure the amount of various neutrino types emitted in a supernova explosion and learn about the supernova process itself. This would shed light to one of the key mechanisms that can create elements like the oxygen we breathe, or the iron we build with…

This measurement was a statistics and data analysis challange. Even after eliminating non-neutrino backgrounds from TB’s of data, I was trying to search for an interaction that is expected to happen ~3 times per year (the actual frequency is what I am trying to measure), compared to ~10 total neutrino interactions per day! I had to develop software to model how these rare interactions appear in our detector compared to other events. This is to find out what sort of variables I could construct from the raw data that will give me a high degree of separation. For example, I have found that the signal events tend to generate more gamma rays compared to backgrounds. To leverage this, I developed a new algorithm that first uses pattern recognition techniques to calculate low level features that differentiate between events with and without gamma rays. Later, I have trained a Fisher distriminant (LDA) that distills all these low level features into one (For this I have also experimented with neural networks and boosted decision trees). By following these procedures, I achieved two things: A smaller sample with significantly higher density of signal events than the inital one, and a small subset of inital features that will be used for statistical separation of events within this sample.

Everything above was studied with Monte Carlo simulations. From these simulations, I modeled probability distributions (PDF) of both signal and background over 3 features (energy, neutron count, and gamma ray probability). Utilizing Monte Carlo is very useful for understanding how the data may look, without actually investigating the which may lead to unconscious biases in my analysis. But using Monte Carlo has the downside of depending on parameters that may not be precisely known. Whether these parameters are from other experimental results (which have measurement errors), or theoretical predictions (which are untested), our imperfect knowledge must also be incorporated into my analysis. Therefore simulations are also performed with variation of all these (over 100) parameters, allowing me to model how the PDFs of signal and backgrounds are varying with respect to these parameters.

The final strech is to estimate the interaction rate from this obtained sample and the PDFs. One of the most complete approaches to this problem is using a full unbinned maximum likelihood fit over the 3 remaining features: This is essentially like solving a tangram: Mixing and matching the fraction of signal and background events, as well as all the parameters in the simulations that had significant uncertainties until their combination give the closest match to the data. I have programmed a custom fitter (called, guess what, TANGRAM) that can first construct the signal and background PDFs for a given set of simulation parameters via kernel density estimation, then calculate the likelihood for a given set of signal and background fractions. One note is that simulation parameters are not fully unknown, they are just uncertain around an estimated value, so a gaussian regularization is added the likelihood for them, encoding our prior knowledge. Once I had constructed this likelihood, and was able to calculate it for any combination of input parameters, it can be used as a cost function (specifically its negative log), so that all of its parameters can be optimized with gradient descent.

In principle, everything is ready here, I could put the data and my PDFs and get a measurement. First, however, the method needs to be tested. If new simulations are performed with randomized parameters can this fit extract the parameters in those simulations? After some effort, the answer was yes. In fact a specific version of it can be seen in one of my poster presentations, which I have embedded just below. See the lower central graph: It shows the linear relationship of the simulated versus the extracted signal fraction, which will lead us to the interaction rate directly.

After all this, I have performed this fit and conclusively observed this interaction for the first time, and even showed that it happens almost twice (1.87) as often than predicted! Of course it has a quiet significant uncertainty of ~30%, yet even with that we are more than 3 standard deviations away from the null hypothesis of seeing nothing!

If this explanation was not long-winded and detailed enough, you are welcome to look at my thesis. As you will see in there too, no part of this analysis would be possible without the diligent work, experience, support and useful feedback of the past and present collaborators of the Super-Kamiokande experiment.