 Hello, everyone. Welcome to this version 137 of the Latin American webinars in physics. My name is Walter Pengarife. I work at Loyola University Chicago. And it is a pleasure today to have Professor Lena Nesif from the Massachusetts Institute of Technology, who will be sharing with us her work. Dr. Nesif got her PhD at MIT in 2017, after which she worked as a postdoctoral fellow at Caltech until 2020. And then at UCI from 2020 to 2021, after which she moved to MIT to start working as an assistant professor. She works on very interesting and exciting astro-particle topics, such as dark matter, the Gaia data analysis, galactic dynamics, and putting all these together with machine learning. We are very excited to have her today, sharing her work with us, in which she will tell us about searching for the dark, using light. Thank you, Lena. And welcome to the Latin American webinar for physics. Before you start, let me remind everyone that you can ask and please ask your questions on the channel and YouTube, on the chat. We'll be happy to share these questions with Dr. Nesif after her presentation is done. And please, if you haven't done so, subscribe to all the social networks and to the mailing list, so that you are up to date with the upcoming webinars. All right. Thank you, Lena, and take it away. Thank you so much for having me. I'm really excited to be telling you about all the cool things that I've been kind of thinking about with respect to dark matter and galactic dynamics and particle physics, all the combo. So hopefully you can see the screen. All right. So let's dive right in. And I promise this picture will make more sense in a little bit. So what I would like to do really in the beginning is to kind of put everything to scale. So I usually start with a picture of the Earth. And I'm very excited about this one is that because it shows my home country, Tunisia, right here. The weird thing is that it shows that Tunisia is cloudy for some reason. That doesn't really happen. So, you know, but let's try to kind of get out of our earth and try to see where we are to kind of put everything to scale. So the first thing that you're going to see is the sun that is the closest start to us. But then as we exit the plane of the Milky Way, the galaxy that we're in, we see that there are so many more stars actually, billions and billions of stars. And we're rotating in this huge galaxy that we call the Milky Way. What I would like to emphasize is that we finally have the computational power to have simulations of galaxies like our own, and in particular, to have very high resolution. Not only can we actually simulate the galaxies and the stars, but we can actually simulate properties of these stars. For example, here, what you see is the gas metallicity, which is like the chemical abundances of the gas and that translates into the stars later, as well as the gas temperature, which is very important in key star formation in different spots. And most of all, we can kind of zoom out and see the cosmic web, which is how the particles are all interrelated into our picture of the universe in what we call lambda CDM. So what I would like to do after kind of going at this whole view is really ask the basic questions. What is dark matter? The short answer is we don't know, but I'm going to start telling you the stuff that we actually know. So dark matter is a solution to a problem that showed up in under various forms in the 30s and in the 70s. So I'm going to just focus on one of the aspects of it, which is what we call rotation curves. So since the 1970s, before the 1970s, we thought that we lived in a galaxy that looks like this, where you can see here that the stars at the edges of the galaxy are rotating a lot slower than the stars at the center of the galaxy. With the work of Vera Rubin and collaborators, we found out that that's not really the case. We actually live in a galaxy that looks a lot more like this, where you can see that the stars at the edge, which as you remember from the previous video, the sun actually does live at the edge of the galaxy, are rotating as fast as the stars in the center. So basically, there is a bit of a gap in what we call rotation curves. So here you have that on the x-axis, you have a distance from the center of the galaxy. And on the y-axis, you have this circular velocity. And if you didn't have dark matter, or you just make the mass out of everything that you see, you would expect that the circular velocity to drop after you exists like the amount of mass that you have. And what we're seeing is really that actually is a lot flatter. So dark matter is a solution that makes up for this difference right here. So what the focus of today's talk is going to be is really about where is dark matter or the properties of that density of what we call the face-based distribution of dark matter or like basically mapping out the dark matter. So first things first, our galaxy is living in this huge halo of dark matter. So dark matter basically extends much further than our galaxy does. To put in some numbers just for later, the sun lives at the edge, so at eight equal to five kilofarsec away from the center of the galaxy in a galaxy that has a radius of 10 kilofarsecs. And it is part of the disk of the Milky Way that's rotating at about 220 nanometers a second. The halo of the Milky Way extends about an order of magnitude larger than that. So basically we're swimming in this huge sphere of dark matter. And today's talk is going to have kind of three aspects to it, where I'm going to show you that from simulations we can use dark matter and stars, we can use stars to actually poorly and understand the face-based distribution of dark matter. This is going to be brief where the main focus is where I put the star here is really using these machine learning techniques to be able to disentangle the face-based distribution of these stars from Gaia, so from observations. And then the last part is going to be putting one and two together. So kind of building these correlations and simulations and then testing them in data to really have an understanding of the face-based distribution of dark matter. So let's start with the simulations. How can we correlate the dark matter and the stars? And the key part here is that it's going to be some of the dark matter to some of the stars. So in one of our simulations, our galaxy looks like this. So basically here, what I'm showing you is the density map of the positions of dark matter. So the different colors actually show different densities of dark matter. The first important thing to note here is that this is very different from, you know, the black sphere or circle that I was showing you earlier, that which tells us that dark matter has a lot more structure to it than we expected. So this very heavy, high density spots are basically where our galaxy is going to live. And in particular, the central point here is that of the host halo. So that's where, you know, that's where you can think of it as like the deepest point of the potential well. So the stars and the variables are going to just fall in there. But there are also other structures that I'm calling here, you know, subhalo and dark subhalo. And the reason for it is because I'm going to move away from the map of the dark matter to the map of the stars. And you can see that there is a bit of there are some densities that are somewhat similar. So the maps are not exactly the same, but there's some correlations to it. The higher density of the stars is living in that host halo. And the, that subhalo that I circled earlier has a matching overall density in stars. However, that dark subhalo does not. So the correlation that I'm going to be talking about today is really in terms of velocities, but I would like to actually kind of emphasize this visual to say that there are some connections between the dark matter and the stars. And we can actually use these stars to be able to investigate dark matter. So what we're going to do is actually use cosmological simulations to track where the dark matter is coming from and track the dark the stars that actually came with it. And in particular, I would like to show the evolution of one of the simulation of one of the simulated Milky Way halos that we have from the fire simulation. So fire stands in feedback, feedback and realistic environment. And it was pioneered by multiple people, including my postdoc advisor, Phil Hopkins at Caltech. And you can see how the galaxies that form that I'm going to show you here actually form through a very cannibalistic process. So this is called hierarchical galaxy formation. Stars, sorry, galaxies form by eating other galaxies. So let's see that in action. So here, what you're seeing is the map of the stars, but you can see that there are multiple mergers that happen. So other galaxies that come in and kind of fall into this main halo and things kind of equiligrate at the end of the day to form a spiral galaxy that looks very much like our own Milky Way. So galaxies form through mergers and that's the key point that I will be emphasizing in today's talk. So if they do form through mergers, it means that if we can track the stars that form through the mergers, we can also track the dark matter that came in with them. To do that, we can take a Milky Way like galaxy, and we have the location of the sun over here. And then we look at this. Basically, we kind of look at the equivalent of the solar neighborhood in our simulation and try to track the dark matter in the stars that actually landed in that solar neighborhood at Richard zero. So today, basically, where did my dark matter and star particles come from that ended up in the solar neighborhood? And indeed, a lot of the dark matter is actually because we show that a lot of the dark matter comes from different mergers. For example, this is one of the mergers where here I'm plotting the velocity distribution of in VR in spherical coordinates, as well as be tangential and the full speed distribution at the end of the dark matter and the stars that are within that solar neighborhood that I picked, right? So basically what I'm doing here is tracking back the dark matter and the stars that came from this one specific merger and I'm plotting their velocity distributions today. And what I see is that the dark matter and the stars are actually very correlated. Now we wonder why, like, did we expect this? And indeed, we did because if dark matter is collisionless, which is a big assumption that we're making here, then dark matter and stars had the same initial conditions, they came in from the same merger, and they're only both of them are governed by gravity. So they end up in the same place. There are a lot of caveats to this that I'll mention. But basically, this is the idea. Stars in a particular immigrant stars, the ones that formed in other galaxies emerged in are going to be a good indicator for the dark matter that they came with. So this does not happen for a single merger. For example, this is a pretty, pretty young merger. It came in an equation, a creation range of the point seven. Same thing for an equation, Richard, point eight, or an equation range of the 3.7. So this is much, much older. And in particular, when this happens, when mergers happen that all, you expect their velocity distributions to be more Gaussian or more like that Maxwell-Boltzmann distribution because the entire system has since virialized and thermalized. However, there are exceptions to this. And in particular, for example, this one, this merger here is extremely massive and happened at a very recent time of Richard, of Richard point two. So you can see that the dark matter and the stars, they kind of tried to be following each other, but it's not really the case. And the reason for why this is true is that this is a very recent merger. So the differences in the initial distributions between stars and dark matter actually are still present. So here's how you can think about it. In the earlier, in some of the earlier schematics that I had, the stars are at the center, and their dark matter halos are further extended. As they fall into a new galaxy, the dark matter, like especially on the outskirts kind of gets etched on first, that kind of gets completely tidally disrupted. And then at some point, you're going to start hitting the stars. So if that happened just very recently, then you're still sensitive to, you're still sensitive to the difference in that initial base space distribution between the stars and the dark matter. However, if you've done that after a long time, then the dark matter, you're only sensitive to the dark matter and the stars in the very inner part. I'm happy to kind of discuss this because there's a little bit of a technicality here. And I'm happy to answer questions about it later. But basically, for more generally, anything that is a little bit older in terms of merger, the dark matter and the stars actually are very well correlated. So that leads us to kind of having this picture in mind, where you have the dark matter and the stars and as long as you can isolate the stars that coming from a particular merger, you can identify the velocity distribution of dark matter coming from that same merger. So now the game is trying to get identify which star came from which merger out of all the stars that we're measuring with observations and telescopes. And this is really not an easy especially because we have a lot of data, which is a very, very good problem to have. So the way that I'd like to organize things is in terms of a treaty or kind of like a bit of a family tree for the Milky Way. Because what we're trying to understand here is where what are the mergers that made our Milky Way. So I'm going to organize this by time. So the most recent mergers I'm going to be here on the bottom. You can think of those are like siblings, first cousins versus the older mergers are going to be all the way in the top. Okay. So what do these mergers look like? There are different stages to them. In particular, we're going to have a galaxy. A galaxy that's going to fall into the Milky Way. And if we see it just like this, it means that it's still in very early stages of merging. Then if it's been a long time, it's going to start getting tightly disrupted into a stellar stream, which then loses its face to face distribution and turns into debris flow. So basically, where these are the different stages, this is how time goes. But if you see emerging galaxy means this has is very, very recent in its merging time. And debris flow is much, much older. So let's start kind of breaking this into these different stages. Let's start with merging galaxies. So here, actually, I kind of cheated and showed you evidence for emerging galaxy in that first, first video. And if you don't remember, let's kind of go back to it. We're going to start with the sun and move out. We have our galaxy, but we have something that's smashing right into us that we call the Large Magellanic Cloud. And in particular, there's also the Small Magellanic Cloud, but let's stick with the big one for now. So the Large Magellanic Cloud is what we call in first impulse. It means that it's a galaxy that is just kind of just falling into the Milky Way now, but it has enough, it's making enough of a mess that is affecting the potential of the Milky Way. And it's something that we really need to take into account when we're doing orbit and integration and all of that kind of stuff. So because it is a kind of an intact galaxy, the Large Magellanic Cloud is still very new. So it should be, you know, as the young kind of siblings part of the family tree. It's not the only galaxy that's smashing into us. And in particular, there are a lot of dwarf galaxies that are actually currently orbiting the Milky Way and that we use for different dark matter studies. So it's definitely not unique, but it is actually the most massive one. And I've been told that in some hemisphere you can actually see it by eye, which is amazing because I haven't been there since one of my personal goals. So the next stage of mergers is going to be a stellar stream. So what is a stellar stream? Basically, you have our galaxy in the Milky Way here in the middle and then I'm going to throw in three, these either small galaxies or globulic clusters, basically three bound objects of stars. I kind of let them be. And you can see here that they get, they start getting tidally disrupted. And the reason for it is because there is a bit of a gradient in the gravitational field on the both sides. So then every object has, ends up kind of developing this leading arm and tidal arm. And you can see these as like these large kind of structures in the sky and they form stellar streams. So these are in position. You can actually see them tracing orbits, although experts in the field get very, very mad when you tell them the streams are tracing orbits because they are not exactly orbits. They're almost orbits. Anyways, we can actually see these with telescopes, in particular there is, for example, the field of streams. This is taken by the Sloan Digital Sky Survey when you can see that there are over densities of stars that are kind of like are these in these linear structures actually, which is quite amazing. And indeed, basically this presence of streams is not just like a theoretical concept. We've seen these. And it's very interesting because they tell us about the formation of the Milky Way. However, not every stream is emerging galaxies. In particular, these three are actually globular clusters. Globular clusters are a collection of stars that are born in the Milky Way out of like gas in the Milky Way, which means that they really don't tell us something about the merger of the history of the Milky Way because they were just born here. And they do not belong in our family tree. However, we can use chemical abundances to be able to disentangle them from the true mergers. And for example, just like the Sagittarius stream. So, Sagittarius is indeed a big dwarf galaxy that is currently third in fall, actually, into the Milky Way. In more detail, Sagittarius started falling into our galaxy about eight billion years ago and did one, two, three passages right now. And now it looks like this huge set of stars that are kind of like bounding the Milky Way in this absolutely stunning and beautiful room. So, Sagittarius has started falling a few billion years ago, which means it's a little bit older than the Large Magellanic Cloud and that's why I have it a little bit higher in here. When we go next, we have debris flow. So, debris flow is basically, if you think about a stream after, imagine a stream that has done so many orbits, you can't tell us if a star is like, is it the first in fall? It becomes basically kind of like a collection of stars and you can't really tell these beautiful lines. So, how can we distinguish debris flow? They have a very specific velocity distribution that is very different from the other stars. So, they're coherent in space, but they still have a velocity, a specific velocity distribution. So, in order to do that, we need basically a very expensive speed gun is what I call it and that's where Gaia comes in. So, Gaia is a space telescope which was launched in December 2013 and has given us the goal of like the positions and perpendicular velocities or proper motions, kinematics of 1.8 billion stars. That is an incredible number of stars. The second data release from 2018 also gave us the sixth dimension, so the line of side velocity of 7 million. That 7 million has actually grew to 33 million as of last June with a third data. So, Gaia basically is just, it's incredible because they really can tell us about so much about these stars. So, you can make videos like this where you can just like zoom in on a very small subset of 96 million stars and evolve them forward for the next 800,000 years and you can see how our galaxy is going to move. So, I cannot even overestimate like the power of what came out of this mission and in particular one of the most amazing things that came out of it is what we call Gaia Enceladus. So, Enceladus is a Greek god, son of Gaia and it kind of, you can see him coming out of the Milky Way here because it is actually a merger from the Milky Way that happened about 6 to 10 billion years ago with an initial mass of that was about 1 to 10 percent of the mass of the Milky Way. This was first seen by Vasily Blukarov in 2018 as well as a paper by Amina Helmi that was about the same time, but basically what we could see here is that if you pick a metallicity bin and we don't have to go into these details, you can take a look here at on the x-axis we have the radial velocity in galactocentric spherical coordinates but basically imagine yourself at the center of the galaxy and you have spherical coordinates. This is VR and on the y-axis you have v5, the rotational velocity. So, most of the stars you would expect are rotating just like the Sun at about 220 kilometers a second which means that you would expect them to live about here in this big club. What was new is this whole haze that there is something that has pretty high radial velocities and not that much of the v5 and that just stood out. We didn't know what it was. In a paper that we worked on in 2019 about that was the first kinematic modeling and substructure we actually saw it kind of in the radial velocity to actually have like this little bit of a double peak structure and indeed that kind of hit us a little bit of like wait I've seen this before and indeed we have because the simulation's murderers actually have that very some murderers have that specific double peak structure here in radial velocity. So, this double peak structure is basically coming from the fact that you can imagine that you have a sphere and you're kind of shooting an arrow through it and then the arrow does like a very tight 180 and comes back as it goes in it's going to spike in negative vr and as it comes out it's going to spike in positive vr. So, basically this is telling you that like the this is what we call the very radial orbit. So, you can imagine it's not a circular orbit it's like it actually the complete opposite. It's a very very elongated meanless. So, here is what Gaia Enceladus and the unfortunate name that Vasily Lukarov gave it the Gaia sausage. Unfortunately that sticks a lot more and people remember that a lot better. Personally I do too. So, we have the Milky Way here in blue and in red below we have the Gaia sausage and we let the Gaia sausage kind of walk in and you see it smashing into the Milky Way and kind of falling into this extremely extended and unfortunately correct sausage behavior that kind of if you can imagine the sun being on or out here it actually dominates quite a bit of the stars there. So, this merger has fallen a little bit over a long time ago much older than Sagittarius and has left a huge debris of stars behind it. Which is why it belongs a little bit higher in our tree so that's Gaia Enceladus. Now the question is how do we find the rest? So, if we wanted to be a little bit more pessimistic about it the large Magellanic cloud you can see by eye. Sagittarius you can see it with a nicer telescope that's true but it has kind of like you can actually see it in the sky because there are like there is an over density of stars. Gaia Enceladus you cannot see it very fair but you can kind of if you take the Gaia data and make some cuts it just stands out and I show you the plot you just like stare at it and it just comes out. What about the rest of that 1.8 billion stars? Unfortunately we can't just stare because there are so many limitations to what we can do although our brains are the original the neural network we can only cluster things in our heads in 2D more than that things just really for all like it really kind of we can't do it anymore so that's why when working on work now is actually automating the identification of cells of structure basically teaching a computer to do the hard work for me because there is so much that my brain can do. So I'm going to show you a brief discussion of like three different methods that we used to be able to kind of disentangle all the different structures and trying to come out with more things here and understand what's going on in the galaxy. The first thing is that how to train on simulations to identify accreted stars. So this has a bit of a jargon here but basically as I was saying earlier what we want is to figure out the stars that were born outside the Milky Way. Those stars are also called accreted stars or in my mind immigrant stars. These are they kind of hold the key about what happened to dark matter. The unfortunate part is that only one percent of the stars very close to us that we can measure very well are actually accreted. Everything else is a disc galaxy. It's a disc star that is interesting for other sciences not personally interesting to me because it's not going to tell me much about matter. So this is kind of the true situation of like a needle in a haystack. How can we figure this out? Well we can use machine learning techniques and in particular we can show a neural network what a cat looks like and then it says okay got it that's a cat and this is what a dog looks like and then you just ask it a question is this a cat or a dog and at some point it does this whole thing and then it says okay it is indeed a dog. What we can do is the same thing. Here is I'm going to teach you what an accreted star looks like in simulations because simulations because in our simulations actually we can tell the difference between stars that are accreted versus the ones that are not and then we can look at we can use that to train on new Gaia stars. So the top one is Anankee. So Anankee is actually a synthetic survey built by Robin Sanderson and basically it answers the question what would Gaia see if we lived in one of the galaxies that were simulated? And the bottom one is indeed Gaia. So the goal here is to train on simulations and then label the Gaia stars however we're not really doing this on pictures about the cats of the dog but we're really doing it on kinematics so basically we're adding the positions that and the positions in the velocities that Gaia can measure and ask is this star accreted or not. So using that we could apply it to Gaia especially in the solar neighborhood and kind of build a Gaia catalog. So this whole machinery actually was pioneered by Brian Ossick who's now actually works at Microsoft and he gave me this catalog which the first thing that I did is obviously to plot it and basically this is what our accreted catalog looks like. Here is on the x-axis we have the radial velocity again and on the y-axis we have v5. Basically the disc of the Milky Way lives around here and you don't see it because these are the stars that are born in the Milky Way so they're not accreted and they should not show up. The one thing that needed to show up here no matter what we did was the Gaia Enceladus with the Gaia sausage and indeed you can just see it right like there is that extended structure here and positive and negative vr and which it has like more or less zero v5 and you can kind of throw a Gaussian mixture model at it to kind of pick it up and indeed that's the Gaia Enceladus as well as the Gaia solar halo. One other thing stood out for us is this structure which was not in the literature before it which means we got to name it and this is Nick's Great Goddess of the Night and it is a collection of stars that are in very funky strange orbits that we still are trying to understand. So these stars are basically kind of like this collection here at high radial velocity but also are somewhat prograde motion so if we wanted to compare it with the other disc stars so the stars in the Milky Way are rotating here at about 220 kilometers a second and they so they would be living or back here. The Nick stars also are kind of rotating with it but they have this radial velocity which means that they want to come into the galaxy or out. In terms of position this is x and y so this is the plane of the galaxy this is the Sun I exaggerated this arrow just for visual effects while the Nick stars are kind of moving that way and this is how it would turn so basically kind of it is rotating kind of in the same direction but it's also coming into the Milky Way and finally this is the x and z direction so this is perpendicular plane our galaxy would look like this and you can see that the stars the place that we picked up are actually extending a little bit further than the disc of the Milky Way so the question is what is Nick's is it evidence for emerging galaxy or is it the disc of the Milky Way that is has had some kind of perturbation the way that we could tell this apart these two scenarios apart is actually by observing Nick's with telescopes and trying to understand their chemical abundances to figure out whether or not they're consistent with the stars or with stars that have merged in so stars born in different galaxies the reason I was personally extremely excited about this is that if indeed Nick's is a merger it's going to look like this where something is falling and kind of like got swallowed in and it's co-rotating this has huge applications for dark matter direct detection because it means that dark matter also is kind of doing this co-rotation thing well so the rates of dark matter can be much much higher than expected however we just finished this paper studying the chemical abundances and the here on the x-axis we have the metallicity and on the y-axis we have that rotation in v5 so so in metallicity think of it as kind of like a bit of a probe for age so zero is the metallicity of the sun so these are very very young stars and then here these are very very old stars so the stars here the ones in gray in the background are basically the stars from the thick disk so these are the ones that we're comparing against you can see where this is going already where we ended up with a bunch of stars a bunch of nick stars that look very much like the disk and that was great we were like okay well I guess it is a disk perturbation but then we also ended up with these five stars that are extremely metaphor and would be very much consistent with emerging and with the merger interpretation so after about two years of work we ended up more confused than we started we were like okay what could do what could explain all of this and it's something that we're still interested in and trying to figure out is this a combination of both is it uh is there an explanation about the early formation of the Milky Way that's telling us about this so although it's very hard to tell we still don't know whether or not nicks is one thing multiple thing is it should not should it belong to our tree or not we don't know yet and it's something that we're still gonna be following through especially with cosmological simulations to understand whether or not we can create such an event so now the question is moving on to the second part is that I can take a look at what Gaia gives me and then can I see the different clumps of stars that are coming from different mergers and this is a very difficult question so the good thing about clustering is that it does not use any predetermined knowledge which means that I don't need to know anything about my simulations before so my simulations are not going to bias my results however it also means that I can't know much about the two answers so from a paper here a recent paper by Robin Naidu at all we see that we can kind of sort of clump up the different structures of the stars in this space where this is the total energy on the y-axis and on the x-axis we have l-z or basically the angular momentum in the z direction the rotation and in dynamics if you actually had a constant like a time independent potential objects that are merging in your potential are actually going to stay clumped up in this space however the monkey way potential is a time dependent potential so things kind of end up getting a little bit messier the question about clustering is that if you go on if you're so googling clustering you end up on this page called sklearn that has implementation for you know multiple many many clustering algorithms and the question is which one should I use well if we wanted to cluster your it really the clustering algorithm that you use depends on the data that you want so for example we look at this data right here it's fake data and your brain can already cluster it in quite different spaces right like you can already see that this is there is something in the clump in the middle they're like these very elongated wave shapes they're like sticks over here you can actually kind of do that the reason that you can do that is because this isn't 2d you can imagine if you're trying to cluster in three or more dimensions your brain is not going to be able to do it so you can use k-means clustering here on the left or you can use a different algorithm called the hdb scam and basically you can take a look at this and realize that okay that left one my brain would not have done that that is a little bit strange and while hdb scan you're like yeah that pretty much matches with what I had in mind so hdb scan is actually really good at using at dealing with data that has noise that has different sizes arbitrary shapes different densities in an unknown number of clusters so that's why that is the right test bed for us and indeed there was a paper by Kaylee Brower here at MIT who actually showed that although they do they all do absolutely they all a lot of clustering algorithms do not perform as well hdb scan is still the one that performs best so my student here Xiaowei O has taken the data from the early data release of Gaia so Gaia EDR3 and of course with some cuts and tried to figure out what are the clumps that are making this so the question here for our paper was not just like okay let's throw some clustering algorithm at it and see what happens but rather what are the absolute robust clusters which means what are the clusters that if we vary under uncertainties and keep re-sampling the ones that are always showing up because most of the time we can either see over densities not take into account uncertainties or we can cluster them and still not see those uncertainties right and these uncertainties and measurements are the difficulty in machine learning usage in physics compared to computer science because unfortunately between cats and dogs there are no error bars so Xiaowei basically did this in two iterations first he looked at the stars that no matter how much you re-sample them always are associated with not noise and these are the red stops here and then after that actually re-cluster them again to figure out which ones are stable and found these guys so this cool red one over here is actually a global cluster which is kind of annoying it is interesting for global cluster studies but as I've said before it doesn't tell me much about dark matter these structures here that we find actually multiple of are belong to the Gaia sausage however some of the literature say that that's not one merger that's multiple and we're finding it in multiple so we're trying to kind of reconcile all these different approaches to really understand the local structure of the monkey way so doing that actually requires that we look at the chemical abundances in future work we're trying to kind of put both the chemical abundances and the kinematics together to really try to identify these structures so definitely keep an eye out and the last thing that I'm going to discuss today briefly is going to be this stream finder so completely separately we can take a look at the sky and try to find over densities in streams and those are going to be able to tell us much about merging galaxies and especially the ones that are too faint or have are basically anomalous in different spaces that we can't release so this is Via Machina now it's pioneered by my collaborator David Chee and basically it's built on the anomaly detector that he actually used for LHC purposes in the beginning and you can take the positions not just the positions here but positions and proper motions and colors of these stars and put them through Via Machina to figure out what are the anomalous stars basically you're scoring the stars by how anomalous they are compared to the ones in the neighborhood and then you can actually see things that pop up right here like this one and this is actually a very easy to find stellar stream GD1 which I showed you in that earlier Sloan Digital Sky Survey paper picture and this is the old method this is the one that ours found and you can actually see that there are some structures and density variations along the stellar stream that we're able to also recreate with our streamfinder we're currently in hopefully in the next couple of weeks famous lost words are going to release the next iteration of Via Machina as well as about 80 new candidates of stellar streams I think about 70 of them are new which is pretty exciting and we are able to figure out what else is going on in the GD1 Sky putting this together how can we bring all of this back to dark matter and stars well this happens through direct detection so direct detection I just picked one experiment this is Xenon one ton dark matter comes in hits one of the Xenon particles and kind of goes out and that releases either electrons or scintillation light and that's how we know dark matter was there and I'm a very visual person and I like videos that's dark matter coming in hitting in a Xenon atom and showing like producing the scintillation light that is later created and that's how we know dark matter was there however the amount of scintillation light does depend on the phase space distribution of dark matter so the weights for us to figure out how much dark matter is supposed to be there depends on how fast and how dense dark matter is and although the density of dark matter has been a question that has been tackled for the past 100 years under one shape or another velocity distribution has had just some very major assumptions mostly the speed distribution of dark matter was assumed to be a Maxwell-Voltzmann distribution which is or the assumption of the standard halo model here shown in dash gray versus the distributions that we actually built based on the mergers so in the first part of the talk I talked about correlating dark matter and stars so we can take the velocity distribution of the halo the background dark matter halo that's here in red speed distribution as well as the substructure or the Gaia sausage that's here in blue and we put them together to end up with a speed distribution from luminous dark matter so basic of luminous satellites so this has some caveats to it so basically this is not the full distribution this is just the contribution of the large background as well as the Gaia sausage there will be smaller contributions from other things very locally which is something that we're currently working on but the distribution in general is very different from that standard halo model in particular at the tail here and the rate of direct detection actually is proportional to the density of dark matter as well as this integral of the velocity distribution and this f of v is basically kind of you're integrating from a particular minimum velocity here that depends on the dark matter threshold and the dark matter mass basically how energetic the dark matter has to be to kind of knock off the xenon above a certain threshold that's why you're integrating this part and you can see that if you integrate just this little part here if you for a very light dark matter it's an GEV dark matter then that you're integrating under this black curve versus the dash gray there will be quite a bit of a difference in the rate versus if you had a heavy dark matter mass for example 100 GEV and you're integrating this whole thing then these two distributions are normalized to one this space is almost the same so we are going to end up with more or less the same answer if however you have a co-rotating disk like a dark disk something nicks like distribution then that might actually affect your detection of heavy dark matter but not the light one so this is all to say that it is a very kind of tough problem it's not like an overall rescaling it really depends a lot of different pieces and that's why it's very important for us to really have kind of a global view all of these different structures and really use the astrophysics to our advantage to understand dark matter so hopefully what I've been telling you today is kind of like how to build a map of dark matter and we use very different tools as many as we could get our hands on from high resolution simulations to show up the correlations between dark matter and stars and then we use Gaia Dynamics to actually build and figure out the different star populations and how much do they tell us about different merging dark matter we use stellar spectroscopy to figure out the origin of different structures as well as machine learning because we have way too many stars which is again a really really cool problem to have and of course we're kind of putting all of this back into the particle physics part to figure out dark matter direct detection the story however does not end here our major goals are really to determine the particle nature of dark matter and be able to experimentally detect it mainly talked about the solar neighborhood and a little bit about the streams but we have a lot of outgoing work in correlating dark matter and stars and the galactic center and dwarf galaxies not exactly the same way so I'm going to use this opportunity to advertise some of the work of people here in my group Dr. Nora Ship actually produced this amazing comparative work where she compared the stream populations in the fire simulations to those that she mainly discovered in the Milky Way and we found that although the numbers are the same the orbital parameters of these solar streams are different and we're going to we're actually kind of currently following up on that and trying to figure out what does that really tell us about the dark matter distribution Honshu here is an undergraduate at the University of New Haven she's currently applying to grad school near you and she is building an created catalog of stars based on the new updated Gaia DR-3 Ian Roche is working on better understanding of the escape velocity of the Milky Way and how much of the mass of the Milky Way that actually tells us based on the new data from Gaia and finally last but not least Tree Ian is actually has this amazing work on using graph neural networks to obtain the density dark matter density profile of dark matter in simulations in dwarf galaxies and has shown that actually he has much better handle on the core cuts problem than the standard genes analysis so definitely definitely check it out so I'd like to finish by saying thank you so much for having me today and listening to me talk about all of the aspects of dark matter and how we can hunt with them for stars and I'm going to leave you with Nix the great goddess of the night thank you thank you very much Lena for this amazing talk it was really enjoyable and great to to see every all the progress that you and your collaborators have been doing with this data from Gaia put together with all this machine learning and particle physics so I'm going to open now the floor for questions those who are watching us please write your questions on the chat of YouTube and I'll be happy to read them to I think we lost Walter okay so I'm going to so I see Nicolas have a question thank you Lena for the amazing talk I do have some questions but I'll let Nicolas to start sure thank you thank you Lena super super nice talk enjoyed a lot so I have several questions let me start with my very naive one so you talk about quality cuts but you know my background is more colliders so for me you use cuts to what really to cut off the background from the scene of the wanting colliders but here I don't know exactly what they cut I mean what kind of cuts do you use that's a very good question so basically when we say quality cuts it's really about taking out the badly measured stars so the ones that I have not enough astromatic solution basically like Gaia is not really sure or it has like an artificially high error bars that are not appropriate there so yeah it's more about really kind of getting the data to like basically taking out various points it's not about kind of getting rid of the background in particular in kind of more numbers the Gaia DR3 dataset has about 33 million stars and when you do this various kind of quality cuts you're dropping down to 29 million so not losing much it's just about bad measurements okay but you were not talking about 8 billion or something like that stars sorry at the beginning at the beginning you were not talking about 8 billion stars or something like this oh 1.8 billion yes so 1.8 billion are the ones that we have 5d kinematics for unfortunately we have a very small subset small subset 33 million with a full DR3 that have 6d kinematics and a lot of the work that I'm showing you here actually relies on that sick dimension thank you thanks Alejandro thank you Lina thank you for the amazing talk I have several questions but I guess the one I'll ask is you showed three different methods and where do they coincide and if I did understand correctly nix was found in one of those right but then if I do this other do they where do they converge or should I see them as complementary or as different stories they're definitely complementary basically so the last one when we're trying to look for streams we're looking very very far away so that's based on 5d kinematics and we're looking at very far stars so basically even like the they don't this does not overlap with the first two methods like spatial selection so the first one was based on DR2 and it was very very local and that's what nix was found the second one however when we were doing the clustering we were trying we were trying to kind of basically it was one of the questions that we had where should we use the accreted catalog that we built in the first one or should we do something completely orthogonal just to kind of see whether or not things are matching and we ended up kind of using basically a different cut which is which is we're looking at the stars that have very high vertical motion and that kind of kills nix altogether so it does not show up because we wanted to see what are the other structures that we have so yeah you should think of them as kind of spatially covering slightly different spaces and trying to understand different things okay thank you also like at some point you mentioned like if I one of those immigrant stars but I love the term as an immigrant where does this number come from this one percent or like is it from the simulations or it's yeah so that's definitely from the simulations it is consistent with what we see if we look at the chemical abundances of the stars locally so that really tells us a little bit where it's coming from but the problem you can say okay why don't we just use chemical abundances and we can get all of them is that we have a very small subset of stars that we actually have chemical abundances for about like a few hundred thousand just because these are expensive and you end up you know pointing a telescope at these and even if with a bigger surveys versus Gaia has like orders of magnitude higher numbers so you need to figure out how to use the kinematics for that thank you and the last one before Walter kicks me out what makes is there a way to think I don't even know how to ask this question if it's well posed but then is the Milky Way especially in some sense or will we be able to I know Gaia well I don't know but I can assume that maybe all these kinematics of course can just be done for close stars close what I quote but then of course I cannot do this for a other galaxies but in some sense do we know if our galaxies is special is an average one or like all these that interesting stuff that you are learning from this Milky Way can be extrapolated to the whole universe in these cosmological scenarios that's an excellent question has been kind of like really hitting the field pretty hard especially because before we figured out the Gaia sausage we thought that oh that is very strange because in all our simulations there are a lot more murders and the Milky Way seems to be particularly quiet what happened there and then now we have a little bit more of a merger so we're like okay maybe we're a little bit less special and that kind of makes us sleep at night a little bit better but we've actually Mary-Angela DeSantiah-Princin and the student Dylan Folsom a student of hers at Princeton we're working on trying to identify basically just out of understanding the numbers of how many galaxies share this origin with Milky Way basically how many galaxies have an LMC and have a Gaia sausage merger in a big simulations of illustrious to kind of really figure out and get deeply into this question of how special is the Milky Way and you're right Gaia can only tell us about very local things inside our galaxy and it's very very hard to kind of get to the kinematics or star by star and outside galaxies thank you right before before Nicolás I need to read a question for you from the YouTube channel Douglas asks are there simulations that relax the assumption of collisionless dark matter yes that's a very good question so that's something that we're currently working on that has been some different groups have built simulations that have self-interacting dark matter and in dissipative ones we're more familiar with the ones in the fire group because that's something that we're working on and I did this very quickly so I didn't really publish it as a paper or anything but basically that correlation between the stars and the dark matter really falls apart when you start including self-interaction so that means that you the secret catalog that builds into the assumption really breaks down so that's something definitely that we're following through as we're building new simulations a larger suite of simulation that incorporates different dark matter physics so just a follow-up to that question so if it's supposed that dark matter we're self-interacting then we expect of course the dynamics for the dark matter component of the galaxies and when there is a measure for example to be a bit different but we should expect the stars to trace to track this motion so they don't track them anymore they don't because the because stars are collisionless and if dark matter starts self-interacting with itself it's going to land in different space spaces so then we're going to have to figure out a different way to to really so then you're going to use the stars as tracers like as tracers like as if the stars are sitting in the potential which is what we do now in dwarf galaxies versus getting the kinematics exactly of the stars to be matching the dark matter okay and you didn't mention before Nicolás I know you didn't mention the dark disk and the last part of your talk so I know that so this comes from I mean some some models where you have some self-interactions for dark matter and that so what is what is what you expect to see to set you know constraints or not on on the presence of this dark disk yeah so I think that that's the dark disk has two different meanings and basically depending who you're talking to so for the particle physicists that's basically yeah dissipative dark matter and kind of dark matter kind of falls falls down into like a component of dark matter basically has something very similar to the disk and then there is like a bit of a co-rotation as well but interestingly enough in Lambda CDM's Equalitarianist Dark Matter you in there are there have been some papers actually back in 2008 by Justin Reid and basically that says that it's because you have so many mergers in your galaxies in general and expectation a lot of them end up in like this prograde motion so you end up what we call a dark disk that's really just like a prograde something co-rotating dark matter co-rotating coming out of a merger and these two scenarios are different but they will still blow down to having dark matter that is co-rotating and that will affect your direct detection so basically we're trying to kind of go at it from you know these two different points of view where the particle physics dissipative dark matter models yes they have been talked about but we still haven't had like a good kind of cosmological simulation that takes into account all like both the dark matter physics as well as the bryonic physics to figure out exactly what happens and what are the implications for structure formation and also we're trying to understand like the component the co-rotating component coming out of mergers so either way I think we're going to land in some kind of co-rotating component but better having a better understanding of this is something that we're still working on thank you Nicolas thank you Walter so when you were talking about nicks and you show up plot the metallicity versus regular velocity and you show like these five like problematic points if you remember correctly will you please go back yes I'm trying to find here it is yep exactly so each of these cross are just stars yeah and this is nicks right so five five stars no sorry we so nicks the actual then actual number of stars very high confidence is about a hundred you can lower that confidence threshold arbitrarily but because you observe these stars with Magellan in Chile and with Keck and Hawaii so we needed to figure out exactly the stars that are going to be up in the sky in that part of the sky we landed on these 34 stars so these are 29 stars here in in red and these so we observed all of them and we didn't know where we're going to land until we did the chemical abundances results and then so like I was expecting I was well yeah I don't know what I was expecting but basically most of them ended up being in this higher mass like here so or sorry higher so higher metallicity here so we were like okay this kind of especially it's not just in this one but all the other elements kind of made it overlap with the gray in the background which is the disc of the Milky Way so we thought that they were like okay if we only stopped here we were like okay the nicks is nicks is just a perturbation of the thick disc because it has very funky kinematics the disc does not move that way but somehow disc stars ended up in these different orbits okay that's bizarre but okay but the weird thing is that out of our 34 stars five random stars ended up in this whole filling this whole other space that we just didn't really understand and they are kind of it's very difficult to reconcile these five stars with the big disc of the Milky Way so we're kind of we end up figuring out okay is it something that merged and then kicked some of the stars that's why like is there a scenario like that or it kind of it really requires us to do a lot more simulations to be able to figure it out okay but five events is like a fluctuation oh five out of 34 like if it's seven yeah if a seventh example is like this yeah in in Astro it's because yeah it's a pretty high number right because you still have to explain these like they they're really like it's not it's not bad measurement these are real stars like what are they doing yeah okay thanks I see I think okay I'm back it's hello okay any other questions all right if not let me thank again Dr. Lina Nesif for this great talk everyone who we're watching thank you for joining us today remember if you haven't done so to subscribe to our channel or to through Facebook, Twitter or to join our mailing list we hope to see you all in two weeks when we have our next webinar thank you again Dr. Nesif and good luck with the upcoming paper we are looking forward to see it thank you all right thank you