WEBVTT

1
330W: Dearly.

2
330W: Okay.

3
330W: Dustin.

4
330W: Victor

5
330W: Oh my god.

6
330W: Hi. How you doing? Good, how are y'all? Oh, good.

7
330W: For a couple weeks coming. Moving good.

8
330W: Yeah, there is a movie thing.

9
330W: He looks terrible, yeah.

10
330W: Have you been paid for 2,000 research editors? I'm meeting with her tomorrow at 11 o'clock.

11
330W: Screwed me today. Cool.

12
330W: Everyone, please.

13
330W: Work the craft.

14
330W: I think, through my efforts to do it, that'll be definitely nice. No, it's good.

15
330W: Other people tell me what's going on.

16
330W: Do you forgot students?

17
330W: I had quite a crowd. I graduated them, and now I'm down to one.

18
330W: We'll graduate next year.

19
330W: And, I thought I had more coming in.

20
330W: No.

21
330W: Take me too quickly. Like… I'm just working with them.

22
330W: I have one right now.

23
330W: Eric. Thank you to please us.

24
330W: Please, please just help me.

25
330W: Was my grad student for life.

26
330W: Interact perspective.

27
330W: Bye.

28
330W: I don't think it's a way to get a recording. I don't know if they report.

29
330W: Yeah. They do? This… your talk, yes. And recorded and posted. Posted. On the webpage. Yeah.

30
330W: You have to go teach? Yep. Okay.

31
330W: the sad overlap? Yeah.

32
330W: Glad I don't teach it like that, then. Yeah. It's weird. It's always something.

33
330W: It's slightly warmer in here than in hers.

34
330W: April.

35
330W: Rounding.

36
330W: Bye.

37
330W: How are you doing? Good, how are you all? I'm good. You are living with Alo and Potule, right? Yes, nice.

38
330W: I think that's good. Yeah, I mean, it's been…

39
330W: Three weeks and a half, three weeks ago.

40
330W: Are you taking all the courses, the co-courses? Yeah.

41
330W: Were the instructors for… The courses?

42
330W: It's, diagnostic for DNF.

43
330W: Oh, yeah. And then… God.

44
330W: Or… classical mechanics. I see.

45
330W: policy and a… So far, it's not bad. Okay. I mean… We haven't diverted.

46
330W: I'm insane enough.

47
330W: Brad. Okay. Yeah, at the end, I'm watching a while. Yeah, so… Like, unmanageable.

48
330W: And then fix later.

49
330W: Are you reading any courses, or… Yeah. Which one? Physics 123.

50
330W: That is what I heard. Half-life equals… Hughes wrote cookie, half-life equals 15… I know that goes back, and there's different versions of the cookies, I assume they go…

51
330W: I don't know where they're getting them. It's maybe… it's maybe the, like, out of labs of the worst scene, as in out of all the labs, I think that one of the worst. It takes the…

52
330W: For other labs, I do on 500 plus 6. The labs do just get on 2 plus 6. Maybe in 2 hours.

53
330W: one and a half. Yeah. And then this floor, so you kind of have to lock in for the first one.

54
330W: It's just… If I see you run out of… If they're like… So there are airtight skills.

55
330W: It's hard to, you know, the second law comes for us all.

56
330W: You know, diversification through scale.

57
330W: They're really the same.

58
330W: Well, look, you know, it could be worse, could be lettuce, you know. Yeah, could be lettuce. Gotta eat healthy in America now. Processed chips, and

59
330W: There we go. Yeah, it doesn't set something.

60
330W: Lilo, how are things going? They're good. Yeah? Okay. Do you finally have the quantum that hasn't met yet? Yeah. Had Quantum. Okay.

61
330W: Okay? It was okay. Good, yeah. Not lost yet.

62
330W: We have farm sick groups?

63
330W: Well, I think

64
330W: You can start on your own and compare notes later, that's the best way to do it, but that way you're never feeling like, oh, I just don't know how to say what this thing.

65
330W: Yeah.

66
330W: crap that you still want to get.

67
330W: Yeah. Maybe it changes things, but I was just suggesting group that could be used. Yeah. It works together towards the end. Definitely.

68
330W: In the plans, let's see.

69
330W: How's your start of the semester been? It's good, I was just telling Kareem. Hi, I'm Kareem. Hi, I'm Lila, nice to see you.

70
330W: Are you a first-year practice? Yes, and passwords.

71
330W: Sorry. Yeah, I'm teaching this science of science fiction.

72
330W: class in the Honors College. Oh, yeah.

73
330W: Yeah.

74
330W: I mean, they're on AI policy.

75
330W: Today, we're friends don't make a lot.

76
330W: Discussing the multiverse.

77
330W: I asked them, what does that mean to have free blood?

78
330W: Some days do not have a very well's education on mechanics.

79
330W: Question.

80
330W: That's all light.

81
330W: Not white, but came up with a quad experiment of how to test

82
330W: No, she says that was the discussion after 200 hours.

83
330W: Value test that you don't predict.

84
330W: Well, I claim that the human brain… Okay, good.

85
330W: Okay? Computers are repeatable.

86
330W: shouldn't, and these generally go.

87
330W: I don't know how to test for human race, but if I believe that we have an illusion of free will through plus beta…

88
330W: And their thought experiment was to take some of the parade, copy it completely, make it more promotional, and see what the coffee parade does.

89
330W: In comparison with the upgrades.

90
330W: Which I watch as a boxer.

91
330W: Okay, and of course, we decided we'd have to completely isolate

92
330W: From an outside student body is just being the same as a computer.

93
330W: And then, someone pointed out, like, she gets home. That's gonna say we have to complete the whole human body, not just the brain.

94
330W: And then someone said, well, we could submit the quantum states, too. There's a quantum computer. If you know how many bits you're asking for, forget it. But being as part of our recent dataset, it's not always just…

95
330W: Oh, I didn't say that. I think that's a possible quote.

96
330W: Yeah, and we're starting out with physics. Was it not there? It's really good so far. Yeah, it's gonna be IHS. So far, good.

97
330W: Short story, community bios.

98
330W: Okay, I have people here.

99
330W: There's a version where you can. There's a version where you too, so why not?

100
330W: have this be the crucial, because, like, it really doesn't matter.

101
330W: But do you know Ted Chan? No! Maybe I should read this. So, he writes short stories and stuff.

102
330W: Every single time. He wrote the short stories that I have.

103
330W: Oh, no, okay. He has the best short story on… The rival series. Good. Best short story on, like, the…

104
330W: Except for two of them, and I give you the Clarence.

105
330W: I don't know if that's do anything.

106
330W: And he said, it's like… The story's just gonna be 10. Eric!

107
330W: I'm sorry, Dr. Putney. It is.

108
330W: I know this is going to sound crazy, because we've never done this before. Telling people today what we…

109
330W: No, we go there every time. We usually do.

110
330W: Maybe that's why I'm like, I need a product. That's it, that's a good… I would take excuses in.

111
330W: Good. You guys said that?

112
330W: Oh, nice to meet you. Yeah.

113
330W: Yeah. It's crazy that they and Rachelor's treating a few customers.

114
330W: I've been doing there for 6 years.

115
330W: I saw that when I walked in. I was like, that's been there the whole time.

116
330W: Okay. And they were like, you know, you know… Anyway, I've been emailing.

117
330W: Yeah, I think he's already onto the computer, although he knows we have a battery.

118
330W: We've had those.

119
330W: attention, and you can hear it.

120
330W: Thank you.

121
330W: Okay. Thank you again.

122
330W: Alright. We'll get some cookies, and then suck a little bit.

123
330W: What kind of teaching in 10 points at one point.

124
330W: Yep. Yep.

125
330W: None.

126
330W: Originally, they… There will never be a problem. Jax, sure.

127
330W: Can you tell me more about it? It's, same as the other guy's stuff. It's the same guy.

128
330W: Is it over at Physics? If you look closely, you see bubbles, and you see strings… Oops, the string theory. Okay.

129
330W: It is art, so… No, he's got it framed. There's a few others that are hanging out in a large lab downstairs, opposite the business office. Wait, it's a lab upstairs. So I'm trying to…

130
330W: Bring them out. That's not a permanent.

131
330W: But there was a hook there. When you're buying one, just say a few words about next week, and then…

132
330W: They already have to take over.

133
330W: You'll sit right away.

134
330W: Zoom and just hop on. Yeah, I've… I can't see the chat or if they're all yelling that they can't hear us.

135
330W: I want to calculate a fraction of my time I spend logging through them.

136
330W: You are mobile.

137
330W: Less than 1%. Yeah. Did you change your password recently? I know. Oh, and then it was like… It was a dozen.

138
330W: I lost my phone one day. I mean, I was paralyzed. No work.

139
330W: Okay, well, welcome to this week's, physics and astronomy Colloquium.

140
330W: welcome both those who are actually here in person and those who are online. Before I hand over to Eric to introduce Matt's talk this week, let me draw your attention to next week's Colloquium, which moves from the cosmos to the

141
330W: to Terra Firma, and Sean Oh will tell us about how he develops a system for

142
330W: for engineering, quantum materials with topological properties. So that's next week, Shahno, and now let me hand over to Eric Gawaiso, who's going to introduce Matt.

143
00:00:00.850 --> 00:00:14.399
330W: Okay, I think we are setting a record for the most punctual Colloquium ever, so as people continue to filter in, both in the room and on Zoom, it's a great pleasure to choose today's Colloquium speaker, Professor Matt Buckley.

144
00:00:14.400 --> 00:00:22.190
330W: Matt is, as many of you know, a theoretical particle physicist who specializes in both cosmology and high-energy phenomenology.

145
00:00:22.240 --> 00:00:38.400
330W: A lot of Matt's recent work has applied novel machine learning algorithms to very large datasets, ranging from proton collisions at the Large Hatron Collider to the census of 2 billion stars in the Milky Way galaxy that the Gaia satellite dataset provided.

146
00:00:38.550 --> 00:00:44.409
330W: And that's been able to make a number of breakthroughs with these types of analyses of large public datasets.

147
00:00:44.430 --> 00:00:54.220
330W: To give you a bit of Matt's history, he got his PhD in 2008 from UC Berkeley. He was then the Dew Bridge Prize Postdoctoral Fellow at Caltech.

148
00:00:54.220 --> 00:01:05.090
330W: and the David Schramm Fellow at Fermilab. And then, after that, joined the Rutgers faculty. Back in 2021, he was promoted and tenured by us as an associate professor.

149
00:01:05.130 --> 00:01:12.309
330W: And has now received 3 different teaching awards from Rutgers, along with this year's… I want to get this right…

150
00:01:12.310 --> 00:01:34.310
330W: Provost Award for Excellence in Cross-Disciplinary Research. So, congratulations on that, Matt. And probably I could have just said at the outset here, Matt gives great talks. I'm sure you're going to enjoy this one. Thank you, thank you for that introduction. And, the cross-disciplinary research is that it turns out that they count astronomy and physics as different disciplines.

151
00:01:34.310 --> 00:01:50.690
330W: So I'm going to be telling you about some work that myself and the research group that I'm part of here has been doing for the last, you know, 5-6 years on looking for dark matter using these new techniques that machine learning allows us to do.

152
00:01:52.410 --> 00:02:00.940
330W: I study dark matter, as Eric said, my background, you know, is particle physics, so I started off thinking about looking for things at the Large Hadron Collider.

153
00:02:01.120 --> 00:02:03.529
330W: But over the years, you know.

154
00:02:03.710 --> 00:02:13.600
330W: I got interested in this new physics, and one of the questions you often get when you are studying dark matter is, like, well, do you really know that it's real? Like, how do I know that this is an actual problem?

155
00:02:13.960 --> 00:02:17.820
330W: And there's lots of evidence for dark matter.

156
00:02:18.030 --> 00:02:22.379
330W: And all of that evidence for dark matter comes from astronomy and comes from astrophysics.

157
00:02:22.380 --> 00:02:42.890
330W: Right, so one of the sort of best-known pieces is what I'm showing here, Vera Rubin's work on rotation curves. This is a measure of the speed of stars as a function of radius in a galaxy, much like the Milky Way, a spiral galaxy. And you can see as you go further and further out, the stars kind of flatten out into this rotation curve at a more or less constant velocity.

158
00:02:42.890 --> 00:02:49.690
330W: And then you do a bunch of work, and you say, like, well, where's all the gas? Where are all the stars? How much do they weigh? And you discover that

159
00:02:49.950 --> 00:03:01.019
330W: the gas and the stars, just don't have enough mass to support this rotation curve. The galaxy should fly apart if you don't have dark matter. And so you can fill in just this missing matter, dark matter.

160
00:03:02.650 --> 00:03:19.179
330W: These days, you can actually go and take a picture of dark matter. It is a thing that it doesn't interact with light, but what you can do is you can look at gravitational lensing. And so this is a very particular system, it's called the Bullet Cluster. What is going on is that it is a cluster of galaxies, so…

161
00:03:19.420 --> 00:03:31.169
330W: One cluster of galaxies came in from the left, one came in from the right, they smashed into each other. Most of the normal matter in a galaxy cluster is in hot gas. It's not in the stars, even.

162
00:03:31.780 --> 00:03:46.600
330W: Hot gas, because it has a long-range force, electromagnetism, stops, right? There's, like, they scatter and they slow down. And that forms this kind of ram pressure bullet right here, that's the name. So in the red, we see the X-rays, that's where all the…

163
00:03:46.970 --> 00:03:58.769
330W: all the visible mass of the cluster is, and then you do gravitational densing, and you look at the light coming from behind it and how it is bent, and you discover that the mass is here and here in blue.

164
00:03:58.950 --> 00:04:14.870
330W: And it's not where the visible matter is. And what happened is just two clouds of dark matter passed through each other, the gas stopped, the dark matter kept going, right? So most of the universe, you know, in the galaxy clusters is just made out of a thing that does not care about electromagnetism.

165
00:04:15.940 --> 00:04:18.710
330W: The best evidence for dark matter, though.

166
00:04:18.940 --> 00:04:23.699
330W: This is a picture of the cosmic microwave background taken by the Planck satellite.

167
00:04:23.780 --> 00:04:39.630
330W: What I'm showing is the density… sorry, the temperature contrast, right? If you just kind of look at the CMV, it will be totally uniform across the sky, but if you look very carefully, you'll see one part in 10,000, or sorry, 100,000 fluctuations, where it's hotter or colder.

168
00:04:39.840 --> 00:04:42.249
330W: And that forms a very particular pattern.

169
00:04:43.110 --> 00:04:49.210
330W: You can study that pattern. What you discover is that you're looking at a classical field theory of

170
00:04:49.410 --> 00:05:02.990
330W: over-densities of photons and baryons, so protons and electrons in this case, and dark matter, and they're all kind of sloshing around. There are sound waves echoing through the universe in the photon-electron plasma.

171
00:05:03.260 --> 00:05:05.599
330W: The way those sound waves move.

172
00:05:05.810 --> 00:05:08.660
330W: It's going to be affected by the gravity underneath them.

173
00:05:09.060 --> 00:05:11.180
330W: Right, and also, you know, that…

174
00:05:11.630 --> 00:05:17.129
330W: And also by the number of electrons, the density of electrons in different places.

175
00:05:17.910 --> 00:05:24.300
330W: That pattern can be turned into a power spectrum, and it has these sort of very particular, you know, peaks right here.

176
00:05:24.440 --> 00:05:30.250
330W: If you increase the amount of baryons in the universe, you will push this peak up and that peak down.

177
00:05:30.410 --> 00:05:47.319
330W: Relative to each other. Whereas dark matter will have a different effect on this detailed structure. And so it's a complicated set of analyses, but the end result is this structure here is completely inconsistent with all of the matter in the universe being made out of the same stuff that we are, right? It cannot be…

178
00:05:47.430 --> 00:05:56.259
330W: baryonic matter, it cannot interact with electromagnetism, and so it passes right through us, right, whatever it is. So this is some of the best evidence for dark matter

179
00:05:56.510 --> 00:05:59.360
330W: Even though it is the one that is sort of the hardest to explain.

180
00:05:59.680 --> 00:06:10.859
330W: These sound waves eventually stop, because the universe cools, and the universe becomes electrically neutral, the plasma stops being a plasma, the electrons combine with the protons.

181
00:06:11.560 --> 00:06:16.429
330W: And you bake in a very particular length scale, which basically corresponds to around this first peak.

182
00:06:16.880 --> 00:06:31.389
330W: On that length scale, there is going to be a slightly larger number of protons than there were at other length scales, which means if you wait several billion years, you're going to find more galaxies separated by a particular length scale than on any other length scale.

183
00:06:31.550 --> 00:06:43.729
330W: And how that appears in the universe today will depend on the amount of dark matter in the early universe and the amount of dark matter between, you know, us and that distant object. What I'm showing you here is simulation.

184
00:06:44.000 --> 00:06:55.819
330W: This is a simulation of large-scale structure. Every black dot you see here is a simulated galaxy. And you can see there's kind of this classic spiderweb pattern across the sky, the large-scale structure.

185
00:06:55.820 --> 00:07:04.380
330W: If you took, the two-point correlation function, you would discover a peak at a very particular length scale. This is a peak of baryon acoustic oscillation.

186
00:07:04.500 --> 00:07:14.310
330W: That length scale is telling you something about the evolution of the universe, both during the time of the CMV and since then, and this is consistent with there being dark matter in the universe.

187
00:07:14.820 --> 00:07:25.100
330W: Right? So these are all the pieces that really tell us that there is a thing out there that does not interact with the standard model forces in the way that we're used to, and so you need new physics.

188
00:07:25.140 --> 00:07:41.800
330W: Question? Yeah, go ahead. What are the units on the X scale? Oh, over here? Sorry. This is the multipole, so you do, like, it's on the sky, so you do a, spherical harmonic decomposition, just L of the spherical harmonic. Yeah, so it…

189
00:07:41.800 --> 00:07:45.589
330W: can be turned into an angle. This is about a degree.

190
00:07:45.920 --> 00:08:00.019
330W: Question? Yeah, so the spectra that are extracted here are model independent? The spectra… so if you looked really carefully, and I apologize, it's hard to see here, there's data. So that's just the data. The fit requires model…

191
00:08:00.110 --> 00:08:02.949
330W: assumptions. There is a basic…

192
00:08:03.010 --> 00:08:20.890
330W: the one that works the best is called Lambda CDM, dark energy plus cold dark matter, right? I'm going to actually return to that, this is from a recent paper where we fiddle around with the model. But yes, you do have to make some assumptions, but this… to really fit it, but the assumption that seems to work requires something like cold dark matter.

193
00:08:24.980 --> 00:08:38.600
330W: So, the fact that dark matter seems to exist, and we as theorists come up with lots and lots of great ideas about what it could be, has motivated, a very extensive particle physics experimental program to go and look for this thing.

194
00:08:38.730 --> 00:08:56.549
330W: Right? And there's lots of ways you can do it. For one thing, you can take, like, the LHC, and if dark matter happens to interact with us just a little bit, when you smash protons together, occasionally you'll create dark matter. And you can't see the dark matter, but if the dark matter shoots this way, and something visible goes that way, you can see sort of missing energy.

195
00:08:56.550 --> 00:09:07.590
330W: And so you can look for missing energy in the… at the LHC. This is an extremely difficult search, because if anything goes wrong, that looks like missing energy. But what I'm showing here is just one particular model

196
00:09:07.860 --> 00:09:09.559
330W: And an exclusion plot.

197
00:09:10.750 --> 00:09:22.679
330W: Dark matter could be something called axions, in which case you can build very, very sensitive sort of radio detectors. You can look for dark matter kind of moving through your detector. This is ADMX, and again, what I'm showing here is an exclusion plot.

198
00:09:22.750 --> 00:09:39.220
330W: Usually I show an exclusion plot from the large direct detection experiments that are deep underground, where you just take, like, a very large bat of something that is mostly non-radioactive, bury it very deep so you don't have cosmic rays, and you look, and you look, and you look, and we haven't found anything.

199
00:09:39.380 --> 00:09:45.620
330W: I am going to point out that this is the LZ experiment, this is the result from two weeks ago, and there's one

200
00:09:45.830 --> 00:09:52.730
330W: scatter in a 7-ton tank of art… of helium… Sorry, of xenon.

201
00:09:53.050 --> 00:10:14.509
330W: This is sitting right at a point that you might expect dark matter interactions to look like. It's one event, so it is deep learning on one event, and I will be spending most of my talk about deep learning on many, many events. I don't know what this is, it's been very interesting, but our group has had a lot of fun, or I've had a lot of fun, I don't know if they have, trying to figure out…

202
00:10:14.510 --> 00:10:19.240
330W: is this… Possibly dark matter? The answer is, maybe.

203
00:10:19.620 --> 00:10:37.020
330W: The other thing I'll say is you can look for dark matter, out in the universe today, annihilating still, right? Whatever it is, it's mostly stable, it's lasted for most of the history of the universe, but occasionally one piece of dark matter could hit another one, it could annihilate and maybe give you something you can look for, like gamma rays.

204
00:10:37.480 --> 00:10:55.710
330W: And one way to read this plot is that this is an exclusion plot where we've looked at dwarf galaxies, and everything above this, the larger cross-sections as a function of mass, have been ruled out. The other thing to notice is that there is a signal from the center of the galaxy, of gamma rays, which I will be talking a lot more about.

205
00:10:55.710 --> 00:11:03.340
330W: We don't know if it's dark matter, but if you want to interpret it that way, that's where it would sit, and then you could argue about whether it's excluded by other searches.

206
00:11:04.580 --> 00:11:12.310
330W: But, as you were kind of asking, like, is, you know, are these constraints from the CMV and things like that, are they model independent? And their answer is, they're not.

207
00:11:12.500 --> 00:11:32.260
330W: Astrophysics can do more than just tell you that dark matter exists. It can tell you things about the physics of dark matter, the particle physics of dark matter, the things that I, sort of, started off being really interested in. As an example of that, this is a simulation of the cosmic web, sorry, maybe you can see it over there better.

208
00:11:32.340 --> 00:11:47.420
330W: And you can see, this is CDM, so cold, dark matter, you can see these kind of over-densities separated by big voids, and smaller and smaller overdensities next to each other. The power spectrum, the number of large galaxies and small galaxies, that is a prediction

209
00:11:47.660 --> 00:11:49.439
330W: Of a model of dark matter.

210
00:11:49.610 --> 00:11:56.679
330W: If you make your dark matter not what is called cold, but warm, if you look over there, you should start seeing things kind of smear out a little bit.

211
00:11:56.860 --> 00:12:08.160
330W: Right? And what that tells you is if you could see the really small halos of dark matter, you could tell the difference between the particle physics that created this and the particle physics that created that.

212
00:12:08.330 --> 00:12:11.510
330W: Other things you can do, so this was a paper,

213
00:12:11.710 --> 00:12:19.980
330W: myself and some other colleagues, if you have, like, dark matter that has its own long-range force, like dark radiation.

214
00:12:20.160 --> 00:12:20.970
330W: Right?

215
00:12:21.130 --> 00:12:33.280
330W: You're going to get the kind of acoustic oscillations in the early universe that you see in the baryons, and that will change what this is the power spectrum, the number of halos that you're going to get as a function of wave number, so 1 over weight.

216
00:12:33.280 --> 00:12:43.049
330W: small objects, big objects. And you could start perturbing the number density over here, and if you could go and see those, you could tell the difference between cold, dark matter and your particular favorite model.

217
00:12:43.800 --> 00:13:08.690
330W: Other things that can happen is that if dark matter scatters against itself, so bangs against itself, which I'm never going to see in the lab here, because my lab is not made of dark matter, right, if you look at the dense centers of galaxies, or the dwarf galaxies in particular, you might see that the density, sort of, instead of going like this in the standard scenario, starts either, you know, getting more dense or less dense as you turn these knobs

218
00:13:08.690 --> 00:13:09.989
330W: comes into particle physics.

219
00:13:10.600 --> 00:13:19.130
330W: So this is something that's really interesting, like, there's all these ways that I, as a particle physicist, could learn about dark matter, not by building an experiment here.

220
00:13:19.330 --> 00:13:22.390
330W: But by looking out in the universe and understanding what's going on there.

221
00:13:22.640 --> 00:13:25.670
330W: This was a realization a number of us were kind of

222
00:13:25.850 --> 00:13:40.269
330W: inching our way towards, over time, and in 2017, myself and Annika wrote a very long paper trying to condense all this down and kind of create a picture on our own head, all of the different ways that we would be able to look for these things.

223
00:13:40.270 --> 00:13:47.490
330W: And what we're showing here is just, like, this is a sort of a measure of how big the cluster of dark matter

224
00:13:47.510 --> 00:13:53.059
330W: you're looking at, so remember I showed those power spectra. There's different structures at different length scales.

225
00:13:53.660 --> 00:14:00.389
330W: the Milky Way is, like, here, the Large Magellanic Cloud is here, dwarf galaxies are here.

226
00:14:00.890 --> 00:14:21.059
330W: there's a lot of very interesting physics you could learn if you could see the smaller objects. And so we started just collecting up, oh, these are all of the things that the astronomers and the other… and the physicists are doing in the next, sort of, decade or two that we could maybe use to try to probe those smaller and smaller scales, with the goal of then going back and saying, this tells you something about the physics.

227
00:14:22.320 --> 00:14:39.650
330W: This is a wonderful time to try to be doing that, because there's a tremendous amount of data that we either have, or we're going to have soon, right? This is the Fermi gamma Ray Telescope, this is a picture of the sky in gamma rays, this is the galactic disk, so this should wrap around your head like your head is in the middle of a globe.

228
00:14:39.650 --> 00:14:42.920
330W: Right? That's the galactic center, that would be the anti-center.

229
00:14:43.020 --> 00:14:48.920
330W: And there's all sorts of complicated astrophysics going on there, but maybe there's dark matter annihilating in there, too.

230
00:14:49.290 --> 00:15:09.209
330W: This is a picture of the Milky Way from the Gaia telescope. I'm going to talk a lot more about that coming up soon. Right, but then there's these studies of large-scale structure, which, remember, tell us things about the early universe and how dark matter is maybe moving in that. And Vera Rubin and LSST is coming up, DES and DESI sort of exist now, and they have

231
00:15:09.620 --> 00:15:18.729
330W: you know, this is the coverage on the sky, this is the coverage from Jorubin, and then this is just sort of a picture out in cosmic history from DESI of all the structure that you could see.

232
00:15:20.020 --> 00:15:21.780
330W: Do you have all of this data?

233
00:15:22.300 --> 00:15:40.450
330W: And we have machine learning, which is eating the world, right? So, machine learning, for physicists, I think the really interesting thing is that you don't go and ask the AI, like, what is dark matter? You use the fact that there are these very complicated, very interesting algorithms that give you new ways to look at data sets.

234
00:15:40.450 --> 00:15:47.130
330W: And new ways to kind of look at the data that implements physics at a very deep level, and I'll explain what I mean there.

235
00:15:47.130 --> 00:16:02.769
330W: So what I'm going to do is I'm going to talk in some detail about two of the things that we've been working on using machine learning in large datasets to study dark matter, one of which is mapping dark matter in the Milky Way, and the other is looking at this galactic center gamma ray access.

236
00:16:03.190 --> 00:16:04.990
330W: And then I'll end with…

237
00:16:05.190 --> 00:16:12.659
330W: some work that we're kind of… I'm starting off in interesting new directions, then I'll connect back to, like, actual pure theory, rather than just looking at data.

238
00:16:13.840 --> 00:16:27.569
330W: So the Gaia Space Telescope was a mission, this is an artist's rendition of the telescope, it just sat up in space and kind of pivoted around and looked at stars over and over and over and over again.

239
00:16:27.700 --> 00:16:39.380
330W: This is a picture of the galaxy, but if you zoomed in very carefully, this is not a snapshot. It is a map of where every single star that Gaia saw is. Individual points, right?

240
00:16:39.580 --> 00:16:48.690
330W: It measured about 1.5 billion stars, that is 1% of all the stars in the galaxy, so that sounds like a lot, but it's not really all of them.

241
00:16:48.690 --> 00:17:00.520
330W: And Gaia is so good at measuring stars that it can tell where a star is, and then when it comes back around, it can tell if the star has moved across the plane of the sky, which is a tremendously difficult thing to do.

242
00:17:00.710 --> 00:17:15.429
330W: It can tell you the distance, because it measures parallax, like how much a star seems to have moved as, you know, we orbit the sun. And so, combined with these things, you can get a map of stars, of where they are and how they're moving in three-dimensional space.

243
00:17:15.540 --> 00:17:27.319
330W: It can only get the line-of-sight motion for about 34 million stars. Sounds like a lot, a lot fewer than a billion. That will go up a lot in a couple months, and we're very excited about that for reasons we'll talk about.

244
00:17:27.500 --> 00:17:34.970
330W: And so the question that we were asking ourselves is, what can you do with millions of stars with their position and their velocity?

245
00:17:35.910 --> 00:17:41.090
330W: And one of the things you can do is you can try to build a map of the gravity in the galaxy.

246
00:17:41.270 --> 00:17:58.070
330W: you can think of these stars moving through the galaxy as basically a cloud of gas, right? They are tracers, and the way that they distribute themselves in position and velocity space has to do with the force they're feeling, and the force they're feeling is gravity.

247
00:17:58.150 --> 00:18:04.730
330W: So you can write down just the Boltzmann equation, which is just saying, like, if the whole system is in equilibrium.

248
00:18:04.910 --> 00:18:11.709
330W: right, then DFDT, this is the phase space density. How likely is it to find a star in position and velocity

249
00:18:11.950 --> 00:18:14.689
330W: Well, if it's an equilibrium, it doesn't change with time.

250
00:18:14.920 --> 00:18:22.309
330W: And it will only change in position space and velocity space in this very particular way, related by the acceleration, which is given by graphic.

251
00:18:22.770 --> 00:18:34.439
330W: So if you knew F, and you knew the derivatives of F with respect to position and velocity, you would learn the acceleration due to gravity. You'd get a gravitational map of the entire galaxy, or at least where you had data.

252
00:18:35.220 --> 00:18:37.290
330W: This has been known for a very long time.

253
00:18:37.650 --> 00:18:50.420
330W: It's very hard to do, and so traditionally, what you do is you take moments of it, you get what's called the genes equation, and this is an example of a relatively recent genes analysis, and the problem is that you have to kind of… you don't have enough data.

254
00:18:50.810 --> 00:18:57.799
330W: So you have to take these big blocks of stars and say, I'm going to figure out, kind of, the derivative of the…

255
00:18:57.800 --> 00:19:10.970
330W: integral of a phase-based density from here to here to here to here, and kind of do the derivatives in blocks. And you have to take very large blocks of stars in order to get a number that doesn't kind of blow up, because derivatives are hard to take.

256
00:19:11.310 --> 00:19:12.260
330W: Right?

257
00:19:12.390 --> 00:19:17.000
330W: What we're going to do is say, This is a loss function.

258
00:19:17.270 --> 00:19:32.079
330W: I can bring this over to the other side, and something should be equal to zero if you have the right acceleration. So I'm just going to say there's a neural network that's going to spit out phi. You take the derivative of that neural network, that should satisfy this expression.

259
00:19:32.610 --> 00:19:38.289
330W: Right? So you build a term, it's just a… you build a neural network, it has a loss term that's just…

260
00:19:38.890 --> 00:19:42.469
330W: the Boltzmann equation, right? So it knows about physics.

261
00:19:42.650 --> 00:19:53.370
330W: Right? And that seems like a great thing to do, but I require something there. I need to know what F is. And I need to know the gradients of F. I need to know its derivative with respect to position and velocity.

262
00:19:54.390 --> 00:20:10.839
330W: And that's okay, there are new machine learning algorithms called normalizing flows. These are some animations that my student and now postdoc Eric made for me. This is what the data looks like in position X and Z, so X is away from the galactic center, Z is off the plane.

263
00:20:11.120 --> 00:20:15.979
330W: And what a normalizing flow does is it says, I have no idea what this data is, I know nothing.

264
00:20:16.110 --> 00:20:23.480
330W: But I know that I'm supposed to spit out something that minimizes a particular loss term, and the loss term you pick is the entropy.

265
00:20:23.850 --> 00:20:25.879
330W: And you just start with an analytic.

266
00:20:26.100 --> 00:20:36.680
330W: phase-based density, Gaussians usually, and you just learn the transformations that take you from that initial thing to the final thing. And the end result is you have learned F,

267
00:20:36.930 --> 00:20:49.259
330W: as a function of as many variables as you want, right, and you've learned its gradients. And so, if you're following this animation, what it should do is start with a blur, and by the end of all the transformations, it looks like this.

268
00:20:49.470 --> 00:20:55.840
330W: Even though it doesn't know what it's supposed… all it knows is that I want to minimize the entropy of the system.

269
00:20:56.050 --> 00:21:11.739
330W: So we use a particular state of flows called Mast Autoregressive flows from 2017, which means if you talk to a machine learning person, they're like, oh my god, you are working in the dark ages, right? We have looked at more recent ones, they're useful for certain things, but this turns out to be really, really useful.

270
00:21:12.210 --> 00:21:22.900
330W: So I will say, if you are, you know, interested, if you need to know a face-based density of data, and you like it better than you had any right to expect you could, these are incredibly useful tools.

271
00:21:24.280 --> 00:21:34.190
330W: So with that, we have a phase-based density, we have its derivatives, we can do what I just said, you throw it into the Boltzmann equation, and to die, you get an acceleration, you get the gravity of the galaxy.

272
00:21:34.320 --> 00:21:35.750
330W: There is a problem.

273
00:21:36.210 --> 00:21:52.839
330W: The galactic disk has dust in it. If you've seen the Milky Way at night, there are all these beautiful black patches in it, the dark spots where there's no stars. This is what our data actually looks like. If you kind of look down into the plane of the galaxy, the galactic center is over here, and you can see there's this striations here.

274
00:21:52.840 --> 00:22:03.269
330W: And there's a cloud of dust somewhere, it's blocking all the stars behind you. The Boltzmann equation doesn't know about dust, and shouldn't work with dust, right? It's just missing data.

275
00:22:03.660 --> 00:22:07.520
330W: Well, what we realized is that we could just add more machine learning.

276
00:22:08.930 --> 00:22:15.620
330W: There's a thing we've measured with our normalizing flows, Fobserved. That's the phase-based sensitive you've seen that has dust in it.

277
00:22:15.850 --> 00:22:26.490
330W: There is a true phase-based density that obeys the Boltzmann equation, and they are related by an efficiency factor that is caused by dust that only depends on position. Is there dust in the way or not?

278
00:22:27.360 --> 00:22:41.300
330W: I have no idea what this is, we are physicists, we resolutely know nothing about dust, right? We are not astronomers at heart. But, if you just take this and stick it into the Boltzmann equation, you get a more complicated version of the Boltzmann equation, where that is something you can measure.

279
00:22:41.600 --> 00:22:52.070
330W: And then you just say, that's a neural network, the thing I want to learn? That's a neural network, the thing I don't really care about, because I'm a physicist and I don't care about dust. I'm going to learn both of them at the same time.

280
00:22:52.440 --> 00:22:58.220
330W: Right? And so you end up with what I will call the magic dust eraser, and you get a slightly more complicated,

281
00:22:58.710 --> 00:23:02.160
330W: Lost term, but you can… You can trade it.

282
00:23:02.550 --> 00:23:19.590
330W: This is what the, you know, the phase-based density of our actual Gaia data looks like. You can see all the dust lanes in it. You apply the magic dust eraser, and that's what the actual stars in the galaxy look like, which is more or less the right answer, right? This is the efficiency function that we get. This is, like, how much you're missing in the data.

283
00:23:19.840 --> 00:23:32.449
330W: This is our mock-up, where we actually had to learn some things about dust, and we kind of learned what we should have seen, and you can see the structure is more or less right. It's a little blurrier, but we have just learned… again, so if you…

284
00:23:32.560 --> 00:23:39.430
330W: Things astronomers want a 3D efficiency map of dust in the galaxy? We have one.

285
00:23:39.570 --> 00:23:42.900
330W: Maybe we do, actually. Isn't there, like…

286
00:23:43.520 --> 00:23:56.380
330W: like, the NED database? Have you mapped it to the NED database? This is… I forget what L… what the first author on L here is. It's the… this is a particular efficient… this is…

287
00:23:56.550 --> 00:24:10.569
330W: This is… this is when we learned, this is… we… we downloaded the best 3D dust efficiency we could get and integrated out, to kind of mock up what we should have seen. Gotcha. Right? But we… the network never saw this.

288
00:24:10.840 --> 00:24:16.009
330W: Right, this is just a pure, like, hey, did we do anything remotely correct? And the answer, I think, is yes.

289
00:24:17.570 --> 00:24:23.540
330W: I will say that what I'm calling the efficiency is not what you're thinking of as a dust map. We have to integrate it.

290
00:24:24.000 --> 00:24:34.849
330W: Okay, so I've done all that work, now I want to do some physics with it, so you can learn the potential. You can compare that potential with, like, models, and we are getting more or less the right shape.

291
00:24:34.920 --> 00:24:53.799
330W: There's… this is the potential… I'm sorry, this is the acceleration as you look in radius and off the galactic plane. This should be zero. This is looking down on the plane. There should be no acceleration sideways in the plane. We get a little bit. You can see maybe there is a little bit of problem towards the galactic center. And then…

292
00:24:54.200 --> 00:25:04.290
330W: you know, this is the Z acceleration, kind of looking sideways, and so we're mapping things that look fairly close to what we should expect, and, you know, there are errors associated with all of these.

293
00:25:04.780 --> 00:25:24.230
330W: It's a neural network. I've learned phi. You can just take gradients of phi, right? You can do backpropagation through the network, so you can just calculate the second derivative of it, and you can get a map of the density. This is a map of the density in a 4 kiloparsec sphere surrounding us. The galactic center is over here.

294
00:25:24.330 --> 00:25:25.729
330W: Right?

295
00:25:25.860 --> 00:25:33.550
330W: You can see we have density rising towards the center, that's as it should be. There's more stars and dark matter over there, and falling as you go off of here.

296
00:25:33.700 --> 00:25:38.460
330W: This, because I'm only showing you the average, looks a lot more wobbly than it,

297
00:25:39.000 --> 00:25:56.940
330W: should. There are error bars associated with all of these measurements, and if you look at the errors, you know, you could see that it would have sort of the shape that you'd expect. But we do go through, and I want to emphasize that we deal with correlations, I'm emphasizing that because I think it took, like, 3 months of Eric's life.

298
00:25:57.090 --> 00:26:00.910
330W: But we fully have, you know, full error model and correlations all the way through.

299
00:26:01.700 --> 00:26:13.679
330W: One thing you could do is, like, this is sort of where we're standing in the galaxy, and then you're just moving above the galactic plane and below the galactic plane. Where we're standing, there's lots of stars.

300
00:26:13.820 --> 00:26:19.510
330W: So there's a model of the star density, the stellar varyonic density here, that's the blue.

301
00:26:21.170 --> 00:26:25.159
330W: Data-driven model is the total density, that is the gray.

302
00:26:25.260 --> 00:26:30.349
330W: If you subtract those off, you discover that we are consistent with a

303
00:26:30.790 --> 00:26:44.489
330W: you know, uniform density at a set radius of a thing that is not being traced by the stars, right? The average of that you can interpret as the average density of dark matter at our galactic radius. So this is a map of dark matter.

304
00:26:44.760 --> 00:27:00.009
330W: You can try going in towards the galactic center. I would… as a particle physicist, I'm very interested in what the dark matter is over here. Right now, we can get about 4 kiloparsecs towards the galactic center. The galactic center is about 8 kiloparsecs away, so we're not there yet.

305
00:27:00.390 --> 00:27:01.300
330W: But…

306
00:27:03.340 --> 00:27:13.390
330W: Coming this December, we're going to get a new data release. This is not the Milky Way, this is a simulation of the Milky Way, but the… this is more or less where the Sun is.

307
00:27:13.540 --> 00:27:18.720
330W: Relative to the galactic center, this is the size of the ball that we've just mapped at.

308
00:27:19.560 --> 00:27:26.710
330W: With the new data release, we estimate we should be able to get about to the red line here.

309
00:27:26.920 --> 00:27:41.409
330W: Which will cover the galactic center. Now, I don't believe I'm actually going to be able to tell you what's going on here, because there's just going to be too much dust, and our dust eraser is going to fail, but if I can tell you how much dark matter there is here and here, that's very, very useful, for reasons I'll explain shortly.

310
00:27:42.020 --> 00:27:50.959
330W: I know this is a model, but your red circle goes past the bulge onto the other side? Yeah. Is that… Yeah, you should… you should put a, like…

311
00:27:51.050 --> 00:28:02.680
330W: a mask here, where it's just… in principle, I could see a star there. In practice, I'm not going to, because the dust there is just going to be a disaster. One of the things we are working on right now is trying to build

312
00:28:02.780 --> 00:28:09.479
330W: Better synthetic models to understand how badly we're going to do, so we're not surprised when we actually get the data.

313
00:28:11.270 --> 00:28:21.230
330W: So, you know, just using what we have already, we should be able to reach the galactic center, or at least radii compatible with the galactic Center.

314
00:28:21.670 --> 00:28:24.530
330W: There's other things you can do,

315
00:28:25.240 --> 00:28:44.979
330W: You can use neural networks to kind of in-paint missing data, right? There's 1.5 billion stars. Only a small fraction of them have radial velocities associated with them, right? This is work that, Eric is doing right now. This is a sample of stars where we actually have the full data.

316
00:28:44.980 --> 00:28:49.020
330W: Right? And then you train a neural network,

317
00:28:49.930 --> 00:29:00.700
330W: to kind of in-paint for the missing ones, the values that you're, like, you're not seeing, right? And this is what you get out of it, right? So it never… this is…

318
00:29:00.700 --> 00:29:15.370
330W: in painted data where there was no measurement, right, but you can see you're getting similar structures in here. Now, I don't want to use this to just say, like, oh, I've made a fake measurement, that must be true, but if you're doing this correctly, you also get uncertainties.

319
00:29:15.540 --> 00:29:31.509
330W: Right? So you have an estimate plus errors, and you can take that, you can put it into your own neural network, and if you're doing it correctly, when you, you know, when you get a measurement out of it, you'll get uncertainties along with it. Where does the data not support your knowledge?

320
00:29:32.290 --> 00:29:51.449
330W: Is this an uncertainty on how well the neural net is actually predicting what these velocities are? Yeah. You can also put in, like, measurement errors on the stuff that you see. Okay. Cool. Yeah. Is there, like, a quick explanation of how those are estimated? There isn't, that's okay. You can talk afterwards, yeah.

321
00:29:51.890 --> 00:30:12.079
330W: Other things you can do, totally different work. You can look for anomalies in the Gaia data, you can use that to find stellar streams. We discovered two new stream complexes, which we named the Passaic and the Raritan, after discovering that the largest burger in New Jersey was named the Delaware, and we're like, well, we're not doing that.

322
00:30:12.650 --> 00:30:28.459
330W: So, using similar techniques to that, we talked to Christy McQuinn, and we used DESI data, and she found some dwarf galaxies, so that's Leo K. There's a dwarf galaxy there. I can't see it, but I don't have a strong MRIs. That's why I needed a neural network. So that was… that was kind of fun.

323
00:30:28.990 --> 00:30:29.960
330W: Okay.

324
00:30:30.120 --> 00:30:40.079
330W: I am going to shift gears into a different set of data and a different set of neural networks, but the thing that's going to be kind of in common here is that I'm looking at

325
00:30:40.110 --> 00:30:53.580
330W: data that is coming from a physical system that I want to understand, and I'm going to try to bake into the neural networks that I'm using to analyze them my knowledge of physics, so that when it spits something out, I can interpret it as a physical answer.

326
00:30:54.410 --> 00:31:02.860
330W: So, I mentioned early on that you can look for dark matter annihilating in the universe, right? And one way that it could annihilate would give you gamma rays.

327
00:31:03.500 --> 00:31:11.839
330W: And you go, you build a telescope, the Ferrick Emirates Space Telescope, and you go and you look at the universe and take an entire map of the sky.

328
00:31:12.000 --> 00:31:31.280
330W: And if you look in the center of the galaxy, where there is the most dark matter, and remember, that's where I want to measure what the dark matter is doing, for the last 17 years, we've known that there is an excess of gamma rays that we cannot identify as coming from any known astrophysical source. It is very hard for you to see it by eye, about 10%,

329
00:31:31.370 --> 00:31:34.749
330W: Going out, you know, falling out this way.

330
00:31:35.060 --> 00:31:44.410
330W: What that signal would look like if it is dark matter is basically a sphere or a ball of annihilation of gamma rays coming from it, where the dark matter is concentrated.

331
00:31:45.230 --> 00:31:59.860
330W: To extract that from this data, what you need to do is you say, well, there's, like, an isotropic component of gamma rays, and then there's all of these horrible point sources of terrible astrophysical objects that are gonna look like, you know, if you give me gamma rays.

332
00:31:59.860 --> 00:32:09.500
330W: Notice they cluster towards the center. There's a thing called the Fermi bubbles, which are these, like, big jets of energy going above and below the galactic center that are emitting gamma rays.

333
00:32:09.500 --> 00:32:19.379
330W: There's pi-Brem, you know, Bremstrong and pions coming from the gas, interacting with cosmic rays, there's inverse Compton scattering, there's all of these backgrounds, so you add these all up.

334
00:32:19.380 --> 00:32:25.700
330W: And you say, oh, and then there's a small component on top of that that is not consistent with any of these.

335
00:32:26.310 --> 00:32:32.779
330W: So we've known this for 17 years, and we've had many, many long arguments about what it is, because it really does look like dark matter.

336
00:32:34.750 --> 00:32:47.719
330W: But if you look in the literature, we've all… in the arguments, there are significant disagreements about the galactic center excess morphology and its intensity, and those seem to be different… due to different choices about the…

337
00:32:47.950 --> 00:33:03.989
330W: Particularly these two, because we don't know these. They're models built on our other observations coming from other, like, studies of where the gas is in the galaxy and how much cosmic rays are in different places, right? So we don't actually know those, we're modeling them.

338
00:33:05.590 --> 00:33:12.140
330W: And in particular, we often add in another component, another background component, which is just called the stellar bulge.

339
00:33:12.270 --> 00:33:24.389
330W: What you do is you just say, like, well, astronomers have told me that there's a bunch of stars around the center of the galaxy. What if there were gamma rays coming from those stars, you know, from tracing those stellar populations? And you have a whole bunch of different ones.

340
00:33:24.420 --> 00:33:34.119
330W: Right? That have different shapes depending on which stellar population you think might be tracing the gamma rays, and then you can argue for a decade about whether your choice or their choice was right.

341
00:33:34.330 --> 00:33:37.689
330W: And so people have been sort of trying to figure out what's going on.

342
00:33:37.990 --> 00:33:42.299
330W: And what we did, with my student, Ed Ramirez,

343
00:33:42.830 --> 00:33:50.399
330W: was we said, okay, I'm going to fit this, but I'm not going to make a choice about what the signal looks like in morphology.

344
00:33:50.520 --> 00:34:03.900
330W: I'm going to give it a completely flexible template that can be whatever it wants, right? So we used a Gaussian process, we had to use some sort of deep learning to actually train that process, because there are a lot of variables that had to be minimized over.

345
00:34:04.680 --> 00:34:06.129
330W: But what you get out.

346
00:34:06.750 --> 00:34:22.700
330W: is a model of everything that cannot be fit by any of the templates. So this is a completely, sort of, agnostic picture of what the excess looks like that doesn't know about any of the shapes of the bulge or the NFW.

347
00:34:23.060 --> 00:34:38.400
330W: you can then fit that to the various templates that we have, and what we discovered is it likes one of the stellar templates with an NFW component on top of it. It does want there to be dark matter, shaped annihilation there. So that sort of was very nice.

348
00:34:38.480 --> 00:34:46.780
330W: It was a lot… a way to kind of analyze this without using any, you know, particular choice of what we thought the templates needed to be.

349
00:34:46.950 --> 00:34:58.969
330W: Right, so we did that first, and we just kind of took all of the gamma rays, and we ignored energy, and part of the problem there is that you need to nail down some of the templates by using data that's sort of outside this circle.

350
00:34:59.810 --> 00:35:05.799
330W: So work that we're doing right now, led by my student Sam, is that we are

351
00:35:05.800 --> 00:35:22.970
330W: extending this out, and we're having, like, a Gaussian process, not just in position, but position in energy space. So you're getting the spectrum of the annihilation as well. So each one of these is a different energy bin, and so now you can see, like, the blur of gamma rays that can't be associated with any background.

352
00:35:23.110 --> 00:35:31.789
330W: in… across different energy banks, right? So we're very… I think we're very happy with these results. It's still Premier. We just had a meeting about it to start writing it out.

353
00:35:31.850 --> 00:35:44.600
330W: Can you repeat what NFW is? Sorry, NFW is Navarro, Frank, and White. It is a standard assumption about what dark matter seems to do in a galaxy towards the galactic center. That's a profile.

354
00:35:44.680 --> 00:35:49.610
330W: Right? It's a power law. Okay. Right? And so, when I say NFW, sorry, I should…

355
00:35:49.790 --> 00:36:02.759
330W: is using, too much lingo, it's the shape of the dark matter. But again, all of these, when you actually analyze the gamma ray annihilation, you get a very particular choice of what that profile should look like.

356
00:36:03.140 --> 00:36:12.449
330W: If I could use the Fermi data to actually tell what the profile looked like close to the galactic center, I could compare to see whether it fits what this is.

357
00:36:12.610 --> 00:36:16.169
330W: Right, there's a power law. This wants a power law slope of 1.2.

358
00:36:16.520 --> 00:36:19.180
330W: Right? Okay, we haven't actually ever measured that.

359
00:36:19.580 --> 00:36:24.900
330W: Right? That's just coming from simulation plus data. If I had the Gaia data, I could tell you, hopefully.

360
00:36:25.020 --> 00:36:25.760
330W: Right.

361
00:36:26.350 --> 00:36:27.600
330W: The guy at the end?

362
00:36:27.720 --> 00:36:35.100
330W: With the mapping of dark matter… sorry, the stars… using the stars to map the dark matter potential closer and closer to the lactic center.

363
00:36:35.220 --> 00:36:41.199
330W: Right? I need to get very close to the galactic center, and I'm not close enough yet to tell me what the…

364
00:36:41.340 --> 00:36:43.970
330W: What dark matter is doing in this region.

365
00:36:48.550 --> 00:36:57.589
330W: One of the other issues going on here is that we have these background templates, as I said, and, like, a lot of this work is just sort of, like, well, you have this template, and what's on top of that template?

366
00:36:57.810 --> 00:37:02.209
330W: And… To get these templates, we have to kind of…

367
00:37:02.950 --> 00:37:13.449
330W: make a bunch of assumptions about how cosmic rays are propagating in the galaxy, and then what that will do for the gamma rays. And you use a code, a cosmic ray propagation code, the canonical one is called Galprop.

368
00:37:13.450 --> 00:37:23.480
330W: And you pick a bunch of parameters, right, the magnetic fields in the galaxy and the diffusion parameters in the galaxy, and you say, okay, go forth and solve, and tell me what the gamma ray spectrum should look like.

369
00:37:23.590 --> 00:37:27.639
330W: Each time you do that is several hours.

370
00:37:27.780 --> 00:37:43.780
330W: And that is too slow to be able to do, like, an MCMC, kind of like, you want to regress towards the right answer, where you vary your backgrounds. What you do is you just say, like, okay, here is a choice, or maybe two choices, or if you're really, really, working hard, you might have, like, 80 choices.

371
00:37:44.620 --> 00:37:54.429
330W: We would like to not have to do that. We would like to have a better way to do background propag… the diffusion propagation to give me the gamma rays. And then we remember.

372
00:37:54.550 --> 00:38:04.150
330W: Galprop is just solving a diffusion equation. It is solving this diffusion equation, right? Where there's a bunch of parameters that are, like, diffusion parameters in the galaxy, and

373
00:38:04.420 --> 00:38:12.460
330W: When you solve this, it should equal zero. This gives you the cosmic ray flux, and from the cosmic ray flux, you can get the gamma rays.

374
00:38:13.130 --> 00:38:28.549
330W: That is a perfect thing for a neural network to do. You can just build a pin, a physics-inspired neural network, where the loss term is basically this thing squared. Because if you've guessed the right cosmic rays, it's zero.

375
00:38:28.950 --> 00:38:44.940
330W: Right? And so you just minimize it. And so you can just kind of minimize this thing. The very cool thing that what you can do with this is that normally you're just thinking of the cosmic rays as a function of, like, position on the sky and the energy. You can make it conditional on all of the shit you don't know.

376
00:38:44.970 --> 00:39:02.219
330W: all of the diffusion parameters, all of these things that are unknown, and then you… it is a neural network. It can do backpropagation. It can take derivatives with respect to these things. And so you can just slide them around, and in principle, you can use that… you can have marginalize over them.

377
00:39:02.220 --> 00:39:08.270
330W: as you are trying to fit your data. So you're fitting across all of the maps at once, not just one map and another map.

378
00:39:09.390 --> 00:39:19.029
330W: This is a, our first stab at this. Again, you're probably not going to be able to see it, because the general shape stays the same, but if you look carefully, the intensities are changing.

379
00:39:19.290 --> 00:39:24.699
330W: What we're doing is just varying the parameters, and as you vary the parameter, the map changes.

380
00:39:24.870 --> 00:39:40.360
330W: Right, and so what you'd like to do is you do a FIT2 data, vary those parameters as you're varying all of the other things in your fit, presence of dark matter, things like that, and now your answer is going to be less constrained by, like, I picked this particular diffusion.

381
00:39:40.800 --> 00:39:56.989
330W: And so this will be a really, nice thing to have, and we're starting to think about actually doing that fit right now. Of course, there's issues, but we're hopeful that we'll overcome them. Oh, that's the cosmic ray density.

382
00:39:57.240 --> 00:40:14.030
330W: cosmic ray flux, which then can be converted into the gamma rays that we see here. So it's… it's a whole series… it's actually multiple… I'm simplifying the expression, obviously. There's, like, protons and antiprotons and photons and electrons, and you have to worry about the coupling between all. So it's actually solving several,

383
00:40:14.030 --> 00:40:28.039
330W: terms altogether and has to set all of them to zero. But you can do it, right? And it takes, like, an hour or so to train, and not even, like, an hour, like, of the most expensive GPUs you need, right? Now, we might need to train it more

384
00:40:28.160 --> 00:40:38.639
330W: That's one of the things we're working on right now. My postdocs are staring at me, going… It's all just… yeah. Electrons and protons are all classically… Yeah.

385
00:40:39.470 --> 00:40:43.719
330W: Like, we assume we know all the loss terms, and all the couplings between them.

386
00:40:43.880 --> 00:40:51.309
330W: Yeah. In your previous… Wordsman equation, it was all, like, just interacting through gravity and stuff, so this…

387
00:40:51.310 --> 00:41:07.169
330W: the physics that is going into this is… Just the transport equations, and then loss terms like synchrotron, inverse constant, things like that. Right. And there are inputs that you need, but now we can vary over those inputs. The point is that, like, we don't actually know the cosmic ray density everywhere in the galaxy.

388
00:41:07.170 --> 00:41:09.469
330W: Right? And most we get to measure it here.

389
00:41:09.610 --> 00:41:20.499
330W: you get to measure things that are related to the cosmic ray density elsewhere in the galaxy, like the gamma ray flux, but that's what I want to use, right? So I want to be able to vary over these things so that

390
00:41:20.510 --> 00:41:39.599
330W: my answer is not, like… we have had, like, very long arguments in this community of, like, I don't see this, why don't I see this? Like, when you're using Model O' from Galaproprop, and I'm using Model F from Galaproprop, and we disagree, and I want to know, like, if I didn't care, if I could just fit over all of them at once, what would I get?

391
00:41:40.890 --> 00:41:58.239
330W: And again, I think that's, like, a really general, like, we, in physics, we have things that we want to solve that are differential equations, and we just want the input, right? We don't need a cute, you know, analytic expression for it. This is a very powerful way to spit out those answers from data, or just from

392
00:41:58.240 --> 00:42:12.190
330W: you know, you have some model expectations. These are just gradients, basically, so this is sort of like a way to look at the gradient of the output with respect to different parameters. So the idea being, you have this, you take these gradients.

393
00:42:12.220 --> 00:42:14.959
330W: And then you just kind of marginalize over them when you're doing this.

394
00:42:16.220 --> 00:42:25.129
330W: So this will be pincer when we get around to it, to putting it out, but we're very excited about, the opportunity there.

395
00:42:26.270 --> 00:42:38.779
330W: Okay, so in the last part of the talk, I'm just going to kind of talk about some things that we are interested in and that we're working on, but don't have published results yet, and then I'll talk about some connections back to theory.

396
00:42:38.780 --> 00:42:51.499
330W: So one of the things that I'm really interested in, because there's a tremendous amount of data coming out in this, and it is connected to the physics of dark matter, is this large-scale structure, which I touched on at the very beginning as part of the evidence for dark matter.

397
00:42:51.890 --> 00:42:53.980
330W: If you look at galaxies.

398
00:42:54.420 --> 00:43:12.000
330W: over very, very large distances, and this is, again, simulations coming from the abacus simulation. I'm standing here, this is just the galaxy stretched out away from me, right, over millions of light years, or billions of light years, at least. And if you kind of squint, you can see that cosmic web pattern.

399
00:43:12.210 --> 00:43:21.200
330W: And buried within that pattern is information about baryon acoustic oscillation, and therefore information about the history of the universe.

400
00:43:21.960 --> 00:43:33.719
330W: There is an underlying assumption that goes into all of, like, cosmology, which is that the universe started out very, very uniform. So you have uniform initial conditions that evolved under a set of

401
00:43:33.939 --> 00:43:39.300
330W: Field equations that are actually a perturbative field theory that give you this.

402
00:43:39.790 --> 00:43:44.900
330W: Right? And then the game of cosmology is figuring out, like, what are those terms and those equations?

403
00:43:45.569 --> 00:43:49.010
330W: Again, this is a perfect playground for a pit.

404
00:43:49.640 --> 00:44:06.500
330W: If I want to learn the transport, there are a set of neural networks that are called flow matching. You just give it data, and you say you start from this other uniform or Gaussian distribution, just learn the transport from that, where you started, to whatever the data looks like.

405
00:44:07.109 --> 00:44:10.690
330W: If you combine that with an idea called optimal transport.

406
00:44:10.880 --> 00:44:15.639
330W: And what this will do is it will figure out what initial point and what final point to link.

407
00:44:15.760 --> 00:44:18.550
330W: And you can actually show that,

408
00:44:19.149 --> 00:44:31.120
330W: If you have perturbative field equations and, like, you know, conservation laws, this optimal transport will create a unique map from the beginning, the starting point to the end point, and combining with that

409
00:44:31.529 --> 00:44:46.089
330W: you can find the mapping that each kind of tracer halo took from the initial conditions, where the universe was nearly flat, or nearly uniform, to this map today. And if you're looking very carefully… oh, no, it didn't start running, sorry.

410
00:44:47.620 --> 00:44:53.170
330W: This is just a little simulation where you start… every red dot starts as a uniform.

411
00:44:53.340 --> 00:44:59.029
330W: And it just evolves along this optimal transport flow match map to a final state.

412
00:44:59.220 --> 00:45:17.990
330W: And it sort of just pulls in towards the data. It knows that there should be, you know, optimal transport and a flow. It doesn't actually need to know all of the details of the physics, which is really nice, because if you want to vary over different physics, I have a mapping, and I can see how well the physics matches that mapping.

413
00:45:18.700 --> 00:45:21.430
330W: The other kind of cool thing about this.

414
00:45:21.730 --> 00:45:37.009
330W: is that you will never get the fine structure. If you look really carefully, this is just a upsampled version. The red are the sampled flow locations, and the black are the actual, like, tracer halos in the simulation. If you look carefully, the black is a lot more clustered.

415
00:45:38.350 --> 00:45:44.670
330W: This is a field theory, it is a perturbative field theory. As the universe evolves, we are not in a perturbative

416
00:45:45.050 --> 00:45:57.580
330W: initial, you know, the amount of matter around us is not a perturbative deviation from the early universe. There's a lot more matter here than here, right? It's more than an order one deviation.

417
00:45:57.730 --> 00:46:04.829
330W: The universe has gone non-perturbative on small scales, and the field theory breaks down.

418
00:46:05.150 --> 00:46:24.560
330W: And that means that the optimal transport cannot work. It simply is insensitive to it, and the flow matching simply can't work. But what this does is it means that this combination learns the perturbative mapping, not the perturbative plus non-perturbative mapping. And you could say, well, that sounds terrible, I'm not getting the right mapping.

419
00:46:24.700 --> 00:46:26.350
330W: I don't care about the non-perturbative fist.

420
00:46:27.640 --> 00:46:36.440
330W: is part. Non-perturbative physics is the thing that, like, all of the… That's a trophy?

421
00:46:36.570 --> 00:46:46.409
330W: The non-perturbative physics is like shell crossing, where, like, galaxies are starting to actually form and, like, build up, like, really large structures at very high densities.

422
00:46:47.170 --> 00:46:59.099
330W: If you understand the perturbative physics, that's where you can kind of modify the early universe physics by changing how dark matter works, or changing the struct by dark energy, or something like that, and that will show up in the perturbative structure.

423
00:46:59.110 --> 00:47:03.840
330W: And where I am right now is that this low-matching architecture

424
00:47:03.840 --> 00:47:21.589
330W: seems to mimic the statistics of the perturbative field theory in the theories that we've looked at, and so that gives me some opportunity to then say, like, I've learned that in a very model-independent way, and now I could fit different models to it and see how, if I change physics, what happens.

425
00:47:22.810 --> 00:47:29.830
330W: So that's a, kind of a direction that I'm really interested in. There's a lot of data coming with that, and I think there's a lot of very cool machine learning connections.

426
00:47:30.370 --> 00:47:42.160
330W: I'm gonna end just by saying, well, I actually am a theorist, and I started off doing theory, and machine learning is a way to get a handle on what is going on in dark matter and in the universe, right?

427
00:47:42.160 --> 00:47:54.600
330W: But when you get observations, you know, that may point us to how should we think about dark matter, and then maybe what new observations we have to make. I'm just going to talk about a couple of things that I've been thinking about of late.

428
00:47:54.630 --> 00:48:07.490
330W: One of which are these little red dots from JWST, so those are the little red dots. Very, very high redshift objects that appear to be supermassive black holes that are kind of very massive, very, very early on.

429
00:48:07.850 --> 00:48:09.030
330W: And…

430
00:48:09.360 --> 00:48:21.150
330W: there seems to be some argument, and the astronomers, I'm gonna say, seem to go, you know, like, we're still figuring this out, but there's some suggestion that these black holes have grown faster than what's called the Eddington limit.

431
00:48:21.590 --> 00:48:34.500
330W: Eddington limit is if you have a black hole and you try to throw mass into it in a spherically symmetric way, the gas is going to heat up, and the gas as it heats up, will blow the rest of the gas behind it out. So you can't feed a black hole arbitrarily fast.

432
00:48:34.680 --> 00:48:43.919
330W: you can supersede the Eddington limit, you can, like, throw a jet of gas into a black hole faster than that, so you can do super Eddington accretion in regular physics.

433
00:48:44.550 --> 00:48:54.759
330W: But, you know, these are really massive, and maybe it's really hard to build these things using standard physics. So it was interesting to ask, can you do this with dark… can dark matter help?

434
00:48:55.250 --> 00:49:01.610
330W: And one thing to know is it's actually really hard to throw things into a black hole, because they're really small, and you tend to miss.

435
00:49:01.730 --> 00:49:21.370
330W: Right, so it zips around the black hole and comes right back out. To throw it into a black hole, you need a cooling mechanism. Cold dark matter does not have a cooling mechanism. But what we discovered with my postdoc, Nico and I, was that you can kind of build a version of the standard model in dark matter, where there is, like, dark atoms that have their own cooling mechanism.

436
00:49:21.370 --> 00:49:32.190
330W: And in that… there is an open region of parameter space where you can cause small clusters of dark matter to cool and collapse, and would form black holes very, very early on.

437
00:49:32.370 --> 00:49:42.430
330W: And because, it turns out with these parameters, this is just a cooling curve of, like, temperature and density for dark matter structures. So, like, an object would start here, cool.

438
00:49:43.450 --> 00:49:52.289
330W: the cooling becomes inefficient, and so you're just, like, your density spikes. This is exactly how stars form. You fragment, you create all these little things that would form black holes.

439
00:49:53.140 --> 00:50:10.610
330W: And at that point, because you have this cooling mechanism, it becomes very easy to have very, very high accretion rates for dark matter onto black holes, and so you could build a seed of a black hole that could go on to be these little red dots. So I'm not arguing that little red dots are black hole fed… they're dark matter fed black holes.

440
00:50:10.610 --> 00:50:17.139
330W: But if, you know, more and more evidence from the astronomers becomes, like, look, this is really hard to do in standard, you know.

441
00:50:17.140 --> 00:50:22.309
330W: scenarios, then you're gonna have to get weird. And one of the weird things you can do is look at dark.

442
00:50:23.060 --> 00:50:25.049
330W: The very last thing I'll tell you about

443
00:50:25.180 --> 00:50:30.189
330W: is, some work that I've been thinking about with, the Hubble tension.

444
00:50:30.380 --> 00:50:31.720
330W: Right, so…

445
00:50:31.900 --> 00:50:44.679
330W: I went, you know, I talked about the early universe and the expansion of the early universe, and in that there are all these sound waves, in the plasma, and you can use that to map out knowledge of how the universe was expanding.

446
00:50:45.220 --> 00:50:53.690
330W: From that, you can measure, the… what you think the expansion rate of the universe, what's called H0, today should be.

447
00:50:54.040 --> 00:51:10.240
330W: You can also go and measure H0 today by looking at supernova around, you know, close to the Earth. And by close, I mean only a few billion light years, right? So you have two different measures of H0, and for a long time now, we've known they disagree, and they disagree at something like 7 sigma.

448
00:51:10.380 --> 00:51:12.379
330W: This is called the Hubble tension.

449
00:51:12.830 --> 00:51:17.409
330W: And the problem is, is it's very hard to fix this with new physics.

450
00:51:17.580 --> 00:51:30.429
330W: Because anytime you try to introduce new physics, you change how these acoustic oscillations in the early universe go, and so you start breaking all of the other defined structure in these,

451
00:51:30.810 --> 00:51:32.380
330W: In these power spectra.

452
00:51:32.780 --> 00:51:35.309
330W: So what we did is we said, okay.

453
00:51:35.700 --> 00:51:38.079
330W: We're going to introduce dark matter.

454
00:51:38.400 --> 00:51:44.599
330W: And the standard model's not boring. We have forces, right? We have a force that's called QCD.

455
00:51:44.910 --> 00:51:48.749
330W: I'm going to give Dark Matter its own QCD, so dark QCD.

456
00:51:48.960 --> 00:52:03.740
330W: When you do that, you're going to create… you can create light relativistic particles around the dark matter, which we call dark radiation. And that will change the evolution of the Hubble parameter in the real universe, right? And that's been known for a long time.

457
00:52:04.150 --> 00:52:07.920
330W: If you do it with dark QCD, that dark radiation becomes sticky.

458
00:52:08.100 --> 00:52:09.699
330W: Becomes like molasses.

459
00:52:09.890 --> 00:52:25.200
330W: And that allows energy to flow in a different way than any other model of the early universe has allowed. And what that does is, if you look carefully, like, this is sort of the… zero here is, by definition, the best fit to the,

460
00:52:25.530 --> 00:52:31.620
330W: to Lambda CDM, this is the standard cosmology. And if you look carefully, you can see the data kind of fluctuates around it.

461
00:52:32.180 --> 00:52:43.310
330W: Our model is this red, so it stays very, very close to the, Lambda CDM. There's a little bit of drop-off here, but the data is absolute trash over there, so you can't tell.

462
00:52:43.420 --> 00:52:56.109
330W: Right? And it's the stickiness of this dark radiation that keeps it so close, but while it is doing that, it's keeping the CMB the same, it's doing that with a much larger Hubble parameter.

463
00:52:56.340 --> 00:52:58.120
330W: So here's H0.

464
00:52:59.470 --> 00:53:18.449
330W: This region here, this value of H0 around 67 kilometers per second per megaparsec is the value you want from standard lambda CDM fits the rule universe. This value up here at 73 is what you want from studies of the supernova, from an experiment called Shoes, and the fact that they are different is the entire problem.

465
00:53:18.510 --> 00:53:24.350
330W: Right? With our new scenario, you can move the,

466
00:53:24.460 --> 00:53:28.300
330W: H0 value up to about 71, 72.

467
00:53:28.450 --> 00:53:35.190
330W: It's not where it needs to be, it is lower, it's about 3-signal low. 3-signal low is the best that anyone can do.

468
00:53:35.320 --> 00:53:40.349
330W: The other thing that our model does is that if you are aware of the DESE,

469
00:53:40.520 --> 00:53:58.919
330W: survey that, the DESI anomaly, there's been a lot of news recently. By studying large-scale structure over, you know, redshifts from half up to two, there's been these kind of anomalies where it seems like the way that,

470
00:53:59.090 --> 00:54:08.489
330W: the length… the barrenacoustic oscillation length scale is changing over cosmic history seems to be a little bit different from what Lambda CDM wants.

471
00:54:08.610 --> 00:54:17.270
330W: That's shown here. Again, 0 would be if lambda CDM was perfectly right from the Planck values, and the data are all these circles here.

472
00:54:17.430 --> 00:54:19.500
330W: And you can see they drift way off.

473
00:54:20.220 --> 00:54:33.400
330W: This has caused a lot of interest, and people have introduced, like, changing dark energy, to try to fit this. The problem with that is that if you change dark energy in a way to fit this, you actually need to have dark energy that is phantom.

474
00:54:33.530 --> 00:54:41.810
330W: Which, as a general relativist, that is the kind of dark energy you need to build a time machine. I believe that many things are possible in this universe, I don't believe that is.

475
00:54:42.150 --> 00:54:49.209
330W: Our model, because it changes the early universe physics, brings the measurements that you would see in DESE down.

476
00:54:49.550 --> 00:55:01.860
330W: And it's not as good of a fit as that phantom Dark Energy model, but it's actually very close. There's a couple points you missed, but if this was the only miss you had gotten, no one would have cared, right? That's like a less than one sigma deviation.

477
00:55:02.870 --> 00:55:11.769
330W: The point of this is not that, like, this is definitely the right model for everything, but there are these really interesting things we are learning from these large astrophysical data sets.

478
00:55:11.770 --> 00:55:28.700
330W: The work that I've shown you with machine learning is trying to learn more things from those large astrophysical data sets. One of the reasons I'm interested in flows applied to large-scale structure is I want to know what the BAO scales are doing to higher accuracy than we do now, because that's the only way that that's going to be resolved.

479
00:55:28.770 --> 00:55:43.389
330W: Right? And so I've shown you a whole bunch of things, on a whole bunch of different topics, right? But they are related in the sense that dark matter physics comes into all of them, and in, you know, using machine learning, using these new techniques gives you a new way to look at it.

480
00:55:43.850 --> 00:55:45.700
330W: I will flash off.

481
00:55:45.720 --> 00:55:56.880
330W: On the right, here's a bunch of stuff I didn't talk about, that are… I couldn't fit into this talk. But I also want to thank, of course, the people that I've been working with.

482
00:55:56.880 --> 00:56:09.169
330W: So starting with the students, who have graduated are up here, students who are still working here. Blue, are technically David Cheese students, but I've worked with them a lot, and it's just absolutely tremendous to have.

483
00:56:09.170 --> 00:56:20.510
330W: such a great group around the postdocs, and then the faculty, both here and elsewhere. Especially thanks to David Shee, who got interested in machine learning well before it ate the world.

484
00:56:20.520 --> 00:56:28.429
330W: And so we had a leg up of all sorts of interesting things you could do, and that's been a lot of fun and a lot of opportunities.

485
00:56:28.450 --> 00:56:30.090
330W: been able to do. So, thank you.

486
00:56:40.330 --> 00:56:44.059
330W: You can take questions in the room as well as online. Let's…

487
00:56:45.070 --> 00:56:47.220
330W: Go ahead and start. Josh.

488
00:56:47.350 --> 00:56:59.969
330W: Very great talk, I've enjoyed it a lot. At the… in there, when you were doing the… the dark QCD axia, or the dark QCD, I don't know… There's no Axion on the use case, yeah. Yeah, no Axion, okay.

489
00:57:00.380 --> 00:57:03.310
330W: Couple of questions,

490
00:57:03.570 --> 00:57:23.940
330W: is this doing things similar to, like, early dark energy models, where you have, like, a dark energy component and… Our dark energy is completely boring, it just stays constant. But it… does it have the effect of lowering the sound horizon as well? Yes. All of these have the same… like, the goal is always the same, find a way to lower the sound horizon. Yeah, yeah, yeah.

491
00:57:24.010 --> 00:57:26.100
330W: Does it,

492
00:57:27.380 --> 00:57:40.269
330W: Have you tried also including the, supernova data, not the shoe sample, but the… Yeah, this fit includes Pantheon. Okay, great. Is it, Pantheon Plus? Correct.

493
00:57:42.280 --> 00:57:59.269
330W: You discussed dark matter annihilation, I don't know what that is, but I would like you to explain what that is, but can you explain it in the context of the bullet cluster, where you said that there are two clouds of dark matter that just pass through each other?

494
00:57:59.480 --> 00:58:05.169
330W: The bullet cluster is kind of… you're on this very, very large scale, and on those scales.

495
00:58:05.870 --> 00:58:19.069
330W: even stars… like, if you throw two galaxies at each other on those scales, the stars are just going to pass right through each other. They very rarely hit. So what you're looking at is just tracing by the fact that there's gravity there, right?

496
00:58:19.220 --> 00:58:23.510
330W: Now, if you go way, way back, or go way back to the beginning.

497
00:58:24.460 --> 00:58:26.859
330W: There is one thing that interacts.

498
00:58:27.210 --> 00:58:32.279
330W: And that's the, the hot gas?

499
00:58:32.570 --> 00:58:40.709
330W: So, when you throw a cloud of hot gas at another cloud of hot gas on infographic scales, they slow down, because they have a long-range force.

500
00:58:41.140 --> 00:58:59.580
330W: Right? They have electromagnetism. So that is why this cloud of gas didn't end up over here, they're separated. And that's why this one is not over there. They've slowed down and exchange momentum. So actually, you can write down a whole bunch of really interesting article theory knowledge from this, namely that dark matter doesn't do that.

501
00:58:59.820 --> 00:59:06.490
330W: Dark matter never slowed down, doesn't have a long-range force, at least that is useful on this scale. So you set it up earlier.

502
00:59:06.810 --> 00:59:22.490
330W: Dark matter annihilation is just like… it's like electron-positron annihilation. If you just have a particle, and it's antiparticle, and dark matter could be its own antiparticle, or you could have a mixture of dark matter and anti-dark matter, both be what we call dark matter in these contexts.

503
00:59:22.490 --> 00:59:29.109
330W: If they get close enough, then there's, you know, the chance of a, you know, exchange of a force carrier in the IFA.

504
00:59:29.110 --> 00:59:30.840
330W: What they annihil it into.

505
00:59:30.970 --> 00:59:48.930
330W: Well, if I knew, that'd be great. What you do is you say, like, well, what could it annihilate into, right? And because you're looking at astrophysical scales, you need is something that will get to you, right? So they annihilate into many things that won't get to me, I don't care. So it's mostly photons and neutrinos are what we care about.

506
00:59:49.180 --> 00:59:55.170
330W: Now, if they annihilate into any standard model particle, you will, in the end of the day, get photons out.

507
00:59:55.390 --> 01:00:07.479
330W: how many depends on your choice. What we do is we just model a bunch of different options. And so this is just a game of, like, I have no idea what the dark sector is doing. Let me make a guess and check.

508
01:00:07.580 --> 01:00:15.270
330W: Right? And… the Fermi signal is actually consistent with one of, like, the easiest guesses you could have made.

509
01:00:15.460 --> 01:00:24.199
330W: Right? It's not a particularly… you don't have to get real weird with it. If you just annihilate into a very wide selection of standard model particles, you'll get gamma rays that look more or less correct.

510
01:00:24.610 --> 01:00:33.960
330W: Right? Now, if I knew for sure that it was dark matter annihilation, then I could really drill down and start ruling things out, but right now, it's just like, you know, here's a whole class of options.

511
01:00:34.860 --> 01:00:36.649
330W: There's a question there, and back there.

512
01:00:36.840 --> 01:00:52.040
330W: So this dark QCD model? Yeah. It should have implications in dense regions in the universe, right? Yeah, so the model that we cooked up is safe for all observations, but makes a bunch of predictions that the scale's right below.

513
01:00:52.160 --> 01:01:10.739
330W: Where you have any. That was not entirely, cooked up for purposes of saying you can look for it. The parameters that you get out of the early universe fits suggest that it would be very easy to have structure. You could change structure at scales right below, like, WorkCo

514
01:01:11.760 --> 01:01:29.270
330W: So, that's a real… that's sort of what I want to look at next, because we've thought about that a lot in different contexts, but this model is actually weird enough that I'm not entirely certain I know what's going to happen. But you could imagine that there would be pretty significant changes. I don't think it's constrained by current data, but it could be abruptly soon.

515
01:01:29.270 --> 01:01:35.520
330W: So you're still talking about structure, though? Yeah. What about in very dense systems, like microwaves?

516
01:01:35.990 --> 01:01:52.140
330W: It's very hard to get enough… so there's a lot of interesting work about, like, trying to squeeze enough dark matter inside of a white dwarf or a neutron star to change its structure, and you can do that, but it's very hard to do that because, white dwarfs are not very big, and matter is not very dense.

517
01:01:52.140 --> 01:02:02.490
330W: So it's hard to crunch enough dark matter into a star to actually change its structure very much. You can, there are models that do that. But even if you just take, like, the sun.

518
01:02:02.490 --> 01:02:14.819
330W: and just sweep out its path through the entire galaxy, over its entire lifespan. You can add up how much dark matter it could have eaten, and it's very, very low. You really would have to know the structure very well.

519
01:02:17.060 --> 01:02:20.150
330W: And then I don't know if there's questions online. Andrew.

520
01:02:20.440 --> 01:02:31.540
330W: Yeah, so Matt, one of the other, proposals that has been made for the Galactic Center access is unresolved millisecond pulsars, and I was just curious if the work that you've done had

521
01:02:31.610 --> 01:02:51.560
330W: has anything to say about that, one way or the other. The exact work that I showed you has not. It is a problem that I'm very interested in. I have some, like, things that I'm working on right now that we're not even in the stage that I could show you for that. I will say there has been advances in the last, like, 3 or 4 years, so the…

522
01:02:51.670 --> 01:02:54.190
330W: The idea here, is that

523
01:02:54.630 --> 01:03:04.930
330W: There's just a bunch of things in the galaxy that give you gamma rays, so the ideas are called what are called millisecond pulsars, which are, you know, rapidly spinning, cores of dead stars.

524
01:03:05.230 --> 01:03:17.690
330W: And this was originally proposed, and there was some data that suggested that there was a population of millisecond pulsars that were right below the resolution of our telescopes. We could barely see them, and that's what we were seeing.

525
01:03:17.950 --> 01:03:23.940
330W: And there was some data in the… the… there was some analysis of the Fermary, data that suggested that.

526
01:03:24.650 --> 01:03:36.149
330W: About 5 years ago, that was reanalyzed, and it said, oh, well, we made a mistake, but there's a population of stars, like, it can't be that bright, but if they're about 10 times dimmer.

527
01:03:36.330 --> 01:03:40.800
330W: maybe that's what's giving you the Fermi,

528
01:03:40.990 --> 01:03:51.549
330W: access. And now it's even dimmer, and we're getting to the point where the number of, like Eugen, my postdoc, had done some work on this, we're getting to the point where every

529
01:03:51.730 --> 01:03:57.059
330W: Pulsar would need to give you one gamma ray, To give you the signal.

530
01:03:57.190 --> 01:04:07.179
330W: And if you think of a signal that comes… every gamma ray comes from one source, that is a smooth, continuous source that, if it was in millisecond pulsars, you'd call that dark matter.

531
01:04:07.180 --> 01:04:18.460
330W: Right? There's no difference between a dark matter particle giving you one gamma ray and a millisecond pulsar. But the number of millisecond pulsars now goes up. So you're at several hundred thousand, I think.

532
01:04:18.480 --> 01:04:27.300
330W: So I would really love to know if there's several hundred thousand millisecond pulsars in the Baptic Center, because I could solve this problem real quick.

533
01:04:28.830 --> 01:04:31.340
330W: Alright, we do have a question online.

534
01:04:32.310 --> 01:04:38.389
330W: Which is from Ethan, asking, how close do you think we are to solving the Hubble tension?

535
01:04:38.390 --> 01:05:02.750
330W: And how close are we to discovering dark matter? Well, I mean, obviously I have the right answer for the Hubble tension. I would say the question that I have for the Hubble tension is, I think, the question any theorist has to have when you have this cool signal that's like, oh my god, like, we're, you know, we found something beyond the standard model. Like, as a theorist, you go, like, okay, something has gone wrong here in the observations, and eventually somebody will figure out that systematic that they didn't have.

536
01:05:02.750 --> 01:05:04.950
330W: And they didn't know about, and this will get resolved.

537
01:05:04.950 --> 01:05:14.300
330W: The public attention has been going on for a long time now. I've listened to very, very smart people on both sides of that argue with each other about whether somebody has made a mistake or not.

538
01:05:14.610 --> 01:05:21.949
330W: At this point, I don't know of what could have gone wrong in supernova measurements that would have… that would resolve this.

539
01:05:22.130 --> 01:05:27.889
330W: I think increased knowledge of, like, the barren acoustic oscillations and things like that, that will

540
01:05:28.540 --> 01:05:46.190
330W: That will tell us whether there's something wrong in multiple ways. And if that really pushes us towards there needs to be a solution, then yeah, like, we have an idea that couldn't work, other people have an idea that can work, and we're just going to need more data, both to know whether it's real and know whether any one of these solutions is real.

541
01:05:46.190 --> 01:05:57.550
330W: I'm very happy with our idea because, like, there weren't… dark radiation models tended to not work in very complicated and detailed reasons, and so the fact that ours works as well as it does

542
01:05:57.550 --> 01:06:08.629
330W: at least points you maybe in a direction that we had missed, right? And we'll build on it, and other people will take it and, you know, find a new twist on it, and hopefully that will lead to something interesting.

543
01:06:08.630 --> 01:06:18.479
330W: For dark matter, I don't know, like, we'll open up the LZ data in a couple months, and there'll be many more events, and everything will be sorted, or there won't be,

544
01:06:18.630 --> 01:06:27.799
330W: I don't know. The ride is the fun part, right? Sorry. We've got a raised hand online from Saurabh. Saurabh, can you unmute and ask your question?

545
01:06:28.230 --> 01:06:40.370
Saurabh Jha: Sure, yeah, thanks, Matt. Could you, in discussing the large-scale structure, the learning from the smooth early universe to the data today, you know, you said you want to learn about this further…

546
01:06:40.380 --> 01:06:56.739
Saurabh Jha: perturbative part, which, you know, like, I guess we would call linear growth in astro. But how do you know, like, because your data is already in the nonlinear regime. So, like, how do you know that you're actually learning the linear or the perturbative part correctly, because you have to fit two

547
01:06:56.780 --> 01:06:58.129
Saurabh Jha: The nonlinear part.

548
01:06:59.000 --> 01:07:18.810
330W: Yeah, so what I can say… I mean, this is what I'm trying to wrestle my way through right now. When I have simulation truth, I can compare, and I can see that I'm learning the linear part. There are theoretical arguments about why OT, like the optimal transport, is only going to couple to the linear

549
01:07:20.860 --> 01:07:31.550
330W: I will say… I don't want to say linear, because it's not quite clear to me whether it only goes to the first order, or whether it can do higher order. I think it may be just linear, to be quite honest, but that would be good enough.

550
01:07:31.550 --> 01:07:42.689
330W: And it does seem to… there seems to be, like, good, sort of, theoretical reasons of how OT works and what kind of systems it can apply to. That suggests to me that

551
01:07:42.830 --> 01:07:46.920
330W: As long as your tracers are sufficiently…

552
01:07:47.520 --> 01:07:55.329
330W: they can be biased, but they, you know, you have to have some uniform understanding of that. It does seem that you are picking up this… the thing

553
01:07:55.440 --> 01:08:05.680
330W: that nonlinear perturbations then grew out of. But that is a thing that I am trying to put on very solid foundation before I make that complete argument.

554
01:08:06.330 --> 01:08:13.170
330W: And then we can come back to Ariane's question in the room. Yeah, I actually have, like, two questions, quite similar to it.

555
01:08:13.500 --> 01:08:19.860
330W: The first being, why do we assume that the universe started uniform? And then…

556
01:08:19.950 --> 01:08:31.870
330W: How is the optimal transport the only kind of transformation from that uniform to, like, what we have? So we believe the universe started very, very uniform, because it is consistent starting very, very uniform.

557
01:08:31.960 --> 01:08:37.399
330W: Things like CMB is, it is very uniform.

558
01:08:37.420 --> 01:08:47.299
330W: And it is consistent with being uniform with perturbations on top of it. You can actually tell a lot about the statistics of those perturbations. For example, that

559
01:08:47.300 --> 01:09:00.050
330W: the matter perturbations and the dark matter perturbations and the photon perturbations were all correlated, which is really weird. That's not the most generic way to start. So it's a very specific set of initial conditions if you make that assumption.

560
01:09:00.230 --> 01:09:11.300
330W: gives you something that is consistent with the data. It is an assumption, it's called the Copernican Principle, right, or the cosmological principle. And maybe the universe is conspiring against all of us.

561
01:09:11.300 --> 01:09:27.189
330W: And it didn't start that way, but it is certainly consistent with that. The statement about OTs is just that there are a certain set of systems that have, like, basically energy conservation or, like, perturbative systems, that OT will create a unique mapping from initial to final.

562
01:09:27.460 --> 01:09:40.570
330W: And then in this context, that mapping can be identified as the flow of the galaxies themselves. And then you can… in simulation, you can go and test that. And then there's observational issues, which, of course, you have to worry about.

563
01:09:44.970 --> 01:09:51.770
330W: Nahib question, but I was wondering if this, gamma ray data, does that sort of, is that,

564
01:09:53.060 --> 01:10:08.179
330W: picking one, one candidate microscopic model of that matter versus the other? Like, for instance, is axion that matter? Does it even produce… So, it would not be axions. Axions are a very, very light particle, and these are gamma rays up to the GED scale. So.

565
01:10:08.350 --> 01:10:27.189
330W: it's, like, almost embarrassingly simple what kind of dark matter this would be. It's, like, 100 GeVV dark matter that annihilates through a pseudoscaler to BB-bar pairs, and that gives you the signal. And part of the funny thing with this is that you can write down that theory. We wrote down that theory, like, 15 years ago.

566
01:10:27.250 --> 01:10:30.419
330W: And you can write down other theories, of course, but, like, that works.

567
01:10:30.610 --> 01:10:33.389
330W: And so the argument is all about whether

568
01:10:33.550 --> 01:10:45.929
330W: it's really dark matter. Like, whether there's some other millisecond pulsars or some other signal, it's not about the… the theory would be easy if it's really dark matter. It's a question of, do we believe it's dark matter? Because that's a very strong claim to make.

569
01:10:46.050 --> 01:10:50.400
330W: And we're obviously very loath to make it until we're really, really certain.

570
01:10:51.610 --> 01:10:58.310
330W: But these arguments about morphology and point sources and things like that, you're really down in the weeds of these very, very small

571
01:10:58.330 --> 01:11:13.970
330W: differences in the analysis, but they have to all be consistent. We really want to make sure that we're pushing in the right direction, and, you know, again, like, okay, prob- you know, like, it's just millisecond pulsars, and they've just been ruining my day for a really long time, but, like, there is a signal, and we want to know it is.

572
01:11:14.560 --> 01:11:16.099
330W: I think is what I said.

573
01:11:19.650 --> 01:11:39.339
330W: So, when you were talking about the black holes, were you saying that there is or isn't data for dark matter going into black holes? Right now, what we can say is there is a population of black holes in the early universe that are very, very massive at a size and luminosity that at least surprised the astronomers.

574
01:11:39.340 --> 01:11:46.130
330W: And I will not make the case… I will not say that they are, like, impossible to build in the Standard Mall. They probably are.

575
01:11:46.300 --> 01:11:54.819
330W: capable of doing that in the standard cosmology. But there seems to be, like, they're a little bit bigger than you might have expected before we saw them.

576
01:11:54.980 --> 01:12:02.020
330W: And so it is interesting to ask, could I do this? And the thing is, is that in most dark matter models, you can't.

577
01:12:02.190 --> 01:12:14.870
330W: It's very hard to throw dark matter into a black hole, because black hole is actually small. And if dark… dark matter has to lose its angular momentum to fall in, and most dark matter models simply cannot, because they don't have enough interactions.

578
01:12:15.070 --> 01:12:27.060
330W: So it's just… it's weird, because I think, like, a lot of people, when I explain this problem to them, right, think, like, oh, well, it's dark matter, it's obviously just going to fall into the black hole, and it's just… it's orbiting.

579
01:12:27.100 --> 01:12:44.159
330W: Right? And the orbit is hard to break out of and fall into the center, and so you can do this with normal matter, because normal matter cools through radiation. Dark matter that may have radiation may not, but it's actually highly constrained how much coupling dark matter can have to things like dark radiation coming from there.

580
01:12:44.370 --> 01:12:53.670
330W: So you have to pull all of these things together from, like, 13.8 billion years of cosmic time to say, like, is that model allowed or not? And often the answer is no.

581
01:12:53.840 --> 01:13:02.750
330W: But we did find a corner of parameter space that's totally fine, and so it's an existence proof. You could do this, right? Not saying it is happening, but you could.

582
01:13:03.580 --> 01:13:06.959
330W: A lot of theory is just exploring what is possible.

583
01:13:13.540 --> 01:13:17.749
330W: nothing new, so I think we can thank Matt again for a wonderful talk.

584
01:13:28.930 --> 01:13:29.790
330W: Forget it.

585
01:13:32.870 --> 01:13:34.590
330W: What's so funny?

586
01:13:38.940 --> 01:13:39.680
330W: True.

587
01:13:43.120 --> 01:14:10.309
330W: Yeah, you can… Yeah. Right. Now, yeah, but I'm sure we're going there.

588
01:14:10.950 --> 01:14:27.479
330W: Yeah, no, these are all… these are things I'm thinking about. Yeah, no, it's just… these are all just, like, classical equations. Yeah, yeah. We actually care about, like.

589
01:14:27.650 --> 01:14:35.319
330W: Yeah. Pretty nice.

590
01:14:35.320 --> 01:15:00.050
330W: I train people to show up on time.