WEBVTT 1 00:00:03.399 --> 00:00:07.719 Welcome to Bedtime Astronomy. Explore the wonders of the cosmos 2 00:00:07.759 --> 00:00:12.279 with our soothing Bedtime Astronomie podcast. Each episode offers a 3 00:00:12.359 --> 00:00:16.320 gentle journey through the stars, planets, and beyond, perfect for 4 00:00:16.399 --> 00:00:20.239 unwinding after a long day. Let's travel through the mysteries 5 00:00:20.239 --> 00:00:22.440 of the universe as you drift off into a peaceful 6 00:00:22.480 --> 00:00:23.800 slumber under the night sky. 7 00:00:26.920 --> 00:00:30.239 Today we are undertaking a simulation of well, what was 8 00:00:30.280 --> 00:00:32.439 long thought to be the impossible. When you look up 9 00:00:32.479 --> 00:00:34.880 at the night sky, what you're really staring into is 10 00:00:34.960 --> 00:00:39.200 a massive, almost incomprehensible computation problem. We live in the 11 00:00:39.200 --> 00:00:41.960 Milky Way Galaxy. It's home to over one hundred billion 12 00:00:42.039 --> 00:00:45.079 individual stars, and every single one of those stars, plus 13 00:00:45.079 --> 00:00:47.399 all the gas and dust and dark matter swirling around, 14 00:00:47.840 --> 00:00:51.119 is governed by a complex web of physics. For decades, 15 00:00:51.159 --> 00:00:54.359 the ultimate challenge for astrophysicists, kind of their Mount Everest, 16 00:00:54.679 --> 00:00:57.759 has been to create an accurate star by star model 17 00:00:57.799 --> 00:01:01.159 of this entire system to trace its history, its interactions, 18 00:01:01.159 --> 00:01:03.640 and its future. It's just it's always been out of reach, 19 00:01:04.000 --> 00:01:06.959 but the game has fundamentally changed. Our deep dive today 20 00:01:07.000 --> 00:01:09.480 is focused on a revolutionary breakthrough from a team of 21 00:01:09.560 --> 00:01:13.280 international researchers led by kiya Hiroshima at Reichen in Japan. 22 00:01:13.760 --> 00:01:16.319 Their work was just published, and it describes as simulation 23 00:01:16.439 --> 00:01:20.000 method so fast and so detailed that it has completely 24 00:01:20.040 --> 00:01:23.319 shattered the computational bottlenecks that have held back galactic modeling 25 00:01:23.359 --> 00:01:23.879 for years. 26 00:01:24.120 --> 00:01:26.640 This is genuinely a step change moment. I mean, it's 27 00:01:26.680 --> 00:01:29.200 not just an improvement. It's a completely different league. And 28 00:01:29.239 --> 00:01:32.040 you really need to hear the numbers to appreciate just 29 00:01:32.159 --> 00:01:33.519 how big this achievement is. 30 00:01:33.640 --> 00:01:35.000 Okay, laid them on us well. 31 00:01:35.719 --> 00:01:38.280 Previous state of the art simulations, the best we had, 32 00:01:38.760 --> 00:01:41.400 they struggled to get anywhere close to the true fidelity 33 00:01:41.400 --> 00:01:45.599 of our galaxy. This new model, it successfully and accurately 34 00:01:45.640 --> 00:01:49.719 represents more than one hundred billion individual stars. That right 35 00:01:49.760 --> 00:01:52.319 there is the resolution breakthrough. 36 00:01:52.200 --> 00:01:55.200 One hundred billion, so true star by. 37 00:01:55.040 --> 00:01:58.400 Star, true star by star. But the truly staggering part, 38 00:01:58.519 --> 00:02:01.560 the part that really changes things, is the efficiency game. 39 00:02:02.200 --> 00:02:03.840 They did this at a rate more than one hundred 40 00:02:03.879 --> 00:02:05.640 times faster than was possible before. 41 00:02:05.799 --> 00:02:08.439 Wait say that again, one hundred times the detail and 42 00:02:08.520 --> 00:02:09.439 one hundred times the. 43 00:02:09.400 --> 00:02:12.719 Speed at the same time. Thing about what that means 44 00:02:12.879 --> 00:02:15.800 We aren't talking about a small improvement. We're talking about 45 00:02:15.800 --> 00:02:19.719 going from a theoretical experiment that would take decades to 46 00:02:19.759 --> 00:02:21.919 something you can actually run in a matter of months. 47 00:02:22.680 --> 00:02:25.719 This is, and it's no exaggeration, the world's first simulation 48 00:02:25.800 --> 00:02:29.319 to tackle the Milky Way with true starbuy star fidelity. 49 00:02:29.560 --> 00:02:32.639 Okay, we have to unpack this because a one hundredfold 50 00:02:32.719 --> 00:02:36.639 jump in both speed and resolution, it sounds less like 51 00:02:36.719 --> 00:02:39.439 engineering and more like some kind of magic trick. We 52 00:02:39.520 --> 00:02:43.039 need to understand what the roadblock was exactly, and then 53 00:02:43.439 --> 00:02:47.120 how this team managed to combine artificial intelligence with traditional 54 00:02:47.120 --> 00:02:50.280 simulations to just bypass it completely. 55 00:02:50.639 --> 00:02:53.560 Absolutely, and you're right. To really appreciate the solution, we 56 00:02:53.639 --> 00:02:56.439 have to get deep into the problem because the limitation 57 00:02:57.039 --> 00:02:59.520 wasn't just a lack of computing power. It was a 58 00:02:59.560 --> 00:03:02.479 fundament old conflict baked into the physics itself. 59 00:03:02.599 --> 00:03:05.919 Let's start right there. Then, let's define this astronomical hurdle. 60 00:03:06.479 --> 00:03:09.479 Why was simulating the Milky Way star by star with 61 00:03:09.639 --> 00:03:12.840 just conventional methods considered practically impossible for so long? 62 00:03:12.960 --> 00:03:16.000 Well, the whole point of these efforts is pure scientific discovery, right, 63 00:03:16.199 --> 00:03:20.719 It's about understanding how galaxies evolve. Astrophysicists build these incredibly 64 00:03:20.759 --> 00:03:25.280 complex models of galaxy formation, structure, and how stars evolve 65 00:03:25.479 --> 00:03:27.000 so they can test their biggest theories. 66 00:03:27.159 --> 00:03:29.120 So they build a digital universe to see if it 67 00:03:29.159 --> 00:03:30.919 matches the real one exactly. 68 00:03:31.039 --> 00:03:34.599 They need a model that's detailed enough to compare its output, 69 00:03:35.080 --> 00:03:38.759 say the distribution of heavy elements, against what we actually 70 00:03:38.800 --> 00:03:42.199 see with telescopes like the Hubble or James Web. If 71 00:03:42.199 --> 00:03:44.919 the model doesn't match reality, then the theory is wrong 72 00:03:45.479 --> 00:03:46.599 or at least incomplete. 73 00:03:46.639 --> 00:03:48.719 And when you say an accurate model of a galaxy, 74 00:03:48.800 --> 00:03:50.840 we're not just talking about plotting points on a map. 75 00:03:50.879 --> 00:03:53.800 You're talking about modeling the entire engine of the system, 76 00:03:53.879 --> 00:03:55.199 all the physics. 77 00:03:54.919 --> 00:03:58.599 That's it is the definition of a multiphysics problem on 78 00:03:58.639 --> 00:04:02.319 a gargantuan scale. To get it right, the model has 79 00:04:02.319 --> 00:04:07.000 to account for multiple wildly complex interacting phenomena all at 80 00:04:07.039 --> 00:04:10.199 the same time, such as you've got gravity, the universal 81 00:04:10.199 --> 00:04:12.759 influence of it. Then you have the fluid dynamics of 82 00:04:12.800 --> 00:04:16.279 these huge turbulent clouds of gas and dust. You have 83 00:04:16.600 --> 00:04:20.680 catastrophic supernova explosions releasing immense energy, and on top of 84 00:04:20.720 --> 00:04:24.160 all that, the continuous process of creating new elements inside 85 00:04:24.199 --> 00:04:25.079 every single star. 86 00:04:25.279 --> 00:04:28.040 It sounds like trying to run a trillion different interconnected 87 00:04:28.079 --> 00:04:29.199 experiments all at once. 88 00:04:29.480 --> 00:04:30.959 That's a really good way to put it. And the 89 00:04:31.000 --> 00:04:32.959 hardest part of it all is what scientists call the 90 00:04:33.040 --> 00:04:34.040 scale disparity. 91 00:04:34.120 --> 00:04:35.079 Scale disparity. 92 00:04:35.399 --> 00:04:38.319 All these things are happening on vastly different scales of 93 00:04:38.319 --> 00:04:41.920 space and time. Think about it, gravity, which dictates the 94 00:04:41.920 --> 00:04:45.879 overall shape of the galaxy, the spiral arms. That's a huge, slow, 95 00:04:46.040 --> 00:04:49.040 large scale thing. It plays out over billions of years 96 00:04:49.040 --> 00:04:50.279 and thousands of light years. 97 00:04:50.360 --> 00:04:51.879 Okay, the big picture, right. 98 00:04:52.000 --> 00:04:54.920 But then you have a supernova that's a small, fast, 99 00:04:55.040 --> 00:04:57.839 local event. It plays out in a tiny region of 100 00:04:57.879 --> 00:05:00.839 the galaxy over maybe tens of thousands of years or 101 00:05:00.879 --> 00:05:04.160 even less a blink of an eye in cosmic time. 102 00:05:04.800 --> 00:05:08.079 And trying to cram those vastly different scales into one 103 00:05:08.399 --> 00:05:13.639 single cohesive model, well, that's what creates a paralyzing computational conflict. 104 00:05:13.800 --> 00:05:15.040 Let me see if I can wrap my head around 105 00:05:15.079 --> 00:05:16.680 this with an analogy. Let's say you want to make 106 00:05:16.680 --> 00:05:19.040 a video of a flower growing from a seed all 107 00:05:19.079 --> 00:05:20.959 the way to a full bloom that takes months. 108 00:05:21.040 --> 00:05:22.600 Okay, yeah, but while. 109 00:05:22.360 --> 00:05:25.439 You're filming that flower, a tiny hummingbird flashes past the 110 00:05:25.480 --> 00:05:28.120 frame in one hundredth of a second. If your goal 111 00:05:28.160 --> 00:05:30.240 is to capture both the slow growth of the flower 112 00:05:30.319 --> 00:05:32.680 and the precise rapid wing beats of the hummingbird with 113 00:05:32.759 --> 00:05:36.160 equal detail, the speed of your camera, your frames per 114 00:05:36.199 --> 00:05:39.360 second has to be dictated by the fastest event, by 115 00:05:39.360 --> 00:05:40.040 the hummingbird. 116 00:05:40.240 --> 00:05:42.439 That is precisely the problem that is the challenge for 117 00:05:42.480 --> 00:05:46.480 the supercomputer. The flower is the galaxy structure, slow and 118 00:05:46.519 --> 00:05:50.600 governed by gravity. The hummingbird is the supernova that localized 119 00:05:50.720 --> 00:05:53.920 violent explosion. If the model has to track what's happening 120 00:05:53.920 --> 00:05:56.480 locally at the speed of that explosion, it has to 121 00:05:56.600 --> 00:06:01.000 use impossibly short timesteps, little tiny snap shots in time. 122 00:06:00.920 --> 00:06:03.360 And doing that for the whole system is the killer. 123 00:06:03.480 --> 00:06:06.639 That's what killed it. Demanding those ultra short time steps 124 00:06:06.680 --> 00:06:09.920 across a system of one hundred billion stars is what 125 00:06:10.040 --> 00:06:12.319 made simulating the entire galaxy. 126 00:06:11.879 --> 00:06:14.480 Impossible, which leads us right to the limitations of the 127 00:06:14.480 --> 00:06:16.759 previous state of the art. If you couldn't run the 128 00:06:16.800 --> 00:06:19.360 whole galaxy with those short timesteps, you had to compromise 129 00:06:19.360 --> 00:06:21.240 on the detail, on the resolution. So what was the 130 00:06:21.240 --> 00:06:22.120 compromise the. 131 00:06:22.079 --> 00:06:25.360 Source material is very clear on this previous high resolution attempts, 132 00:06:26.120 --> 00:06:28.120 they had an upper mass limit of only about one 133 00:06:28.199 --> 00:06:29.279 billion sons. 134 00:06:29.319 --> 00:06:32.240 One billion when the Milky Way has over one hundred 135 00:06:32.279 --> 00:06:33.439 billion exactly. 136 00:06:33.560 --> 00:06:37.839 So they were forced into this will disastrous compromise. They 137 00:06:37.839 --> 00:06:41.480 had to sacrifice individual star resolution. They use what the 138 00:06:41.519 --> 00:06:45.480 researchers called the star cluster particle. The smallest unit in 139 00:06:45.519 --> 00:06:48.720 the model wasn't a single star. It was a cluster 140 00:06:48.800 --> 00:06:52.000 of stars, often weighing one hundred sons or more, all 141 00:06:52.040 --> 00:06:53.399 treated as a single particle. 142 00:06:53.560 --> 00:06:56.680 Wait a minute, If the whole point is starby star fidelity, 143 00:06:56.879 --> 00:06:59.399 how can you justify lumping one hundred sons together and 144 00:06:59.439 --> 00:07:02.680 calling it one thing. Doesn't that just defeat the purpose 145 00:07:02.720 --> 00:07:03.360 from the start. 146 00:07:03.560 --> 00:07:06.480 It does, and that's exactly why the fidelity was so low. 147 00:07:06.720 --> 00:07:09.040 Will you average out one hundred separate stars at one 148 00:07:09.120 --> 00:07:12.319 data point? You lose the ability to model the detailed 149 00:07:12.319 --> 00:07:15.720 physics of how stars actually evolve. What happens to the 150 00:07:15.720 --> 00:07:18.439 individual massive stars, the ones that live fast, die young, 151 00:07:18.519 --> 00:07:22.120 and go supernova that just gets completely lost. It's averaged 152 00:07:22.120 --> 00:07:22.800 out by the group. 153 00:07:22.879 --> 00:07:25.519 And soupernovad are critical, aren't they They're the element factories. 154 00:07:25.680 --> 00:07:29.839 They're everything. They are the primary way heavy elements get 155 00:07:29.879 --> 00:07:33.480 created and distributed. They're the raw ingredients for new stars, 156 00:07:33.480 --> 00:07:37.040 for planets, and ultimately for life. So if you're averaging 157 00:07:37.079 --> 00:07:39.959 out the physics of that explosion, because your smallest particle 158 00:07:39.959 --> 00:07:43.720 represents one hundred suns, you miss the crucial feedback loops. 159 00:07:44.079 --> 00:07:46.759 You might see the general shape of a spiral arm shore, 160 00:07:46.920 --> 00:07:51.120 but the precise local dynamics, the energy injection, the shockwaves, 161 00:07:51.199 --> 00:07:53.279 all of that is just lost to the averaging. 162 00:07:53.759 --> 00:07:56.439 So you get the big picture, the slow rotation of 163 00:07:56.480 --> 00:07:59.240 the galaxy, but you completely miss the details of how 164 00:07:59.360 --> 00:08:01.879 stars are born or how the space between stars gets 165 00:08:01.959 --> 00:08:05.439 enriched with new elements. It's a macromodel that can't answer 166 00:08:05.480 --> 00:08:06.360 the micro questions. 167 00:08:06.519 --> 00:08:09.199 It's exactly right, and the reason you can't have both 168 00:08:09.279 --> 00:08:13.240 high resolution and the full galaxy. It all circles back 169 00:08:13.240 --> 00:08:15.920 to that time penalty. If you try to use short 170 00:08:15.920 --> 00:08:18.560 time steps to see those fast local changes at the 171 00:08:18.600 --> 00:08:21.279 star level, you have to recalculate the physics of one 172 00:08:21.360 --> 00:08:25.720 hundred billion stars and all the gas way way more often, 173 00:08:25.839 --> 00:08:29.879 the computational cost just grows. Well, it's faster than exponentially 174 00:08:29.920 --> 00:08:33.039 as you crank up the resolution. High resolution demands short 175 00:08:33.080 --> 00:08:36.919 time steps, which exponentially increases the work, pushing the total 176 00:08:37.000 --> 00:08:40.480 run time beyond anything feasible. They were trapped, completely trapped 177 00:08:40.480 --> 00:08:43.120 in a classic computational trade off. You could have high 178 00:08:43.120 --> 00:08:45.639 resolution over a tiny piece of the galaxy or low 179 00:08:45.679 --> 00:08:48.600 resolution over the whole thing, but you absolutely could not 180 00:08:48.639 --> 00:08:49.679 have both at the same time. 181 00:08:49.759 --> 00:08:51.639 So they needed a way to satisfy the demands of 182 00:08:51.679 --> 00:08:54.960 the big slow gravity part of the simulation while still 183 00:08:55.000 --> 00:08:58.519 getting the results of the small fast supernova part, but 184 00:08:58.600 --> 00:09:01.799 without actually having to calculate the supernova physics from scratch. 185 00:09:02.039 --> 00:09:04.360 For every single star that explodes. 186 00:09:04.039 --> 00:09:06.919 And that leads us directly to the cold hard numbers, 187 00:09:06.960 --> 00:09:09.759 to the computational wall they hit. Let's quantify what this 188 00:09:09.840 --> 00:09:12.759 trade off really meant in terms of real world time. 189 00:09:13.039 --> 00:09:16.879 This is where the scale of the challenge just becomes prohibitive. 190 00:09:17.000 --> 00:09:19.879 Right If a team wanted to model the Milky Way 191 00:09:19.919 --> 00:09:23.360 star by star using the best conventional methods before this breakthrough, 192 00:09:24.039 --> 00:09:26.399 what was the actual cost in computer time? 193 00:09:26.759 --> 00:09:31.120 The barrier was well astronomical. Literally, if you took the 194 00:09:31.159 --> 00:09:34.519 best traditional physical simulation and tried to run it with 195 00:09:34.559 --> 00:09:37.639 the high resolution needed for individual stars. Yeah, it would 196 00:09:37.639 --> 00:09:40.399 require three hundred and fifteen hours of compute time, three 197 00:09:40.480 --> 00:09:43.000 hundred and fifteen hours for every one million years of 198 00:09:43.039 --> 00:09:44.200 simulated galactic history. 199 00:09:44.240 --> 00:09:46.799 Okay, three hundred and fifteen hours for a million years. 200 00:09:46.840 --> 00:09:49.360 That seems huge, but it's hard to grasp until you 201 00:09:49.399 --> 00:09:52.120 put it in the right context. We're not simulating just 202 00:09:52.120 --> 00:09:55.039 a million years. How long do astrophysicists need to model 203 00:09:55.080 --> 00:09:56.559 to see anything meaningful happen? 204 00:09:56.879 --> 00:09:59.639 To see the big structural changes things like spiral arms 205 00:09:59.639 --> 00:10:02.399 forma or stars migrating, or the long term impact of 206 00:10:02.399 --> 00:10:05.840 all that supernova feedback, researchers really need to simulate at 207 00:10:05.919 --> 00:10:08.679 least a billion years of evolution. A billion Okay, So 208 00:10:08.840 --> 00:10:10.879 now let's scale up those three hundred and fifteen hours. 209 00:10:10.919 --> 00:10:13.120 A billion is one thousand millions, so one thousand times 210 00:10:13.120 --> 00:10:16.840 three hundred and fifteen hours. That's simulating that crucial one 211 00:10:17.000 --> 00:10:21.080 billion year timeframe. Using the old methods would take more 212 00:10:21.120 --> 00:10:25.480 than thirty six years of continuous, dedicated real world compute time. 213 00:10:25.799 --> 00:10:28.879 Thirty six years, thirty six years. That's an entire career. 214 00:10:29.039 --> 00:10:33.080 That's a fundamental barrier to discovery. No research grant, no 215 00:10:33.200 --> 00:10:37.039 PhD program, no single research group, can commit to running 216 00:10:37.120 --> 00:10:39.919 one test for three and a half decades. 217 00:10:39.480 --> 00:10:42.519 The hardware would be obsolete three times over before the 218 00:10:42.559 --> 00:10:44.080 simulation even finished running. 219 00:10:44.240 --> 00:10:46.360 It wasn't just difficult, then, it was a hard stop. 220 00:10:46.399 --> 00:10:48.960 It was a hard stop on certain lines of scientific inquiry. 221 00:10:49.559 --> 00:10:52.600 If you have a brilliant new theory about, say, how 222 00:10:52.679 --> 00:10:56.320 star clusters form, and testing it requires a billion year 223 00:10:56.480 --> 00:11:01.200 high resolution simulation, You're just blocked. You can't do it. Yeah, 224 00:11:01.200 --> 00:11:04.200 you're forced to use those low resolution star cluster particles, 225 00:11:04.480 --> 00:11:06.960 which severely limits the kinds of questions you can even 226 00:11:07.000 --> 00:11:08.000 ask in the first place. 227 00:11:08.080 --> 00:11:11.000 So the obvious and maybe crude answer that usually works 228 00:11:11.000 --> 00:11:14.639 in computing is just throwing even bigger supercomputer at it. 229 00:11:14.720 --> 00:11:17.399 Build a bigger machine, throw ten times the cores of 230 00:11:17.480 --> 00:11:21.080 the problem. And wait, why wasn't that a viable solution here? 231 00:11:21.279 --> 00:11:25.360 Because of the principle of diminishing returns and specifically a 232 00:11:25.399 --> 00:11:30.360 crippling problem called communication overhead. Okay, building a bigger supercomputer 233 00:11:31.120 --> 00:11:35.159 solve the raw processing power issue, sure, but it doesn't 234 00:11:35.159 --> 00:11:38.480 solve the efficiency problem of keeping all those processors in sync. 235 00:11:39.440 --> 00:11:41.840 Think about how these things work. You take your one 236 00:11:41.879 --> 00:11:44.759 hundred billion stars and all the gas and you distribute 237 00:11:44.799 --> 00:11:48.240 the calculations across millions of process or cores or nodes. 238 00:11:49.039 --> 00:11:52.480 But here's the catch. Every single core needs to know 239 00:11:52.519 --> 00:11:55.279 what every other core is doing at every single timestep. 240 00:11:55.279 --> 00:11:57.120 There's a constant chatter between them. 241 00:11:57.039 --> 00:12:00.000 A constant massive flow of data to synchronize the physical 242 00:12:00.200 --> 00:12:03.200 state of the galaxy. So even if each individual core 243 00:12:03.320 --> 00:12:06.159 is calculating its little piece of a puzzle incredibly quickly, 244 00:12:06.519 --> 00:12:08.360 the system as a whole has to wait for everyone 245 00:12:08.360 --> 00:12:09.679 to catch up and share their results. 246 00:12:09.759 --> 00:12:11.919 So the whole system moves at the speed of the 247 00:12:12.000 --> 00:12:14.120 slowest link in the chain, precisely. 248 00:12:14.759 --> 00:12:17.320 And as you add more and more cores, the amount 249 00:12:17.360 --> 00:12:20.440 of time the system spends just communicating between them, managing 250 00:12:20.480 --> 00:12:24.679 that data flow, synchronizing the calculations, it starts to outweigh 251 00:12:25.120 --> 00:12:28.039 the time spent actually doing the physics. The system gets 252 00:12:28.080 --> 00:12:29.559 bogged down in its own overhead. 253 00:12:29.720 --> 00:12:31.919 It's like having a meeting with a million people. You'd 254 00:12:31.960 --> 00:12:33.759 spend all your time just trying to get everyone on 255 00:12:33.799 --> 00:12:35.919 the same page, and no time making decisions. 256 00:12:36.039 --> 00:12:39.639 That's a perfect analogy. The synchronization costs is the bottleneck. 257 00:12:40.200 --> 00:12:43.960 Adding more people or more cores just increases the noise 258 00:12:44.039 --> 00:12:46.919 and the waiting time, not the actual speed of the work. 259 00:12:47.440 --> 00:12:49.840 So simply scaling up the hardware was not a path 260 00:12:49.879 --> 00:12:52.279 out of that thirty six year weight. They needed a 261 00:12:52.320 --> 00:12:54.759 paradigm shift, not just a hardware upgrade. 262 00:12:54.840 --> 00:12:57.720 The conventional equations, no matter how fast you ran them, 263 00:12:57.879 --> 00:13:00.720 demanded that you calculate every tiny, dear detail of the 264 00:13:00.759 --> 00:13:04.159 fastest event across the whole system, at every single snapshot. 265 00:13:04.240 --> 00:13:06.360 That was the core inefficiency correct. 266 00:13:06.039 --> 00:13:08.759 The computational wall wasn't just a lack of transistors, It 267 00:13:08.799 --> 00:13:12.759 was a fundamental efficiency problem in the method itself, trying 268 00:13:12.759 --> 00:13:17.399 to model these vastly different time scales simultaneously with rigid 269 00:13:17.519 --> 00:13:18.960 traditional physics equations. 270 00:13:19.200 --> 00:13:22.360 And this brings us to the game changing solution, the 271 00:13:22.399 --> 00:13:25.440 AI shortcut. This is where the innovation team comes in 272 00:13:25.480 --> 00:13:28.440 and essentially tells the computer, you don't need to calculate 273 00:13:28.440 --> 00:13:30.879 this from scratch. Every time we've taught an AI, the 274 00:13:30.960 --> 00:13:34.399 answer tell us about the core methodology they used. 275 00:13:34.840 --> 00:13:38.080 So the team led by Hiroshima at Reichen's Ethem Center, 276 00:13:38.320 --> 00:13:41.000 along with collaborators from the University of Tokyo and University 277 00:13:41.039 --> 00:13:44.279 tak Day Barcelona, they focused on one thing, breaking the 278 00:13:44.320 --> 00:13:47.919 dependency between the fast small scale physics and the slow 279 00:13:48.360 --> 00:13:51.120 large scale physics. Their core idea is to combine the 280 00:13:51.159 --> 00:13:55.200 high fidelity traditional physical simulations for the large scale with 281 00:13:55.279 --> 00:13:58.559 a specialized AI technique known as the deep learning surrogate 282 00:13:58.600 --> 00:14:00.000 model for the small scale. 283 00:14:00.279 --> 00:14:02.919 Okay, the term surrogate model is key here. It's not 284 00:14:03.039 --> 00:14:06.720 replacing the entire physics simulation, right, It's replacing just one component, 285 00:14:06.840 --> 00:14:08.759 the most expensive component exactly. 286 00:14:08.799 --> 00:14:10.639 And which component do you think they targeted? 287 00:14:10.840 --> 00:14:13.559 The hummingbird? The supernova. 288 00:14:13.919 --> 00:14:18.120 The supernova, as we discussed, that's the primary bottleneck. It's rapid, 289 00:14:18.320 --> 00:14:22.360 it's localized, it has incredibly complex fluidynamics, and it demands 290 00:14:22.399 --> 00:14:26.120 those tiny timesteps that paralyze the main simulation. When a 291 00:14:26.159 --> 00:14:29.440 massive star dies, the gas around it is shocked, heated, 292 00:14:29.600 --> 00:14:33.879 and expands rapidly. Calculating that complex feedback with conventional methods 293 00:14:34.000 --> 00:14:36.080 is what was costing three hundred and fifteen hours for 294 00:14:36.159 --> 00:14:38.080 every million years of simulated time. 295 00:14:38.279 --> 00:14:40.360 So how do you train an AI to solve a 296 00:14:40.360 --> 00:14:42.480 problem like that, a fluid dynamics problem. 297 00:14:42.720 --> 00:14:46.759 Well, you leverage existing high resolution simulations. Before this breakthrough, 298 00:14:46.879 --> 00:14:50.360 researchers could run incredibly detailed, short simulations of a single 299 00:14:50.399 --> 00:14:54.120 supernova explosion just in a small isolated box, not inside 300 00:14:54.120 --> 00:14:55.639 a whole hundred billion star galaxy. 301 00:14:55.679 --> 00:14:57.799 Right. They could do the micro just not the micro 302 00:14:57.879 --> 00:14:59.279 and macro together, right. 303 00:14:59.440 --> 00:15:02.519 So the team took terabides of data from these ultra 304 00:15:02.639 --> 00:15:07.360 high resolution standalone supernova simulations and fit it all to 305 00:15:07.399 --> 00:15:11.279 a jeep learning model. The AI wasn't trained to solve 306 00:15:11.279 --> 00:15:14.960 the differential equations of fluid dynamics itself. It was trained 307 00:15:15.000 --> 00:15:18.919 to learn the input output relationship. Basically, given these starting 308 00:15:18.960 --> 00:15:22.519 conditions gas density, stellar mass, et cetera, what will the 309 00:15:22.519 --> 00:15:25.519 precise state of the surrounding gas be ten thousand years, 310 00:15:25.840 --> 00:15:29.559 fifty thousand years, and one hundred thousand years after the explosion. 311 00:15:30.039 --> 00:15:33.960 So it's pattern recognition, but for physical outcomes. The AI 312 00:15:34.120 --> 00:15:36.600 learned to predict the result of a month of calculation 313 00:15:36.799 --> 00:15:38.039 in just a few milliseconds. 314 00:15:38.039 --> 00:15:40.639 That's it exactly. That is its predictive power. The deep 315 00:15:40.720 --> 00:15:44.279 learning model learned to accurately predict how the gas expands 316 00:15:44.480 --> 00:15:46.799 and how energy is injected back into the galaxy over 317 00:15:46.799 --> 00:15:49.399 that critical one hundred thousand year period after a supernova. 318 00:15:49.960 --> 00:15:53.799 This allowed the main simulation to bypass millions of repetitive, 319 00:15:53.879 --> 00:15:56.879 tiny fluid dynamics calculations every single time a star died. 320 00:15:57.039 --> 00:16:00.320 That makes the computational advantage crystal clear. Instead of the 321 00:16:00.320 --> 00:16:03.879 main supercomputer grinding to a halt to calculate a shockwave 322 00:16:03.879 --> 00:16:06.200 for one hundred thousand years the old way, it just 323 00:16:07.279 --> 00:16:11.159 it flags the event. The AI surrogate model spits out 324 00:16:11.240 --> 00:16:14.639 the accurate high resolution outcome in a fraction of a. 325 00:16:14.559 --> 00:16:18.240 Second, and the main simulation just inserts that result and 326 00:16:18.320 --> 00:16:21.559 keeps on chugging along with its large scale calculations. Brilliant 327 00:16:21.679 --> 00:16:24.440 it is the main simulation can now take these big, 328 00:16:24.519 --> 00:16:28.000 comfortable time steps based on the slow pace of galactic gravity. 329 00:16:28.440 --> 00:16:30.519 It would only have to momentarily check in with the 330 00:16:30.559 --> 00:16:34.519 AI to inject the necessary fine scale physics the supernova 331 00:16:34.559 --> 00:16:38.799 feedback as needed. This ability to delegate the toughest computational 332 00:16:38.840 --> 00:16:42.000 lift is what unlocked the one hundredfold speed boost. They 333 00:16:42.080 --> 00:16:44.840 kept the high resolution detail because the AI is providing that, 334 00:16:44.960 --> 00:16:46.960 but they completely avoided the time penalty. 335 00:16:47.039 --> 00:16:50.279 The strategic delegation keeping the traditional physics for the macro 336 00:16:50.399 --> 00:16:53.639 structure and using AI for the microphysics. That feels like 337 00:16:53.639 --> 00:16:56.840 a fundamental re architecture of how we approach computational science. 338 00:16:57.240 --> 00:16:59.799 It is, But of course the very next question has 339 00:16:59.799 --> 00:17:03.480 to be how can you trust it? How can you 340 00:17:03.519 --> 00:17:06.599 trust a result provided by an AI shortcut? Right? 341 00:17:06.880 --> 00:17:09.920 If the model is predicting outcomes instead of calculating them 342 00:17:09.960 --> 00:17:12.440 from first principles, how do you know the physics is 343 00:17:12.440 --> 00:17:13.079 still valid? 344 00:17:13.279 --> 00:17:16.359 Trust is everything here, and the team knew that they 345 00:17:16.400 --> 00:17:19.720 put the model through a rigorous verification process. They didn't 346 00:17:19.759 --> 00:17:22.640 just accept the AI's output as gospel. They ran their 347 00:17:22.680 --> 00:17:26.319 new AI accelerated simulation and then compared its results against 348 00:17:26.359 --> 00:17:29.000 large scale tests using some of the world's most powerful 349 00:17:29.039 --> 00:17:30.519 conventional supercomputers. 350 00:17:30.640 --> 00:17:32.920 So they ran the old slow method on a smaller 351 00:17:32.960 --> 00:17:35.400 scale to check the AI's work exactly. 352 00:17:35.519 --> 00:17:39.039 They used Reichen's own Fugaku supercomputer, one of the fastest 353 00:17:39.039 --> 00:17:41.720 machines on the planet, and the University of Tokyo's Miami 354 00:17:41.759 --> 00:17:46.119 supercomputer system. They'd run test simulations of smaller galaxies or 355 00:17:46.160 --> 00:17:50.240 specific regions using the old full physics calculations, and the 356 00:17:50.240 --> 00:17:54.039 comparison showed that the AI surrogate model was indeed accurately 357 00:17:54.079 --> 00:17:57.599 representing the physics it was designed to emulate. That verification 358 00:17:57.720 --> 00:18:00.880 step was critical. It validated the AI as a genuine 359 00:18:00.920 --> 00:18:04.160 physics tool, not just a data analysis engine. It proved 360 00:18:04.160 --> 00:18:06.240 the shortcut maintained scientific fidelity. 361 00:18:06.359 --> 00:18:08.920 Okay, so after validating the physics we get to the payoff. 362 00:18:09.119 --> 00:18:11.559 Let's talk about the breakthrough results and what this means 363 00:18:11.559 --> 00:18:14.680 for the science of the cosmos and maybe for things 364 00:18:14.680 --> 00:18:15.160 beyond that. 365 00:18:15.440 --> 00:18:19.200 We can finally put that transformative speed increase into concrete terms. 366 00:18:19.440 --> 00:18:22.240 Remember the old timeline three hundred and fifteen hours of 367 00:18:22.240 --> 00:18:26.039 compute time just to simulate one million years of galactic evolution. 368 00:18:25.839 --> 00:18:28.400 The number that was the single biggest inhibitor to this 369 00:18:28.480 --> 00:18:29.079 kind of research. 370 00:18:29.279 --> 00:18:32.039 That time requirement dropped from three hundred and fifteen hours 371 00:18:32.319 --> 00:18:34.079 to only two point seven eight hours. 372 00:18:34.319 --> 00:18:39.000 Wow, two point seven eight that is just an unbelievable acceleration. 373 00:18:39.079 --> 00:18:41.720 You're trading nearly two weeks of waiting for a single 374 00:18:41.799 --> 00:18:45.559 data point for less than three hours. That changes the 375 00:18:45.720 --> 00:18:47.400 entire flow of how you do research. 376 00:18:47.480 --> 00:18:49.359 And now let's scale that up to the big prize 377 00:18:49.400 --> 00:18:53.200 that one billion years of galaxy evolution. The waytime collapses 378 00:18:53.240 --> 00:18:56.119 from thirty six years to something actually. 379 00:18:55.759 --> 00:18:57.680 Manageable, don't leave us in suspense. 380 00:18:58.000 --> 00:19:00.480 The thirty six year simulation time would, which was an 381 00:19:00.519 --> 00:19:03.920 impossibility for any research group, can now be accomplished in 382 00:19:03.960 --> 00:19:05.720 a mere one hundred and fifteen days. 383 00:19:05.519 --> 00:19:08.400 One hundred and fifteen days less than four months. That is, 384 00:19:08.720 --> 00:19:11.880 that's truly incredible. A researcher can now propose a fundamental 385 00:19:11.960 --> 00:19:15.039 question about how galaxies form, start the simulation, when their 386 00:19:15.079 --> 00:19:17.559 grant gets approved, and how the complete results before the 387 00:19:17.599 --> 00:19:18.880 academic year is even over. 388 00:19:19.279 --> 00:19:25.000 It completely transforms theoretical astrophysics from this multigenerational pursuit into 389 00:19:25.000 --> 00:19:29.400 a rapid response science. It absolutely does. The scientific impact 390 00:19:29.480 --> 00:19:33.720 is finally the ability to achieve individual star resolution across 391 00:19:33.759 --> 00:19:40.480 an entire large realistic galaxy with over one hundred billion stars. Previously, 392 00:19:40.480 --> 00:19:43.480 we had models of galactic structure, or we had models 393 00:19:43.519 --> 00:19:47.839 of stellar evolution, but never a single comprehensive model that 394 00:19:47.920 --> 00:19:49.240 linked the two in real time. 395 00:19:49.440 --> 00:19:51.559 So now you can see the whole system working together. 396 00:19:51.680 --> 00:19:54.519 Now researchers can run these simulations where they can literally 397 00:19:54.559 --> 00:19:57.839 trace the path of individual stars, watch them explode, and 398 00:19:57.880 --> 00:20:00.640 then observe precisely how the energy and elements from that 399 00:20:00.680 --> 00:20:05.039 explosion interact with the surrounding gas clouds, influencing the birth 400 00:20:05.039 --> 00:20:08.000 of the next generation of stars. Yeah, this level of detail, 401 00:20:08.079 --> 00:20:11.680