1 00:00:00,040 --> 00:00:02,280 Welcome back to the Deep Dive. We are. 2 00:00:02,640 --> 00:00:05,240 We're doing something a little specialized today and frankly, I 3 00:00:05,240 --> 00:00:07,120 think it's about time we did. I agree. 4 00:00:07,440 --> 00:00:11,000 We're looking at a pretty fundamental pivot in how well 5 00:00:11,240 --> 00:00:14,760 how humanity invents things, how we create new materials. 6 00:00:14,960 --> 00:00:17,880 It's a huge topic and our source material for this is really 7 00:00:17,880 --> 00:00:19,520 interesting. We're looking at a lecture from 8 00:00:19,520 --> 00:00:21,480 Shiki Wang, right? It was delivered at the 9 00:00:21,480 --> 00:00:25,080 University of Helsinki for Professor Pedro Camargo's 10 00:00:25,080 --> 00:00:27,320 course. And the date on the material is 11 00:00:27,320 --> 00:00:30,840 February 13, 2026. So yeah, we're getting a little 12 00:00:30,840 --> 00:00:34,160 glimpse of the very near future. We're basically stepping into a 13 00:00:34,160 --> 00:00:36,920 university classroom, which is a fun change of pace for us. 14 00:00:37,240 --> 00:00:39,080 The title is, and it's a mouthful. 15 00:00:39,240 --> 00:00:42,160 Development and application of computational methods on 16 00:00:42,160 --> 00:00:43,920 catalysis research. Right. 17 00:00:44,200 --> 00:00:46,600 And I know as soon as you say catalysis, you can almost hear 18 00:00:46,600 --> 00:00:50,480 people's eyes glazing over. It sounds so, so high school 19 00:00:50,480 --> 00:00:53,480 chemistry. It does, but The thing is, this 20 00:00:53,480 --> 00:00:56,280 is actually about the engine of the entire modern world. 21 00:00:56,440 --> 00:00:59,440 I mean literally everything. The fuel in your car, the 22 00:00:59,440 --> 00:01:02,720 medicine you take, the plastic in your phone case, it all 23 00:01:02,720 --> 00:01:06,480 exists because of catalysis. It's the secret ingredient that 24 00:01:06,480 --> 00:01:08,480 makes things happen. Exactly. 25 00:01:08,680 --> 00:01:11,600 And the computational part of that title, that's the real 26 00:01:11,600 --> 00:01:14,040 story here. We are witnessing the end, or 27 00:01:14,040 --> 00:01:17,520 maybe the beginning of the end of the shake and bake era of 28 00:01:17,520 --> 00:01:19,360 chemistry. The shake and bake era. 29 00:01:19,360 --> 00:01:21,400 I love that. That's the Edisonian method, 30 00:01:21,400 --> 00:01:23,080 right? Trial and error 100. 31 00:01:23,400 --> 00:01:25,240 I have a beaker. I have a hunch. 32 00:01:25,280 --> 00:01:28,160 I mix some things together and I just, you know, I hope nothing 33 00:01:28,160 --> 00:01:29,720 explodes. Precisely. 34 00:01:30,120 --> 00:01:32,000 And for the last, what, 200 years? 35 00:01:32,280 --> 00:01:34,880 That's genuinely how we found almost everything. 36 00:01:35,000 --> 00:01:38,360 It's incredible to think about. From penicillin to Teflon, it 37 00:01:38,360 --> 00:01:42,000 was largely based on educated guessing, a lot of serendipity, 38 00:01:42,160 --> 00:01:43,640 and frankly, a whole lot of. Luck. 39 00:01:43,640 --> 00:01:47,040 But Shiki Wang's lecture makes it painfully clear that that 40 00:01:47,040 --> 00:01:49,440 method is, well, it's hitting a hard wall. 41 00:01:49,560 --> 00:01:52,160 It is. It's just too slow, it's way too 42 00:01:52,160 --> 00:01:54,800 expensive and honestly all the low hanging fruit has already 43 00:01:54,800 --> 00:01:57,040 been picked. We can't rely on luck anymore 44 00:01:57,040 --> 00:01:58,720 and the. Stats and the source material to 45 00:01:58,720 --> 00:02:02,120 back this up are they're wild. I mean, they mentioned that 46 00:02:02,120 --> 00:02:05,080 moving from that traditional trial and error to what they 47 00:02:05,080 --> 00:02:09,039 call theory guided experiments. So using computers to guide you. 48 00:02:09,039 --> 00:02:14,400 Exactly that simple shift cuts the development process by over 49 00:02:14,400 --> 00:02:18,480 60%. 60% just stop and think about what that means in 50 00:02:18,480 --> 00:02:19,560 practice. It's. 51 00:02:19,560 --> 00:02:22,360 Huge. If it takes, say, 10 years to 52 00:02:22,360 --> 00:02:26,440 bring a new battery material to market, which is pretty 53 00:02:26,440 --> 00:02:29,800 standard, cutting that by 60% means you're doing it in four 54 00:02:29,800 --> 00:02:31,000 years. You're getting it six years 55 00:02:31,000 --> 00:02:32,720 earlier. That's the difference between 56 00:02:32,960 --> 00:02:36,320 maybe solving a major part of the climate crisis and 57 00:02:36,320 --> 00:02:39,160 completely missing the boat. It's a revolutionary speed up. 58 00:02:39,160 --> 00:02:41,840 And then there's the cost metric, which let's be honest, 59 00:02:41,840 --> 00:02:44,080 is probably the real clincher for the industry. 60 00:02:44,080 --> 00:02:45,560 Oh, absolutely. It always comes down to the 61 00:02:45,560 --> 00:02:47,840 money. The lecture notes state that a 62 00:02:47,840 --> 00:02:51,520 single calculation using DfT, that's density functional 63 00:02:51,520 --> 00:02:53,800 theory, and we will get deep into what that is. 64 00:02:54,360 --> 00:02:58,280 It costs roughly one one thousandth of the equivalent 65 00:02:58,280 --> 00:03:02,040 experimental expense. 1000 I mean, that's not an incremental 66 00:03:02,040 --> 00:03:03,600 improvement. No, it's a phase shift. 67 00:03:03,600 --> 00:03:05,680 It's the difference between buying a house and buying a cup 68 00:03:05,680 --> 00:03:07,120 of coffee. It's just not on the same 69 00:03:07,120 --> 00:03:08,920 planet. And that completely changes the 70 00:03:08,920 --> 00:03:10,760 risk profile for research, doesn't it? 71 00:03:11,000 --> 00:03:12,880 How so? Well, if I'm running a physical 72 00:03:12,880 --> 00:03:15,840 lab, every single experiment burns cash. 73 00:03:16,640 --> 00:03:19,040 You've got the chemicals, the waste disposal, the equipment, 74 00:03:19,040 --> 00:03:21,480 time, the personnel. The energy costs, which are 75 00:03:21,480 --> 00:03:22,240 massive. Right. 76 00:03:22,240 --> 00:03:25,160 So I have to be conservative. I can really only afford to test 77 00:03:25,160 --> 00:03:26,960 the things I'm pretty sure are going to work. 78 00:03:27,120 --> 00:03:30,520 You stick to what you know. You make small changes to proven 79 00:03:30,520 --> 00:03:32,400 formulas. Exactly. 80 00:03:33,160 --> 00:03:37,520 But if the cost of an experiment drops by a factor of 1000. 81 00:03:37,640 --> 00:03:39,240 Then you can test the crazy ideas. 82 00:03:39,240 --> 00:03:41,280 You can test the things that shouldn't work just on the off 83 00:03:41,280 --> 00:03:43,880 chance they do. You can explore the entire 84 00:03:44,160 --> 00:03:47,480 chemical universe, as they call it, rather than just, you know, 85 00:03:47,480 --> 00:03:49,160 the little neighborhood you're already familiar with. 86 00:03:49,160 --> 00:03:51,680 It's like going from a map of your town to Google Earth. 87 00:03:51,840 --> 00:03:54,840 That's a perfect analogy, and that leads us right into the 88 00:03:54,840 --> 00:03:58,360 first big theme of this lecture, which is just the why. 89 00:03:58,560 --> 00:04:01,600 Why are we doing all this? And aside from the speed and the 90 00:04:01,600 --> 00:04:04,440 money, it's about overcoming our own human limitations. 91 00:04:04,880 --> 00:04:08,320 The lecture highlights overcoming experimental limits 92 00:04:08,320 --> 00:04:10,880 as a primary driver. This is so crucial. 93 00:04:11,000 --> 00:04:13,960 There's a fundamental resolution problem in physical chemistry. 94 00:04:14,360 --> 00:04:16,760 You know, we talk about atoms like the little Legos we can 95 00:04:16,760 --> 00:04:20,480 move around, but in a real experiment, you can't just you 96 00:04:20,480 --> 00:04:22,440 can't just look at a single atom and watch what it does. 97 00:04:22,440 --> 00:04:24,000 No, you're always looking at averages. 98 00:04:24,000 --> 00:04:26,600 Averages of trillions upon trillions of atoms all doing 99 00:04:26,600 --> 00:04:28,480 something slightly different. Exactly. 100 00:04:28,800 --> 00:04:31,320 But all the interesting stuff, the chemistry, the reason a 101 00:04:31,320 --> 00:04:34,680 reaction works or fails, it happens at the margins. 102 00:04:34,720 --> 00:04:37,560 It happens with that one weird atom on the corner of a 103 00:04:37,560 --> 00:04:39,200 nanoparticle. And you can't see that. 104 00:04:39,320 --> 00:04:42,560 You can't. The source specifically mentions 105 00:04:42,680 --> 00:04:46,920 resolving atomic scale dynamics, and they give the example of 106 00:04:46,920 --> 00:04:49,800 dislocation motion. Now for the material scientists 107 00:04:49,800 --> 00:04:52,760 listening, you know the dislocation motion is 108 00:04:52,840 --> 00:04:55,520 everything. It's how metals bend instead of 109 00:04:55,520 --> 00:04:57,440 snapping. It's how things fail. 110 00:04:57,440 --> 00:04:58,760 But it. Happens at what? 111 00:04:58,800 --> 00:05:03,120 The speed of sound inside a solid block of metal at the 112 00:05:03,120 --> 00:05:06,080 scale of a nanometer? You can't exactly stick a GoPro 113 00:05:06,080 --> 00:05:07,160 in there to see what's happening. 114 00:05:07,160 --> 00:05:10,400 I really, really can't. But in a simulation in a 115 00:05:10,400 --> 00:05:12,480 computer model, you're basically. 116 00:05:13,280 --> 00:05:16,800 You control the universe. You can pause time, you can 117 00:05:16,800 --> 00:05:20,200 rewind it, you can strip away the outer layers of the metal 118 00:05:20,200 --> 00:05:23,040 and justice look at the core. You can see the exact moment a 119 00:05:23,040 --> 00:05:27,320 bond breaks or a new one forms. That kind of visibility is 120 00:05:27,320 --> 00:05:30,520 something no physical experiment on Earth can ever give you. 121 00:05:30,760 --> 00:05:34,200 And the other limitation the lecture brings up is the 122 00:05:34,200 --> 00:05:36,400 environment, what they call the dangerous factor. 123 00:05:36,400 --> 00:05:39,320 Right, things that are too hot, too high pressure or too 124 00:05:39,320 --> 00:05:42,760 radioactive to study easily. And the example they use is 125 00:05:42,760 --> 00:05:45,880 predicting irradiation damage in nuclear reactors. 126 00:05:45,880 --> 00:05:49,120 A classic problem, and a super important one if you want to 127 00:05:49,120 --> 00:05:52,600 build a fusion reactor. The dream of clean energy. 128 00:05:53,040 --> 00:05:55,440 You need to design materials that can withstand a constant 129 00:05:55,440 --> 00:05:59,200 bombardment of high energy neutrons that would basically 130 00:05:59,200 --> 00:06:01,160 disintegrate almost anything we can currently. 131 00:06:01,160 --> 00:06:04,200 Make and you can't exactly build a test fusion reactor just to 132 00:06:04,200 --> 00:06:06,080 see if your new wall material melts. 133 00:06:06,080 --> 00:06:08,600 Well, you can, but it costs billions of dollars and takes a 134 00:06:08,600 --> 00:06:11,320 decade or more. Or you use computational 135 00:06:11,320 --> 00:06:13,680 modeling. You build the wall inside the 136 00:06:13,680 --> 00:06:16,840 supercomputer. You simulate a single neutron 137 00:06:16,840 --> 00:06:19,440 hitting the atomic lattice. You calculate the energy 138 00:06:19,440 --> 00:06:22,000 transfer. You watch the cascade of damage 139 00:06:22,240 --> 00:06:24,920 as that one neutron knocks other atoms out of place. 140 00:06:24,920 --> 00:06:28,200 And you can fail 1000 times in the simulation. 141 00:06:28,200 --> 00:06:30,440 A million times. And you do it without ever 142 00:06:30,440 --> 00:06:32,920 creating a single speck of radioactive waste. 143 00:06:33,120 --> 00:06:34,720 We're spending billions of dollars. 144 00:06:34,760 --> 00:06:37,400 So it's safer, it sees the invisible, and it's 145 00:06:37,400 --> 00:06:40,160 astronomically cheaper. That's a pretty compelling case. 146 00:06:40,320 --> 00:06:42,240 But then there's the acceleration aspect. 147 00:06:42,240 --> 00:06:44,760 Yes, the lecture talks about high throughput screening. 148 00:06:44,880 --> 00:06:46,840 This is the brute force approach. 149 00:06:46,840 --> 00:06:48,360 Right. It is brute force, but it's 150 00:06:48,360 --> 00:06:50,640 smart brute force. I mean, think about it. 151 00:06:50,640 --> 00:06:53,800 In the old days, a PhD student might synthesize and 152 00:06:53,800 --> 00:06:56,880 characterize maybe maybe one new material a week. 153 00:06:57,000 --> 00:06:59,640 If they were lucky and all the equipment was working and 154 00:06:59,640 --> 00:07:02,120 nothing broke, right? And they didn't contaminate 155 00:07:02,120 --> 00:07:04,520 their sample. High throughput screening is 156 00:07:04,520 --> 00:07:07,560 more like a machine gun. You're not firing one bullet, 157 00:07:07,640 --> 00:07:11,640 you're firing off thousands or 10s of thousands of calculations 158 00:07:11,640 --> 00:07:13,760 simultaneously on a supercomputer cluster. 159 00:07:14,040 --> 00:07:17,240 And the lecturer sites a specific MIT study that really 160 00:07:17,240 --> 00:07:19,160 just it illustrates this perfectly. 161 00:07:20,080 --> 00:07:24,920 They use this method to look for new electrolytes for lithium ion 162 00:07:24,920 --> 00:07:27,920 batteries. And they didn't just find a few, 163 00:07:28,280 --> 00:07:31,920 they identified 21 promising new candidates. 164 00:07:32,280 --> 00:07:35,080 And these weren't just, you know, a 5% improvement. 165 00:07:35,320 --> 00:07:37,520 The source says they found candidates that could lead to 166 00:07:37,520 --> 00:07:41,680 100 times efficiency gains 100 X. 167 00:07:41,920 --> 00:07:44,840 And here's the kicker, the part that really matters. 168 00:07:45,680 --> 00:07:48,120 They likely found them in parts of the chemical space that 169 00:07:48,120 --> 00:07:50,280 humans were completely ignoring. Why? 170 00:07:51,000 --> 00:07:53,800 Because humans are biased. We have chemical intuition, 171 00:07:53,800 --> 00:07:55,560 which is great, but it's also a blinder. 172 00:07:56,160 --> 00:07:58,360 We tend to modify things we already know work. 173 00:07:58,360 --> 00:07:59,920 We think, oh, this electrolyte works pretty well. 174 00:07:59,920 --> 00:08:02,520 Let me just add a methyl group over here and see if it gets a 175 00:08:02,520 --> 00:08:04,400 little better. You stick to your comfort zone, 176 00:08:05,440 --> 00:08:07,240 the computer. The computer has no comfort 177 00:08:07,240 --> 00:08:08,920 zone. It doesn't care about tradition 178 00:08:08,920 --> 00:08:12,000 or intuition, it just methodically churns through all 179 00:08:12,000 --> 00:08:14,760 the possibilities you give it. It looks in the weird corners of 180 00:08:14,760 --> 00:08:17,560 the periodic table that no human would ever think to look. 181 00:08:17,720 --> 00:08:18,960 Exactly. And that's where the 182 00:08:18,960 --> 00:08:21,320 breakthroughs are hiding. Which leads us, I think, to the 183 00:08:21,520 --> 00:08:25,280 biggest, most mind bending number in the entire fly deck. 184 00:08:25,480 --> 00:08:28,920 The deep mind stat. Oh yeah, the 2.2 million 185 00:08:28,920 --> 00:08:31,800 structure. It's just it's staggering. 186 00:08:32,440 --> 00:08:36,480 Deep Mines AI, their material science model, predicted 2.2 187 00:08:36,480 --> 00:08:39,240 million new stable crystal structures. 188 00:08:39,360 --> 00:08:42,720 OK, put that in context for us. How many did we as a species 189 00:08:42,720 --> 00:08:45,400 know about before that? In the entire history of 190 00:08:45,400 --> 00:08:49,360 inorganic chemistry leading up to that point, humanity had 191 00:08:49,680 --> 00:08:54,920 painstakingly identified and catalog maybe, maybe 50,000 192 00:08:55,280 --> 00:08:58,800 stable inorganic crystals. So in one project, one AI 193 00:08:58,800 --> 00:09:02,240 basically did what, 1000 years of human scientific work in an 194 00:09:02,240 --> 00:09:03,840 afternoon? Essentially, yes. 195 00:09:04,240 --> 00:09:06,200 It's a breathtaking expansion of our knowledge. 196 00:09:06,200 --> 00:09:09,080 Now, we have to be careful here. Predicting a structure exists on 197 00:09:09,080 --> 00:09:11,520 a computer is not the same as holding in your hand. 198 00:09:11,520 --> 00:09:13,400 There's still a lot of work to do to synthesize them, of 199 00:09:13,440 --> 00:09:16,200 course, but it gives us a map and almost complete map of 200 00:09:16,200 --> 00:09:18,560 what's possible. It turns us from explorers 201 00:09:18,560 --> 00:09:21,000 hacking our way through the jungle with a machete into 202 00:09:21,000 --> 00:09:22,840 navigators with a full blown GPS. 203 00:09:23,000 --> 00:09:25,480 And this concept of inverse design that the lecture 204 00:09:25,480 --> 00:09:28,320 mentions, that's the GPS, right? That's the navigation system. 205 00:09:28,320 --> 00:09:30,560 That's exactly traditional science. 206 00:09:30,560 --> 00:09:33,200 The shake and bake method is direct design. 207 00:09:33,480 --> 00:09:36,360 You have a material and you measure its properties to see if 208 00:09:36,360 --> 00:09:37,320 it's any good. Yep. 209 00:09:37,760 --> 00:09:39,360 Does this rock conduct electricity? 210 00:09:39,400 --> 00:09:41,400 No. OK, how about this rock? 211 00:09:41,480 --> 00:09:43,840 No, you're just testing what you find. 212 00:09:44,040 --> 00:09:46,840 Inverse design completely flips the script. 213 00:09:46,920 --> 00:09:48,560 It does. You start with the answer. 214 00:09:48,880 --> 00:09:52,120 You go to the computer and you say I want a material that is 215 00:09:52,120 --> 00:09:55,520 transparent, conducts electricity like a metal, is 216 00:09:55,520 --> 00:09:58,760 flexible and cost less than $5 a kilogram. 217 00:09:58,840 --> 00:10:01,320 So you define the properties you need first. 218 00:10:01,520 --> 00:10:05,600 And the computer searches that massive database of 2.2 million 219 00:10:05,600 --> 00:10:08,120 structures. It works backward to tell you 220 00:10:08,120 --> 00:10:10,000 OK. To get those properties you need 221 00:10:10,000 --> 00:10:14,280 to make strontium, titanium and oxygen in this precise crystal 222 00:10:14,280 --> 00:10:16,440 structure. It's the ultimate shortcut to 223 00:10:16,440 --> 00:10:19,080 innovation. It's designing with intent 224 00:10:19,080 --> 00:10:21,040 instead of by accident. That's the goal. 225 00:10:21,160 --> 00:10:24,240 But let's let's get under the hood here, because I don't want 226 00:10:24,240 --> 00:10:26,840 to just hype the results. We need to understand the tools 227 00:10:26,840 --> 00:10:29,760 that make this possible. The lecture uses this wonderful 228 00:10:29,760 --> 00:10:33,320 phrase from Newton to Schrodinger, which I think is a 229 00:10:33,320 --> 00:10:35,800 great way to visualize the hierarchy of methods. 230 00:10:35,800 --> 00:10:39,120 It's the scale adder, and this is probably the most critical 231 00:10:39,120 --> 00:10:42,440 concept for anyone trying to get their head around computational 232 00:10:42,440 --> 00:10:45,440 chemistry. There is no single master 233 00:10:45,440 --> 00:10:48,760 algorithm that simulates everything from an electron to 234 00:10:48,760 --> 00:10:51,000 an airplane. It just doesn't exist. 235 00:10:51,080 --> 00:10:54,480 Right, you can't use the same math to model a single electron 236 00:10:54,480 --> 00:10:56,800 that you use to model, say, a bridge. 237 00:10:56,960 --> 00:10:59,120 Exactly. The physics that matters changes 238 00:10:59,120 --> 00:11:01,080 depending on the size of the thing you're looking at and the 239 00:11:01,080 --> 00:11:03,560 speed at which it's moving. The lecture breaks this down 240 00:11:03,560 --> 00:11:06,000 into 4 main tiers. OK, so let's start at the 241 00:11:06,000 --> 00:11:07,720 bottom. The absolute basement of 242 00:11:07,720 --> 00:11:10,280 reality. Tier 1 quantum mechanical 243 00:11:10,280 --> 00:11:12,360 methods. This is the electronic scale. 244 00:11:12,480 --> 00:11:15,720 We are dealing with length scales of angstroms, that's .1 245 00:11:15,720 --> 00:11:18,640 nanometers and time scales of femtoseconds. 246 00:11:19,120 --> 00:11:21,600 We're talking about the behavior of individual electrons. 247 00:11:21,600 --> 00:11:24,600 And the boss down here, the governing law is Schrodinger, 248 00:11:24,800 --> 00:11:26,440 the. Schrodinger equation, yeah, It's 249 00:11:26,440 --> 00:11:28,240 the fundamental equation of quantum mechanics. 250 00:11:28,520 --> 00:11:30,280 It describes the wave function of a system. 251 00:11:30,560 --> 00:11:34,480 And if you can solve it, you know literally everything there 252 00:11:34,480 --> 00:11:37,840 is to know about that system, the energy, the position of 253 00:11:37,840 --> 00:11:42,320 electrons, how it will react. Everything but, and this is a 254 00:11:42,320 --> 00:11:44,760 big but, there's a catch. There's a huge catch. 255 00:11:44,880 --> 00:11:46,280 It's called the Many Body Problem. 256 00:11:46,800 --> 00:11:49,480 The Schrodinger equation is actually pretty easy to solve if 257 00:11:49,480 --> 00:11:52,120 you only have one electron and one proton, like a hydrogen 258 00:11:52,120 --> 00:11:53,680 atom. But as soon as you have two 259 00:11:53,680 --> 00:11:56,240 electrons. Even just two, like in a helium 260 00:11:56,240 --> 00:11:59,160 atom, it becomes incredibly difficult because now you don't 261 00:11:59,160 --> 00:12:01,680 just have the electrons attracted to the nucleus, you 262 00:12:01,680 --> 00:12:03,720 have the two electrons propelling each other. 263 00:12:04,120 --> 00:12:07,320 And as you add more and more electrons, the number of these 264 00:12:07,320 --> 00:12:09,760 interactions just explodes in complexity. 265 00:12:09,800 --> 00:12:13,480 So for a real material with trillions of electrons, it's 266 00:12:13,840 --> 00:12:16,880 it's just not possible. It's mathematically practically 267 00:12:16,920 --> 00:12:20,280 unsolvable in its pure form. So we have to cheat. 268 00:12:20,600 --> 00:12:23,040 Or to be more polite, we approximate. 269 00:12:23,200 --> 00:12:26,240 And this is where DfT comes in. Density functional theory. 270 00:12:26,520 --> 00:12:28,320 It's all over these lecture slides for a reason. 271 00:12:28,560 --> 00:12:31,880 It's the workhorse. It is the absolute workhorse of 272 00:12:31,880 --> 00:12:34,360 modern computational chemistry and material science. 273 00:12:34,640 --> 00:12:37,880 It won the Nobel Prize in Chemistry in 1998, and it 274 00:12:37,880 --> 00:12:39,480 deserved. It So how does it cheat? 275 00:12:39,600 --> 00:12:42,360 What's the trick? The trick is genius. 276 00:12:42,760 --> 00:12:45,520 Instead of trying to track the position and momentum of every 277 00:12:45,520 --> 00:12:48,600 single electron, which is a nightmare of complexity, Anybody 278 00:12:48,600 --> 00:12:51,920 problem, right, DfT says let's just look at the overall 279 00:12:51,920 --> 00:12:54,880 electron density. OK, so instead of tracking every 280 00:12:54,880 --> 00:12:57,080 single car on the highway, you're just looking at the 281 00:12:57,120 --> 00:12:59,760 overall flow of traffic. The traffic jam itself is a kind 282 00:12:59,760 --> 00:13:02,680 of fluid. That is a fantastic analogy. 283 00:13:03,320 --> 00:13:06,760 By focusing on the density, a property that only depends on 284 00:13:06,760 --> 00:13:11,320 three spatial coordinates XY and Z, you reduce the complexity 285 00:13:11,320 --> 00:13:14,160 from a three end dimensional problem where N is the number of 286 00:13:14,160 --> 00:13:16,400 electrons down to just three dimensions. 287 00:13:16,400 --> 00:13:18,000 And that makes the impossible possible. 288 00:13:18,040 --> 00:13:19,880 It makes it possible on a supercomputer. 289 00:13:19,960 --> 00:13:23,600 The lecture lists the software heavyweights here, VISP, Quantum 290 00:13:23,600 --> 00:13:25,680 Espresso, Cast Step, and Gaussian. 291 00:13:25,680 --> 00:13:27,840 Are they all just different flavors of the same thing? 292 00:13:27,960 --> 00:13:30,000 Not quite. They're distinct tools for 293 00:13:30,000 --> 00:13:33,160 different jobs. VSP, Quantum Espresso, and 294 00:13:33,160 --> 00:13:35,560 Casket are typically used for solid-state physics. 295 00:13:35,800 --> 00:13:39,280 So metals, semiconductors, crystals. 296 00:13:39,280 --> 00:13:42,400 Exactly things that are periodic that repeat in space. 297 00:13:42,960 --> 00:13:46,080 They use something called plane waves to describe the electrons. 298 00:13:46,280 --> 00:13:49,280 Because in a metal the electrons are delocalized. 299 00:13:49,360 --> 00:13:52,040 They act like waves moving through the entire solid. 300 00:13:52,320 --> 00:13:55,720 Whereas Gaussian is different. Gaussian is the undisputed king 301 00:13:55,720 --> 00:13:57,520 of molecular and organic chemistry. 302 00:13:57,520 --> 00:14:00,400 So if you're designing a new drug molecule, for example. 303 00:14:00,400 --> 00:14:02,000 Right. It uses what are called 304 00:14:02,120 --> 00:14:05,760 localized basis sets. It treats electrons as being 305 00:14:05,760 --> 00:14:08,680 more or less attached to specific atoms, which is a much 306 00:14:08,680 --> 00:14:10,600 better description for individual molecules. 307 00:14:10,920 --> 00:14:12,760 So if you're designing a new pharmaceutical, you're probably 308 00:14:12,760 --> 00:14:14,600 using Gaussian. If you're designing a new 309 00:14:14,600 --> 00:14:17,040 battery cathode, you're probably using VSP. 310 00:14:17,280 --> 00:14:20,440 But even with the brilliant cheat that DfT uses, we still 311 00:14:20,440 --> 00:14:22,720 have limits. The lecture notes say these 312 00:14:22,720 --> 00:14:25,840 methods are usually limited to less than 1000 atoms. 313 00:14:25,880 --> 00:14:29,000 And even 1000 atoms is pushing it for a routine calculation. 314 00:14:29,560 --> 00:14:32,520 A standard DfT job might handle a few 100 atoms. 315 00:14:33,360 --> 00:14:36,400 The problem is that the computational cost still scales 316 00:14:36,400 --> 00:14:40,280 badly, usually with the cube or even the 4th power of the number 317 00:14:40,280 --> 00:14:42,800 of electrons. So if you double the size of 318 00:14:42,800 --> 00:14:45,280 your system. The calculation takes 8 or 16 319 00:14:45,280 --> 00:14:48,360 times longer to run. It gets expensive very, very 320 00:14:48,360 --> 00:14:50,560 fast. So you can model a very accurate 321 00:14:50,560 --> 00:14:54,160 picture of the, say, the active site of a catalyst. 322 00:14:54,160 --> 00:14:56,560 The most important part, Yes. But you can't model the whole 323 00:14:56,560 --> 00:14:59,480 nanoparticle it sits on, let alone the whole reactor. 324 00:14:59,760 --> 00:15:02,520 Not with this method. For that you have to climb up 325 00:15:02,520 --> 00:15:05,240 the ladder to Tier 2 atomistic simulation. 326 00:15:05,360 --> 00:15:07,680 And this is where we say goodbye to Schrodinger and quantum 327 00:15:07,680 --> 00:15:10,040 mechanics, and we say hello to good old Isaac Newton. 328 00:15:10,680 --> 00:15:12,320 We do. We're in the world of classical 329 00:15:12,320 --> 00:15:15,120 mechanics now. The main tool here is molecular 330 00:15:15,120 --> 00:15:17,680 dynamics or MD. OK, so what's the big 331 00:15:17,680 --> 00:15:19,120 simplification we're making here? 332 00:15:19,120 --> 00:15:22,680 The big simplification is that we completely stop caring about 333 00:15:22,680 --> 00:15:24,720 individual electrons. We just ignore them. 334 00:15:24,720 --> 00:15:29,560 We treat the atom as a single hard sphere, a little ball like 335 00:15:29,560 --> 00:15:31,640 a billiard ball. We're essentially dumbing the 336 00:15:31,640 --> 00:15:33,600 system down. So how did the balls know how to 337 00:15:33,600 --> 00:15:36,160 interact with each other? We tell them we use what's 338 00:15:36,160 --> 00:15:39,400 called a force field. It's basically a set a simple 339 00:15:39,600 --> 00:15:41,200 mathematical rules. Like what? 340 00:15:41,360 --> 00:15:45,280 Like if ball A and ball B get too close, push them apart with 341 00:15:45,280 --> 00:15:47,280 this much force. If they're connected by a 342 00:15:47,280 --> 00:15:50,080 chemical bond, treat that bond like a spring and pull them 343 00:15:50,080 --> 00:15:53,080 together. It's all just simple springs and 344 00:15:53,080 --> 00:15:55,280 repulsions. And because we're not solving 345 00:15:55,280 --> 00:15:58,600 these incredibly complex quantum equations anymore, this must be 346 00:15:58,600 --> 00:16:01,840 way, way faster. Orders of magnitude faster. 347 00:16:01,960 --> 00:16:03,960 It's not even a comparison with MD. 348 00:16:03,960 --> 00:16:07,280 You can simulate systems with millions or even billions of 349 00:16:07,280 --> 00:16:09,640 atoms. You can watch proteins fold in 350 00:16:09,640 --> 00:16:11,720 water. You can watch polymers entangled 351 00:16:11,720 --> 00:16:13,880 to form plastics. The lecture lists the big 352 00:16:13,880 --> 00:16:17,480 software names here to LAMP, PS, Gromax, Amber. 353 00:16:17,560 --> 00:16:19,160 Right, and again, they have their specialties. 354 00:16:19,480 --> 00:16:23,480 Gromax and Amber are very famous in the biochemistry world for 355 00:16:23,480 --> 00:16:26,840 simulating proteins in DNA. LAMP PS is more of a general 356 00:16:26,840 --> 00:16:29,280 purpose tool, very powerful for material science. 357 00:16:29,440 --> 00:16:30,960 There has to be a huge trade off. 358 00:16:31,160 --> 00:16:34,000 If we're ignoring electrons, how can we possibly do chemistry? 359 00:16:34,000 --> 00:16:36,920 Chemical reactions are electrons moving around, bonds breaking 360 00:16:36,920 --> 00:16:38,480 and forming. If I just have little balls 361 00:16:38,480 --> 00:16:40,440 connected by springs, I can't break the spring. 362 00:16:40,440 --> 00:16:43,280 And that is the classic fundamental limitation of 363 00:16:43,280 --> 00:16:46,520 standard classical MD. It is non reactive so. 364 00:16:46,520 --> 00:16:47,960 It's great for physical processes. 365 00:16:47,960 --> 00:16:51,040 Fantastic melting, boiling, diffusion. 366 00:16:51,400 --> 00:16:53,720 How a material responds to being stretched. 367 00:16:54,400 --> 00:16:57,680 All physical things, but it's terrible for modeling chemical 368 00:16:57,680 --> 00:16:59,840 reactions. So it can't model a fire. 369 00:16:59,880 --> 00:17:03,680 No, but the lecture mentions a bridge, a clever way to get 370 00:17:03,680 --> 00:17:06,400 around this called Reax FF. Reactive force fields. 371 00:17:06,400 --> 00:17:08,560 This is a really brilliant hybrid method. 372 00:17:08,800 --> 00:17:10,760 It's still a classical method at its core. 373 00:17:10,760 --> 00:17:14,000 It's not solving Schrodinger's equation, but the force field 374 00:17:14,000 --> 00:17:16,680 itself. The rules for the springs is 375 00:17:16,680 --> 00:17:19,920 much more sophisticated. It's been trained on a vast 376 00:17:19,920 --> 00:17:22,560 amount of quantum data from DfT calculations. 377 00:17:22,560 --> 00:17:25,160 So it learns how bonds should behave. 378 00:17:25,160 --> 00:17:27,200 Exactly. It has a concept of bond order. 379 00:17:27,200 --> 00:17:30,120 It knows that if you pull 2 atoms apart, the bond between 380 00:17:30,120 --> 00:17:32,640 them gets weaker and weaker until at a certain point it 381 00:17:32,640 --> 00:17:35,360 snaps and it knows how to form new bonds. 382 00:17:35,600 --> 00:17:38,080 O with Rex FF you can simulate a fire. 383 00:17:38,160 --> 00:17:40,320 You can. You can simulate combustion, or 384 00:17:40,320 --> 00:17:43,840 an explosion, or a battery material degrading overtime, all 385 00:17:43,840 --> 00:17:47,000 these complex chemical events involving millions of atoms 386 00:17:47,440 --> 00:17:50,080 which you could never ever hope to do with pure DfT. 387 00:17:50,080 --> 00:17:52,800 It sits right in that sweet spot between the two tiers. 388 00:17:53,000 --> 00:17:55,840 It's an incredibly powerful tool for bridging that gap. 389 00:17:56,040 --> 00:17:58,400 OK, let's keep climbing. We're moving up the ladder to 390 00:17:58,400 --> 00:18:01,320 Tier 3, the mesoscale. This is the middle ground, 391 00:18:01,320 --> 00:18:03,320 right? We're zooming out again, and 392 00:18:03,320 --> 00:18:06,320 here we start to blur the atoms together. 393 00:18:06,760 --> 00:18:09,000 The main technique is called coarse graining. 394 00:18:09,160 --> 00:18:11,360 It sounds like pixelating an image. 395 00:18:11,640 --> 00:18:13,120 That's a great way to think about it. 396 00:18:13,480 --> 00:18:17,440 Imagine you're trying to model a cell membrane which is made of 397 00:18:17,640 --> 00:18:21,840 millions of lipid molecules. You don't need to know the exact 398 00:18:21,840 --> 00:18:25,280 position of every single hydrogen atom on every lipid 399 00:18:25,280 --> 00:18:27,200 tail. It's just too much detail, you'd 400 00:18:27,200 --> 00:18:29,720 never be able to simulate it for long enough to see anything 401 00:18:29,720 --> 00:18:30,920 interesting. Exactly. 402 00:18:31,120 --> 00:18:34,400 So with coarse graining, you might say, OK, let's treat the 403 00:18:34,400 --> 00:18:38,000 head group of a lipid as one big bead, and we'll treat the tail 404 00:18:38,000 --> 00:18:40,240 as two or three other beads connected by a string. 405 00:18:40,640 --> 00:18:42,600 You're simplifying the representation. 406 00:18:42,640 --> 00:18:44,600 Drastically. And by doing that you can 407 00:18:44,600 --> 00:18:48,280 suddenly simulate much larger systems for much longer times. 408 00:18:48,920 --> 00:18:51,800 Instead of nanoseconds, you can watch things happen over 409 00:18:51,800 --> 00:18:53,960 microseconds or even milliseconds. 410 00:18:54,480 --> 00:18:57,080 This is how people study things like protein folding or how 411 00:18:57,080 --> 00:18:59,760 polymers self assemble. And the lecture also mentions 412 00:18:59,760 --> 00:19:02,320 the phase field method here. That sounds a bit different. 413 00:19:02,320 --> 00:19:04,280 It is. This is huge for metallurgy and 414 00:19:04,280 --> 00:19:07,920 material science. Imagine a pool of molten metal 415 00:19:08,000 --> 00:19:10,080 that's cooling down and starting to solidify. 416 00:19:10,080 --> 00:19:13,160 Crystals begin to form. Right, little crystals, what we 417 00:19:13,160 --> 00:19:16,800 call grains, start to grow. They bump into each other and 418 00:19:16,800 --> 00:19:19,360 form boundaries. The Hase field method is a 419 00:19:19,360 --> 00:19:22,760 mathematical way to track the evolution of those boundaries an 420 00:19:22,760 --> 00:19:24,520 the shapes of those grains overtime O. 421 00:19:24,680 --> 00:19:26,920 It's not tracking atoms at all anymore. 422 00:19:27,080 --> 00:19:29,480 Not at all. It's tracking a continuous field 423 00:19:29,480 --> 00:19:33,360 that just tells you this region is solid crystal A, this region 424 00:19:33,360 --> 00:19:36,480 is solid crystal B, and this region is still liquid. 425 00:19:36,480 --> 00:19:39,000 It's another level of abstraction up. 426 00:19:39,280 --> 00:19:42,520 And that leads us to the top of the ladder, Tier 4, the macro 427 00:19:42,520 --> 00:19:45,480 scale or continuum methods. This is the world we actually 428 00:19:45,480 --> 00:19:47,160 see and live in. This is engineering. 429 00:19:47,160 --> 00:19:50,440 This is where the material becomes just a solid uniform 430 00:19:50,440 --> 00:19:53,160 block with certain properties like strength and conductivity. 431 00:19:53,640 --> 00:19:56,920 We use tools like the Finite Element Method or FEM, or 432 00:19:56,920 --> 00:19:59,080 Computational Fluid Dynamics CFD. 433 00:19:59,080 --> 00:20:02,120 This is what an aerospace company uses to design A wing. 434 00:20:02,360 --> 00:20:04,760 Exactly. They are not simulating the 435 00:20:04,760 --> 00:20:07,400 individual iron and aluminum atoms in the wing. 436 00:20:07,600 --> 00:20:09,400 They've smeared all that detail out. 437 00:20:09,720 --> 00:20:13,160 They're solving equations for stress, strain, heat, flow and 438 00:20:13,160 --> 00:20:15,880 things like that. And for CFD it's airflow. 439 00:20:16,160 --> 00:20:19,080 Right, they're solving the Navier Stokes equations for how 440 00:20:19,080 --> 00:20:21,920 air flows over the wing. The software here are giants 441 00:20:21,920 --> 00:20:25,160 like COMSOL and NSYS. So the ultimate goal, and I 442 00:20:25,160 --> 00:20:28,200 think the lecture is really emphatic about this point, is to 443 00:20:28,200 --> 00:20:30,240 link all of these tiers together. 444 00:20:30,280 --> 00:20:33,000 Yes, that's the dream. To take the quantum data about 445 00:20:33,000 --> 00:20:37,040 how 2 atoms bond, feed that information into the force field 446 00:20:37,040 --> 00:20:40,160 for your molecular model. Use the results of the molecular 447 00:20:40,160 --> 00:20:43,680 model to parameterize your mesoscale model, and finally use 448 00:20:43,680 --> 00:20:46,360 that to tell the engineer what the bulk properties of the 449 00:20:46,360 --> 00:20:48,560 bridge will be. That whole process is called 450 00:20:48,560 --> 00:20:51,360 multi scale modeling. It is the Holy Grail of this 451 00:20:51,360 --> 00:20:54,360 entire field. It's incredibly difficult to do 452 00:20:54,360 --> 00:20:57,760 seamlessly to pass information between the scales without 453 00:20:57,760 --> 00:21:00,480 losing accuracy, but it's where everything is headed. 454 00:21:00,480 --> 00:21:03,240 OK, so we have the toolkit. We understand the latter of 455 00:21:03,240 --> 00:21:04,960 complexity. Now let's apply it. 456 00:21:05,240 --> 00:21:07,600 The lecture focuses very specifically on the field of 457 00:21:07,600 --> 00:21:09,840 catalysis. The absolute heart of the 458 00:21:09,840 --> 00:21:13,440 chemical industry, something like 90% of all chemical 459 00:21:13,440 --> 00:21:16,640 products have seen a catalyst at some point in their production. 460 00:21:16,640 --> 00:21:20,400 So if you can improve catalysts, you improve, well, everything 461 00:21:20,440 --> 00:21:22,280 pretty much. And the core question the 462 00:21:22,280 --> 00:21:26,480 lecture poses is what can computational methods do for 463 00:21:26,480 --> 00:21:29,560 catalysis? And from what I can see, it 464 00:21:29,560 --> 00:21:32,680 seems to boil down to understanding and improving 3 465 00:21:32,680 --> 00:21:36,680 key pillars, Activity, Kinetics and Selectivity. 466 00:21:36,920 --> 00:21:39,240 Right, let's break those down because this is where the theory 467 00:21:39,240 --> 00:21:41,080 meets the real world. This is where the money is made. 468 00:21:41,080 --> 00:21:42,040 Let's. Start with the first one 469 00:21:42,160 --> 00:21:44,880 Activity. The slides say this is all about 470 00:21:44,880 --> 00:21:48,320 the active site structure. The active site is the specific 471 00:21:48,320 --> 00:21:51,920 special arrangement of atoms on the catalyst surface where the 472 00:21:51,920 --> 00:21:53,760 chemical reaction actually happens. 473 00:21:54,280 --> 00:21:56,720 You can think of it as a perfectly shaped parking spot 474 00:21:56,720 --> 00:21:59,560 for a molecule. If the molecule parks there, the 475 00:21:59,560 --> 00:22:01,720 reaction happens. If it parks anywhere else, 476 00:22:01,720 --> 00:22:03,960 nothing happens. So a more active catalyst has 477 00:22:03,960 --> 00:22:06,360 more of these special parking spots or better ones. 478 00:22:06,360 --> 00:22:08,520 Exactly, and computation lets us design them. 479 00:22:08,880 --> 00:22:11,520 The lecture mentions heteroatom doping is 1 strategy. 480 00:22:11,520 --> 00:22:14,400 OK, what is that? So imagine you have a surface of 481 00:22:14,400 --> 00:22:16,240 pure carbon, like a sheet of graphene. 482 00:22:16,880 --> 00:22:19,840 It's pretty inert. The carbon atoms are all happy. 483 00:22:19,840 --> 00:22:21,440 They don't really want to react with much. 484 00:22:22,080 --> 00:22:25,480 But what if you use a computer to simulate kicking out one of 485 00:22:25,480 --> 00:22:28,760 those carbon atoms and replacing it with a nitrogen atom? 486 00:22:29,520 --> 00:22:32,200 That's heteroatom doping. You're introducing an impurity 487 00:22:32,200 --> 00:22:34,000 on. Purpose a strategic impurity, 488 00:22:34,520 --> 00:22:36,960 and you've completely changed the electronic landscape right 489 00:22:36,960 --> 00:22:39,040 at that spot. Nitrogen has a different number 490 00:22:39,040 --> 00:22:41,360 of electrons than carbon. It creates a little hot spot on 491 00:22:41,360 --> 00:22:44,240 the surface. Suddenly, that one spot becomes 492 00:22:44,240 --> 00:22:48,000 incredibly attractive to, say, an oxygen molecule in the air. 493 00:22:48,000 --> 00:22:49,760 It activates it it. Activates it. 494 00:22:50,120 --> 00:22:53,160 The reaction can now happen at that site, and with computation 495 00:22:53,280 --> 00:22:55,800 we can screen the entire periodic table to predict 496 00:22:55,800 --> 00:22:58,960 exactly which element to dope with and precisely where to put 497 00:22:58,960 --> 00:23:01,280 it to get the maximum possible activity. 498 00:23:01,640 --> 00:23:04,000 The lecture also mentions defect engineering. 499 00:23:04,360 --> 00:23:07,600 I love this concept because in most parts of life, defects are 500 00:23:07,600 --> 00:23:10,240 bad. You want a perfect diamond, a 501 00:23:10,240 --> 00:23:13,800 flawless crystal. But in catalysis, perfection is 502 00:23:13,800 --> 00:23:16,000 actually failure. It's so true. 503 00:23:16,360 --> 00:23:20,240 A perfectly flat, perfectly ordered crystal surface is often 504 00:23:20,240 --> 00:23:23,240 a terrible catalyst. The atoms are all happy and 505 00:23:23,240 --> 00:23:25,160 comfortable and have the right number of neighbors. 506 00:23:25,400 --> 00:23:27,360 They are low energy. They don't want to bond with 507 00:23:27,360 --> 00:23:28,920 anything else. So you need to break the 508 00:23:28,920 --> 00:23:30,440 surface. You need to introduce 509 00:23:30,440 --> 00:23:32,640 imperfection. You need to rough it up, but in 510 00:23:32,640 --> 00:23:36,560 a very controlled way. You need to create steps, edges, 511 00:23:36,560 --> 00:23:39,480 corners and vacancies where you've plucked an atom out. 512 00:23:39,600 --> 00:23:41,320 And why are those spots so active? 513 00:23:41,840 --> 00:23:45,640 It's because the atoms at those defect sites have missing 514 00:23:45,640 --> 00:23:47,600 neighbors. We call them low coordination 515 00:23:47,600 --> 00:23:50,720 sites, and because they're missing neighbors, their 516 00:23:50,720 --> 00:23:55,120 remaining electrons are sort of dangling, unsatisfied. 517 00:23:55,320 --> 00:23:58,320 They are desperate to bond with something to become stable. 518 00:23:58,440 --> 00:24:00,680 So they'll grab on to passing molecules. 519 00:24:00,680 --> 00:24:02,440 You'll grab on tight and help them react. 520 00:24:02,480 --> 00:24:05,280 Computers can tell us if you create a step edge with this 521 00:24:05,280 --> 00:24:08,440 specific geometry and your nanoparticle, your reaction rate 522 00:24:08,440 --> 00:24:11,800 will increase by 1000%. We can design the defects. 523 00:24:11,880 --> 00:24:14,600 OK, that's activity. The second pillar is kinetics. 524 00:24:14,600 --> 00:24:16,920 This is about movement, about speed, right? 525 00:24:17,320 --> 00:24:19,560 The lecture frames this as support architecture. 526 00:24:20,520 --> 00:24:22,720 It doesn't matter how amazing your active site is if the 527 00:24:22,720 --> 00:24:25,640 reactive molecules can't get to it and the product molecules 528 00:24:25,640 --> 00:24:28,320 can't get away. This is a mass transport 529 00:24:28,320 --> 00:24:30,800 problem, a diffusion problem. And the lecture talks about 530 00:24:30,800 --> 00:24:34,520 designing hierarchical pores. This is a beautiful concept 531 00:24:34,600 --> 00:24:36,840 borrowed from nature. Think about a city's road 532 00:24:36,840 --> 00:24:39,160 network. You need huge highways. 533 00:24:39,160 --> 00:24:42,920 Those are the large pores to get a massive amount of traffic into 534 00:24:42,920 --> 00:24:45,600 the city center quickly. But you can't park on a highway. 535 00:24:45,600 --> 00:24:47,520 Exactly. So then you need the main 536 00:24:47,520 --> 00:24:50,840 streets, the medium sized pores to distribute that traffic 537 00:24:50,840 --> 00:24:54,000 throughout the neighborhoods. And finally, you need the little 538 00:24:54,000 --> 00:24:57,360 driveways, the small pores to get to the individual houses 539 00:24:57,440 --> 00:25:00,440 which are the active sites. If you only had tiny driveways 540 00:25:00,440 --> 00:25:03,000 from the countryside, you'd have a permanent traffic jam. 541 00:25:03,360 --> 00:25:05,400 If you only had highways, you could never stop. 542 00:25:05,600 --> 00:25:07,960 Precisely. You need all three scales 543 00:25:07,960 --> 00:25:10,320 working together. Nature does this brilliantly. 544 00:25:10,320 --> 00:25:13,400 Look at the structure of your lungs or the veins in a leaf. 545 00:25:14,200 --> 00:25:17,200 We use computations, specifically methods like CFD, 546 00:25:17,560 --> 00:25:20,680 to design these hierarchical pore structures in catalyst 547 00:25:20,680 --> 00:25:23,920 supports to maximize the flow of molecules in and out. 548 00:25:24,200 --> 00:25:27,360 So you're engineering the Catalyst's plumbing system? 549 00:25:27,360 --> 00:25:28,880 You're being a civil engineer, yes. 550 00:25:29,040 --> 00:25:32,160 And that brings us to the third pillar, selectivity. 551 00:25:32,720 --> 00:25:35,440 This is the really subtle one, the one that prevents waste. 552 00:25:36,240 --> 00:25:38,560 This is arguably the most important one for green 553 00:25:38,560 --> 00:25:40,960 chemistry. In many complex chemical 554 00:25:40,960 --> 00:25:44,840 reactions, you have a choice. Your reactive molecule A could 555 00:25:44,840 --> 00:25:47,400 turn into product B, which is the valuable pharmaceutical you 556 00:25:47,400 --> 00:25:50,920 want, or could turn into product C, which is a useless toxic 557 00:25:50,920 --> 00:25:52,680 sludge. And traditionally you just get a 558 00:25:52,680 --> 00:25:55,280 mix of both and then have to spend a ton of energy and money 559 00:25:55,280 --> 00:25:57,920 separating the good stuff from the bad stuff later on. 560 00:25:58,000 --> 00:26:01,320 Right Computational catalysis allows us to tune the reaction 561 00:26:01,320 --> 00:26:04,280 to be more selective, to force it down the path we want. 562 00:26:04,720 --> 00:26:07,280 The lecture mentions field effects as a powerful way to do 563 00:26:07,280 --> 00:26:09,800 this. Strain electric fields, Magnetic 564 00:26:09,800 --> 00:26:11,840 fields. Take strain for example. 565 00:26:12,200 --> 00:26:14,320 If you take your catalyst surface and you simulate 566 00:26:14,320 --> 00:26:17,760 stretching it, literally pulling the atoms apart by just a few 567 00:26:17,760 --> 00:26:21,520 percent, you change the exact distance between the binding 568 00:26:21,520 --> 00:26:24,840 sites on the surface. So maybe the molecule that leads 569 00:26:24,840 --> 00:26:28,400 to the good product, product B fits perfectly on that stretch 570 00:26:28,400 --> 00:26:31,440 surface now, but the one that leads to the bad product, 571 00:26:31,480 --> 00:26:34,960 product C doesn't fit anymore. That's exactly the idea. 572 00:26:35,160 --> 00:26:38,240 You are physically tuning the lock so that only the right key 573 00:26:38,240 --> 00:26:40,160 will fit. It's an incredible level of 574 00:26:40,160 --> 00:26:42,680 control. Or you could apply an external 575 00:26:42,680 --> 00:26:45,640 electric field, and the simulation can show you how that 576 00:26:45,640 --> 00:26:49,760 field stabilizes the transition state for the reaction you want, 577 00:26:49,920 --> 00:26:52,320 making it happen much faster than the one you don't want. 578 00:26:52,560 --> 00:26:55,600 And the lecture also mentions using light photocatalysis. 579 00:26:55,680 --> 00:26:58,280 This is a huge area, especially for green energy. 580 00:26:58,560 --> 00:27:02,040 Think about splitting water into clean hydrogen fuel using only 581 00:27:02,040 --> 00:27:04,640 sunlight. The simulation can track what 582 00:27:04,640 --> 00:27:07,200 happens when a photon of light hits an electron in the 583 00:27:07,200 --> 00:27:09,600 catalyst. It enters an excited state. 584 00:27:09,600 --> 00:27:11,800 And you need that excited electron to go to the right 585 00:27:11,800 --> 00:27:12,960 place. You do. 586 00:27:13,120 --> 00:27:16,280 Does it jump to the right place on the surface to drive the 587 00:27:16,280 --> 00:27:19,400 water splitting reaction or does it just fall back down and waste 588 00:27:19,400 --> 00:27:22,440 that energy as heat? We can use quantum mechanics, 589 00:27:22,760 --> 00:27:26,600 specifically time dependent DfT, to design materials that Channel 590 00:27:26,600 --> 00:27:29,440 that light energy exactly where we need it to go. 591 00:27:29,760 --> 00:27:32,760 O We're designing the activity, the plumbing, and the final 592 00:27:32,760 --> 00:27:34,320 outcome. It's total control. 593 00:27:34,480 --> 00:27:38,200 In theory, yes, that's the goal. OK, section five of our 594 00:27:38,200 --> 00:27:41,160 discussion here brings us to what the lecture calls the 595 00:27:41,160 --> 00:27:44,080 Trinity diagram. It's a flow chart that connects 596 00:27:44,080 --> 00:27:47,280 insights, computations and guidelines specifically for the 597 00:27:47,280 --> 00:27:49,720 task of Electro catalyst discovery. 598 00:27:50,000 --> 00:27:51,640 This is basically the modern workflow. 599 00:27:51,920 --> 00:27:54,680 It shows how a computational chemistry lab actually operates 600 00:27:54,680 --> 00:27:56,520 day-to-day. It starts with insights on the 601 00:27:56,520 --> 00:28:00,080 left, things like understanding the activity origin, reaction 602 00:28:00,080 --> 00:28:04,520 mechanisms and active sites. So this is the human part, the 603 00:28:04,520 --> 00:28:06,920 scientists intuition and existing knowledge. 604 00:28:07,160 --> 00:28:09,320 What do we think is making this catalyst work? 605 00:28:09,560 --> 00:28:12,360 Then that feeds into the middle box computations. 606 00:28:12,800 --> 00:28:15,080 This is the engine this. Is the supercomputer doing the 607 00:28:15,080 --> 00:28:16,600 heavy lifting? We've been talking about renting 608 00:28:16,600 --> 00:28:20,280 the DfT, the MD, all of it. And that produces the output on 609 00:28:20,280 --> 00:28:23,640 the right guidelines for the experimentalist and there are 610 00:28:23,640 --> 00:28:27,000 three types listed, rational design, high throughput 611 00:28:27,000 --> 00:28:30,800 screening and machine learning. Right, so rational design is the 612 00:28:30,800 --> 00:28:33,680 classic approach. The computer tells you the 613 00:28:33,680 --> 00:28:36,200 binding strength of oxygen is too strong. 614 00:28:36,520 --> 00:28:39,680 Try adding a little bit of gold to your platinum catalyst to 615 00:28:39,680 --> 00:28:42,280 weaken it. It's a specific logical 616 00:28:42,280 --> 00:28:44,320 suggestion. High throughput screening is 617 00:28:44,320 --> 00:28:47,320 what we talked about before, testing thousands of options 618 00:28:47,320 --> 00:28:49,480 virtually to find the top candidates. 619 00:28:49,480 --> 00:28:53,000 And machine learning is the new frontier, using AI and neural 620 00:28:53,000 --> 00:28:56,320 networks to find patterns in the data that a human might miss. 621 00:28:56,680 --> 00:28:59,240 I want to focus on the battery application here because the 622 00:28:59,240 --> 00:29:02,360 lecture notes mention computational material science 623 00:29:02,360 --> 00:29:05,280 in the context of anodes, cathodes, and electrolytes. 624 00:29:06,000 --> 00:29:09,280 Everyone listening has a lithium ion battery in their pocket or 625 00:29:09,280 --> 00:29:12,240 on their desk right now. What is the computer actually 626 00:29:12,240 --> 00:29:14,920 calculating inside that battery? One of the most important 627 00:29:14,920 --> 00:29:17,320 properties it's calculating is the migration barrier. 628 00:29:17,360 --> 00:29:21,600 OK, unpack that term for us. A battery works by moving ions, 629 00:29:21,680 --> 00:29:24,720 in this case lithium ions, from one electrode to the other. 630 00:29:25,160 --> 00:29:28,560 To do that, a lithium ion has to hop from one empty spot in the 631 00:29:28,560 --> 00:29:31,720 crystal lattice of the electrode material to the next available 632 00:29:31,720 --> 00:29:33,840 spot. And that hop isn't free. 633 00:29:33,960 --> 00:29:35,680 It's not free. It cost a certain amount of 634 00:29:35,680 --> 00:29:37,360 energy to squeeze through the other atoms. 635 00:29:37,600 --> 00:29:39,960 That energy cost is a migration barrier. 636 00:29:40,080 --> 00:29:43,800 And if that barrier, that energy cost is too high? 637 00:29:43,920 --> 00:29:46,920 The lithium ion moves very slowly, it struggles to make the 638 00:29:46,920 --> 00:29:50,160 hop and that means your battery takes forever to charge and 639 00:29:50,160 --> 00:29:51,720 can't deliver power very quickly. 640 00:29:51,760 --> 00:29:54,640 But if the barrier is very low. You get fast charging, 641 00:29:55,040 --> 00:29:57,520 supercharging. The lithium ions can just glide 642 00:29:57,520 --> 00:29:58,240 through. The material. 643 00:29:58,480 --> 00:30:01,480 Computation allows us to map out the entire energy landscape 644 00:30:01,640 --> 00:30:04,880 inside a crystal and find the pathways of least resistance. 645 00:30:04,880 --> 00:30:07,880 Yeah, we can design materials where the lithium just zips 646 00:30:07,880 --> 00:30:09,400 through like it's on a slip and slide. 647 00:30:09,400 --> 00:30:11,520 Amazing. The other term listed is 648 00:30:11,520 --> 00:30:14,600 absorption energy. That's basically how sticky the 649 00:30:14,600 --> 00:30:17,000 surface of the material is for the lithium ions. 650 00:30:17,320 --> 00:30:18,920 You want the lithium to stick to the material. 651 00:30:18,920 --> 00:30:22,000 That's how it stores energy, but not so much that it gets stuck 652 00:30:22,000 --> 00:30:24,640 forever and can't leave. It's the Goldilocks principle 653 00:30:24,640 --> 00:30:26,680 again. Not too hot, not too cold. 654 00:30:26,760 --> 00:30:29,120 Exactly. And the final one here is phase 655 00:30:29,120 --> 00:30:31,760 transformation. This is the swelling problem, 656 00:30:31,760 --> 00:30:32,640 right? The reason? 657 00:30:32,640 --> 00:30:34,840 Old phone batteries sometimes puff up. 658 00:30:34,880 --> 00:30:36,680 That's it. When you shove a bunch of 659 00:30:36,680 --> 00:30:40,400 lithium ions into a material like graphite or silicon, the 660 00:30:40,400 --> 00:30:42,400 material has to expand to make room for them. 661 00:30:43,000 --> 00:30:45,520 It physically changes its structure, its phase. 662 00:30:45,600 --> 00:30:48,560 And that repeated swelling and shrinking with every charge 663 00:30:48,560 --> 00:30:52,080 cycle causes stress. It causes immense stress which 664 00:30:52,080 --> 00:30:55,320 leads to tiny cracks which grow overtime and actually kill the 665 00:30:55,320 --> 00:30:57,640 battery. We can use simulations to 666 00:30:57,640 --> 00:31:00,520 predict how much a material will expand and if it's likely to 667 00:31:00,520 --> 00:31:03,520 crack, all before we ever have to build a single physical 668 00:31:03,520 --> 00:31:05,680 prototype. It's just incredible how much 669 00:31:05,680 --> 00:31:09,360 violent physics is happening inside that little quiet black 670 00:31:09,360 --> 00:31:12,720 brick in our phones. It is a dynamic and frankly a 671 00:31:12,720 --> 00:31:16,080 pretty violent environment at the atomic scale, and we are 672 00:31:16,080 --> 00:31:18,320 finally learning how to tame it with mathematics. 673 00:31:18,480 --> 00:31:20,600 So let's look forward. We're heading into the future 674 00:31:20,600 --> 00:31:23,040 now with Section 6, the future landscape. 675 00:31:23,400 --> 00:31:28,000 The lecture actually references papers from 2024, 2025, which in 676 00:31:28,000 --> 00:31:31,880 the context of our 2026 source is the very recent past. 677 00:31:32,160 --> 00:31:34,960 What is the cutting edge? The cutting edge is the 678 00:31:34,960 --> 00:31:39,120 convergence of all these tools. The 2024 Advanced Materials 679 00:31:39,120 --> 00:31:42,560 paper mentioned in the slides talk specifically about merging 680 00:31:42,560 --> 00:31:45,880 first principles. That's our high accuracy DfT 681 00:31:45,920 --> 00:31:48,280 with machine learning. And this brings us back to that 682 00:31:48,280 --> 00:31:51,000 really important shift you mentioned earlier, the move from 683 00:31:51,000 --> 00:31:53,760 descriptive science to predictive science. 684 00:31:53,760 --> 00:31:56,040 That is the single most important take away from this 685 00:31:56,040 --> 00:31:58,800 whole discussion. For decades, computation and 686 00:31:58,800 --> 00:32:00,320 chemistry was largely descriptive. 687 00:32:00,600 --> 00:32:03,360 An experimentalist would run a reaction, something weird and 688 00:32:03,360 --> 00:32:05,440 unexpected would happen, and they would go to the 689 00:32:05,440 --> 00:32:07,640 computational chemist and ask them to explain it. 690 00:32:07,680 --> 00:32:10,680 It was a post mortem. Why did my catalyst turn blue 691 00:32:10,680 --> 00:32:12,320 and stop working? Exactly. 692 00:32:12,360 --> 00:32:16,120 It was forensic science. Now we are firmly moving into 693 00:32:16,120 --> 00:32:19,120 the predictive era. The computer is no longer the 694 00:32:19,120 --> 00:32:22,440 last resort for explanation, it's the first step in design. 695 00:32:23,200 --> 00:32:26,120 The computational chemist tells the experimentalist what to 696 00:32:26,120 --> 00:32:30,000 build before they go into a lab. If you put a single cobalt atom 697 00:32:30,000 --> 00:32:33,400 right here surrounded by 4 nitrogen atoms, it will work. 698 00:32:33,720 --> 00:32:36,400 And the surrogate models that are mentioned in the lecture. 699 00:32:36,840 --> 00:32:39,560 This is a fascinating idea. This is the solution to the 700 00:32:39,560 --> 00:32:42,840 scaling problem we talked about. You see, even though DfT is a 701 00:32:42,840 --> 00:32:46,400 million times cheaper than an experiment, it's still too slow 702 00:32:46,400 --> 00:32:50,240 to test, say, all 2.2 million of those deep mind structures. 703 00:32:50,240 --> 00:32:53,440 It would still take forever. So a surrogate model is an AI. 704 00:32:53,680 --> 00:32:55,920 You take a few thousand of those structures, a representative 705 00:32:55,920 --> 00:32:59,560 sample, and you run the full expensive DfT calculation on 706 00:32:59,560 --> 00:33:01,800 them. Then you feed all that data, the 707 00:33:01,800 --> 00:33:04,440 inputs and the outputs to a machine learning model. 708 00:33:04,520 --> 00:33:08,120 So the AI learns the underlying laws of physics from the DfT 709 00:33:08,120 --> 00:33:10,240 data. It learns an approximation of 710 00:33:10,240 --> 00:33:12,160 the physics. It doesn't understand 711 00:33:12,160 --> 00:33:14,960 Schrodinger's equation, but it learns the pattern. 712 00:33:15,240 --> 00:33:18,320 It learns to recognize oh when the atoms are arranged in this 713 00:33:18,320 --> 00:33:21,120 particular way, the binding energy is usually X. 714 00:33:21,560 --> 00:33:25,000 And once it learns that pattern. It can predict the properties of 715 00:33:25,080 --> 00:33:27,640 a brand new structure it's never seen before in a matter of 716 00:33:27,640 --> 00:33:30,520 milliseconds, instead of the hours or days it would take to 717 00:33:30,520 --> 00:33:34,280 run a full DfT calculation. It's a virtual virtual 718 00:33:34,280 --> 00:33:36,760 experiment, a simulation of the simulation. 719 00:33:36,760 --> 00:33:40,440 You have a perfect description. You can use this cheap, fast 720 00:33:40,440 --> 00:33:44,120 surrogate model to scan billions of potential compounds in a 721 00:33:44,120 --> 00:33:46,840 single day. Now it's less accurate for sure, 722 00:33:47,240 --> 00:33:50,440 but it acts as a massive filter. To find the needles in the 723 00:33:50,440 --> 00:33:51,480 haystack. Exactly. 724 00:33:51,640 --> 00:33:54,760 You use the AI surrogate to scan the billion options and find the 725 00:33:54,760 --> 00:33:57,400 top say 10,000 most promising ones. 726 00:33:57,800 --> 00:34:00,840 Then you use the more accurate DfT to check those top 10,000 727 00:34:00,840 --> 00:34:03,720 and narrow it down to the top 100, and then you go into the 728 00:34:03,720 --> 00:34:06,520 lab and synthesize the top 10. That is the funnel. 729 00:34:06,520 --> 00:34:09,600 That is the modern pipeline for scientific discovery. 730 00:34:09,639 --> 00:34:11,600 It's the future and it's happening right now. 731 00:34:11,840 --> 00:34:14,639 It really makes you wonder about the changing role of the 732 00:34:14,639 --> 00:34:17,840 scientist in all of this. Let's try to synthesize this 733 00:34:17,840 --> 00:34:22,400 deep dive now as we wrap up. We started with the old way, the 734 00:34:22,400 --> 00:34:27,199 slow, expensive luck based trial and error method. 735 00:34:27,199 --> 00:34:28,320 The shake and bake. Era. 736 00:34:28,320 --> 00:34:31,440 Then we walk through the modern computational toolkit, the whole 737 00:34:31,440 --> 00:34:34,360 ladder from Schrodinger's quantum equation for electrons 738 00:34:34,679 --> 00:34:38,639 all the way up to massive fluid dynamics for engineering. 739 00:34:38,639 --> 00:34:41,960 And we saw how applying these tools creates better catalysts, 740 00:34:42,080 --> 00:34:45,040 which leads to cleaner energy, more efficient industry, and 741 00:34:45,040 --> 00:34:47,639 things like faster charging, longer lasting batteries and. 742 00:34:47,639 --> 00:34:51,480 We ended on this incredible idea of AI surrogates learning the 743 00:34:51,560 --> 00:34:54,239 ways of physics to do most of the heavy lifting. 744 00:34:54,239 --> 00:34:57,520 It is a complete, undenied paradigm shift in how we invent 745 00:34:57,520 --> 00:34:59,240 the future. I want to leave our listener 746 00:34:59,240 --> 00:35:01,200 with a final provocative thought here. 747 00:35:01,280 --> 00:35:05,080 The lecture mentions inverse design and this AI generated 748 00:35:05,080 --> 00:35:08,680 library of 2.2 million materials that have never existed before. 749 00:35:08,760 --> 00:35:11,280 If we have computers that can predict a material that has 750 00:35:11,280 --> 00:35:15,160 never been seen on Earth, and then simulate its properties 751 00:35:15,160 --> 00:35:18,800 down to the individual electron, and then verify that prediction 752 00:35:18,800 --> 00:35:22,320 with a trained AI, are we rapidly approaching a point 753 00:35:22,320 --> 00:35:24,480 where the lab is entirely digital? 754 00:35:25,040 --> 00:35:26,480 It's a question of the century, isn't it? 755 00:35:26,800 --> 00:35:29,920 If the simulation becomes perfect or near perfect, does 756 00:35:29,920 --> 00:35:33,240 the physical experiment even matter anymore except as a final 757 00:35:33,240 --> 00:35:34,760 confirmation? Right. 758 00:35:34,800 --> 00:35:37,080 I mean, if the computer model, which is validated against 759 00:35:37,080 --> 00:35:39,920 decades of data says this new battery design will last for 760 00:35:39,920 --> 00:35:44,040 10,000 charge cycles, do I really need to sit in a lab for 761 00:35:44,040 --> 00:35:47,280 three years cycling it 10,000 times to believe it? 762 00:35:47,280 --> 00:35:51,480 Today, yes, you absolutely do. Verification is still king and 763 00:35:51,480 --> 00:35:54,320 our models still have flaws and approximations. 764 00:35:54,760 --> 00:35:57,760 But 50 years from now, when we have fault tolerant quantum 765 00:35:57,760 --> 00:36:00,280 computers running these simulations with near perfect 766 00:36:00,280 --> 00:36:01,640 accuracy. What happens then? 767 00:36:01,800 --> 00:36:04,160 Then the physical world might just be from manufacturing. 768 00:36:04,520 --> 00:36:07,200 All the discovery, all the R&D, might happen entirely in the 769 00:36:07,200 --> 00:36:09,360 cloud. The Eureka moment moves 770 00:36:09,360 --> 00:36:11,840 permanently from the lab bench to the server room. 771 00:36:11,960 --> 00:36:13,840 For many researchers, it already has. 772 00:36:14,040 --> 00:36:16,920 It certainly has. Cheeky Wang's lecture is a clear 773 00:36:16,920 --> 00:36:20,840 road map to that future. It's a dense, highly technical 774 00:36:20,840 --> 00:36:24,160 field, but the impact of this revolution is going to be felt 775 00:36:24,160 --> 00:36:26,320 by every single person on the planet. 776 00:36:26,320 --> 00:36:29,240 Absolutely no question. Thank you for joining us on this 777 00:36:29,360 --> 00:36:32,720 deep dive into the atomic scale. Next time you plug in your phone 778 00:36:32,720 --> 00:36:35,800 to charge, just take a second to remember there is a whole 779 00:36:35,800 --> 00:36:38,960 universe of mathematics and quantum physics going on inside 780 00:36:38,960 --> 00:36:41,600 that little battery. And chances are, a computer 781 00:36:41,600 --> 00:36:44,920 choreographed that entire dance. Thanks for listening, We'll see 782 00:36:44,920 --> 00:36:45,600 you next time.