Story 1 wrap-up — one more bottleneck may now be solved

50:19we have four really good rundown stories that you guys are going to love this week. So, our first story in the rundown is about teaching computer simulations about atoms to learn from over 150 years of hard-won experimental data and why that matters for building the next generation of materials. So, what is this first story about? >> Yeah, so they've developed a new machine learning method. It's called DDOS, which is descriptor density of states. For those who are in solid state physics, they'll know what density of state means. I don't want to get into it, but effectively, what it's doing is it's bridging the gap between atomic

50:59simulations where we have simulations of materials, right, where atoms interact one way or the other, and then we've got experimental data, and we want to make a prediction about undiscovered materials. This particular paradigm, which uses AI in the loop with experimental data in the lab, makes that discovery of undiscovered materials 10 times more efficient. It's It's very cool. The idea is, you know, when you're in search of a metallic alloy, let's say for like, okay, I want to build like a turbine or like a fusion energy component, what I really care about is whether this material is not

51:40going to warp, it's not going to crack, if there's heat fluctuations, it's not going to act weird, right? The essential step in understanding these properties is to calculate something called a vibrational free energy, which is the idea of these atoms sort of vibrating at certain energies, and whether that is going to rattle the crystal lattice that the atoms are positioned in, right? That crystal lattice and that vibrational energy strongly influences something called phase stability. We've heard of phases in normal matter like solid, liquid, gas. Well, in solid materials you can have like micro phases in some sense because like the atoms can be in

52:20one figuration in one configuration in the crystal lattice, and then at certain temperatures they can transition to another configuration. And that's going to change stuff like heat capacity or like the stress or the bending ability of these materials, and that's really important to figure out, right? For example, that I mean the thing that comes to mind is like iron workers >> Mhm. who then heat like a thing a piece of iron on or any metal in super high temperatures and then they start banging on it to shape it because it's now more malleable Yes. because of the temperature that it's at. >> Yeah, yeah, yeah. And that And what you're actually doing there is you're

53:01like getting rid of impurities because at that temperature the impurities have a much higher probability of like leaving and so on and so forth, right? That's a That's a fundamental phase that the material is in. Mhm. Now, for the longest time what we've done is do simulations of these materials based on how the atoms interact, and then we've tried to compare it to experiment, and we've been off. Yes. Okay? Because there's a lot of things in between simulation that we're maybe not capturing that actually happen in experiment. That I'd just briefly We actually did a great episode on this exact concept, which if I recall correctly was when we talked about hypersonics >> Yes. trying to do simulations with the

53:43Navier-Stokes equation and where for supersonic flight our interpretation which is not real world accurate to the actual environment because it's so complex that we can't actually model it to the level of sophistication necessary to do a direct real world sim. So, it's there's a little fuzziness. Yes. And at different levels of stuff if you get a little bit faster the the simulation kind of breaks down a little bit. >> Yes, that's exactly right. And and what these guys have done with this D-DoS machine learning paradigm is they've captured all the possible ways that the atoms can be arranged in a material via

54:24some kind of probability distribution that is inside the latent space which is like the middle of your machine learning network, okay? So, the machine learning is learning how the atoms interact in the middle of its machine learning framework. And then from there, we can now start predicting what the phase diagram is going to be. The phase diagram is something like, you know, if I have pressure and temperature, where is it going to be a liquid? Where is it going to be a solid? Now, that's a traditional phase diagram, but if you imagine for these solid state materials, where is it going to be in a certain crystalline phase versus another crystalline phase? What is the heat capacity going to be in these two regions, right? And what this machine learning paradigm can do is now instead

55:05of a just a feed forward mechanism where I start from the atoms and I try to figure out what the phase diagram is going to be I try to figure out what the phase diagram is going to be. It's going to be inevitably off. >> Mhm. I can feed that back and inform my machine learning to do better. Mhm. Right? And the fact of the matter is this D-DoS system is a smooth function which means I can actually go backwards. I can start asking, "Hey, what if I want a certain material that is that has a certain property at 2000° C, right? It has a certain stability. I can go backwards and do an inverse design. I can start

55:46with the material and what the properties that I want and