 Calculate anything at all, okay? And I guess the answer to this question depends on what is really your question. So if your question is doing a prediction ahead of the experiment, sometime maybe your electron structure calculation is not the right things to do, okay? So what I'm plotting here is a machine learning model to predict the code temperature of ferromagnet. And every point you see here is an experimentally known ferromagnet. There's about 2,500 of them, okay? And as you see, this machine learning model that I don't have any time to describe in detail, predict on average with an accuracy of 80 Kelvin, okay? 80 Kelvin is better than what you can do from DFT if you don't know the answer. From DFT, you can predict accurate code temperature only if you know what the code temperature is. So if your question is, can we predict new magnet? Can we predict the new code temperature for magnet? The answer is probably you don't need to do DFT, you need to do something else. Now, I'm going to come back to this point at the very end and try to show that in reality, you need a combination of the traditional I-fruput and methods like machine learning and so on to really achieve knowledge and precision, okay? So the two objective of this talk are essentially to show that you have an I-fruput strategy to design your magnet and I'm going to tell you in a second why we pick up magnet. And hopefully that if you're clever, the databases can be augmented with different methods and sometime you can become hopefully more clever to use it and to construct those. So very, very quickly, as I said, we want to show you that you discover new magnet. Now, there is, if I have to sell this to a funding agency, there's a lot of argument like you want to do new magnet. There is a big push for diversity in design of magnetic material because the element are in supply from area which sometimes are geopolitically not very friendly. Also because a lot of magnet are made by rare earth. So these are expensive, this table, the color is the cost of an element per kilogram and sometime you have also new technologies that require to develop magnets that you don't have available at the moment. So there's a lot of reason why you want to develop new magnet. Now, if you scrap all this, the reason why we do that, the magnet is very interesting, okay? And so these are two main things that you can think about magnet. So first of all, funding a new magnet is a relatively rare event. So this is an Instagram that we put together many years ago where we list any known magnet as a function of the ordering temperature and you see that the magnet that you can use which means some things that survive well above room temperature, not that many. And in fact, most of them, magnetite, carbon and so on, are known since ancient time. And in fact, the earliest magnet I know about, it was reported in a book written in 79 AD by Plinius Vielder, is the guy that described the eruption of Pompeii in fact. And it report magnetite in fact. So 2,000 years, 2,500 magnet is what we know. So it's not really great throughput if you think about it. Now, the other thing is magnet at some time define completely your chemical physical intuition. So if you pick up this perovskite here, this perovskite is a 1,000 Kelvin or so antiferromagnet. And if you look at the chemistry of this perovskite and you look around the technician in the pithodic table, you don't find any hyperforum magnet at all in fact. So you find other paramagnet or a very lousy antiferromagnet or in the case of slonsilutonite, a very lousy ferromagnet. So somehow this seems to be an anomaly in your chemical physical intuition. So magnetism is defined intuition and is rare. So that's why we want to look at it. Now, if this turn, okay. So here what I want to give you is some feeling of how we construct a pipeline to look at new magnet. This was really work done in collaboration with my friend in Duke Stefano. And so let me very quickly go through the methods. The method was described in this paper some time ago and more or less is still the same. So essentially you have three steps to do. First of all, you have to calculate things. So you have to construct libraries of materials which would include existing material. So something you know about it and then library of new one that are constructed according to some criterion that you decide. So and this is up to you and up to what you want to look for. And what we do here, we use the most probably rock solid electron instruction method. We can think about it now. The point is this database grows. So once you start with a method, you are more like stuck because otherwise you have to restart from the very beginning. So this is where we are at the moment. You have to put this in a database and you need a very efficient manager to construct this database to handle the databases. So this is done in this code, it's called a flow developing Duke. And then this is really the interesting part that then you have to search. And search means that you have to come up with the criterion to take material on or throw material away depending on what you're looking for. So let me very quickly describe these three steps. The first two are the, I'm going to go very, very quickly because many of you might have seen this. First of all, you have to calculate everything is being made. And if you take one of the crystallographic, probably the most comprehensive crystallographic database you find about 150,000 in organic compound. This is, everything has been very well characterized crystallographically. So that's it. So that's, you can calculate all of those and you can read what it is. So this is not redundancy material come many times because the same material has been characterized differently over the years. If you remove what is a duplicate and maybe you have few more that don't make the database my kind of assessment that this number is probably right. So, okay. So that's more or less what has been done. And essentially we have about 100,000 of them. So we calculate all of those. We left some things out simply because there was no immediate interest to look at those particular material not for any other reason. Now, and just want to put this slide on because this is of course is a lot of calculation. So would make a lot of sense to share this calculation among different partners. So this is a, I think this has grown now but this was a snapshot of the consortium that was put together by Stefano. And those are people who essentially agree to share the data. There's actually not much more behind that to participate of our program to share the data. So every time I calculate something I just pass the information to the database. Of course, as a standard, you cannot just do the calculation as you like. So as a standard about convergence about, well many of them. And convergence about methods to calculate this on. But I know the data share. Now, then you have to construct new libraries for new materials and then you have to come up with a criterion. So you have to decide how you want to make this library. And what we decide to do is to construct a library of Oisler Alloys. And the reason for that is, well first of all, Oisler Alloy are this essentially intermetallic compound. If you take four FCC lapses, you interpenetrate it to each other. In such a way that you ended up with two tetradylacognitive side and two of tetradylacognitive side, you made an Oisler. And depending on how you occupy these sites, you have the regular, the half, where you miss one of the tetradyl occupation and then you have the inverse where you mixed the atoms that appear with a twice in the formula unit. You put one in the tetradylacognitive coordination. Now, we start from those because experimentally you can plug element according to the color code in the various site and you have a lot of them. So in ICS-D these are about 2,000. A good assessment tell you that are probably about, sorry, they're about 200. A good assessment tell you that are about 1,000. So 800 have not been characterized crystallographically well enough to make ICS-D, okay? And there's about, probably it's about 80, which are magnetic. So what we thought was, this is a great place to start because we know there are already a lot of magnets. And the reason is simply because you can pick up a lot of those just using element in the 3D part of the periodic table, okay? So what we did was we construct all the Oisler that you can make out of all the element you see in blue. And if you run the combinatorial analysis, you find out that you have about a quarter of a million prototype. If you want to play a little bit with magnetic order and so on, it turns out that we did about half a million calculation. So this is a simple calculation, a unicell type of calculation, but you have to do half a million of them. So that's, I guess, that's the good news and the bad news. Then everything goes in the database and here I don't really want to spend any time because I know nothing about it, but essentially the database is still now in this library. So this is a front-end. If you can do a front-end search, you get a bunch of property. This is just a crop of what you can get. But of course it is a back-end where you really now do the analysis. So when you calculate things, access indirectly the database. Now, and this is actually where now the interesting part comes. So now you have to find out what we have in the database. So the first question you want to ask, remember, we want to find new magnet and I want to go back to this experimentals and ask them to make them, okay? So the first thing I have to ask myself is, can a new magnet be made? Okay, and the very first things, the very first thing you have to do is, you have to find out the thermodynamics of the new materials. Now, what I call the zero descriptor is you have to try the enthalpy of formation. If the enthalpy of formation give you a positive number, so the enthalpy of your new material is larger than the enthalpy of the considered element in the most stable solid phase, then probably there is no go and you can throw the materials away, okay? Now, if you start from that, you start from a quarter of a million and you find out that you have only 35,000 that pass this first criterion. So 90% of the things you have calculated essentially goes away, right before you start. And where there's about 6,000 which are magnetic? Remember, we have about 4,000 magnet. So if I found new 6,000 new, that would be a fantastic deal, okay? You will find out it's not quite as that. Now, a much more serious analysis of the thermodynamic stability can be done by constructing the convex hull. So essentially you have to check whether your prototype decompose against the stable single element solid phases, but also whether it decompose, whether it doesn't decompose across all the possible combination of other phases, other binary or ternary. And when you start to do this analysis, then you have to check for all the binary and for some of the ternary, if you have some of the ternary. And this is of course become much more intensive because it means that for every choice of three, you have to construct the convex hull along the two dimensional edges of your phase space. Now, and then you have to do this. And this is an example of how one of those plot look like the dot are the one that actually are on the convex hull. And for this particular case, we found that this material seems to be stable according to the convex hull. Now, fun enough, I didn't know about it. This was not in my ICSD and we predicted stable and then I find out was reported in some Chinese literature about a year ago. So I consider this as a blind test because I didn't know it existed, okay? Although it existed already. Now, my point essentially is before, so here now you have to do many more calculation this is half a million because you have to construct all this binary. So you estimate that you probably need about between five and 10 million new calculation to map out the entire convex hull of all those oscillators that we made. So to start with, we decided to look only at the subset of them, simply because we have already the binary for those, okay? So then it's something we can carry on the analysis right away. So from now on, I'm going to talk about only a subset of a quarter of a million, this 36,000 that are only the intermetallic. So made of a three, four, and five D transition meter. Now, if you do them, what you find out is then you know what is the convex hull stability, okay? So you start from this 36,000, 248 have been predicted stable, only 22 are magnetic. So 36,000 to 22. Now all those are new, okay? So none of them has been made before. And then if you look at the robustness of this one, which essentially means how stable are from the most stable decomposition, you find out that only eight are within a KT from the convex hull. So actually only eight are something you can probably stand by, okay? Now, if I extrapolate, I can predict about 140. Now we have map out the entire database space and I can tell you that we have about 70 which has been predicted magnetic and stable. Now, if you remember, 80 are the known one. So we predict less than the known one. And if you start to look at the interception between those, what you're going to find is essentially they look more or less like this. So what we have is we predicted correctly about 40 of the known one. We have about 30 or 40 new. But we miss completely about 40 or so of the one that exists and we predict unstable and in countries are stable. Now, of those one, I would say about the third are just about the convex hull. So any sort of a level that you have in your function, any DFT level can swing you across the convex hull. Also, of course, we calculate the total energy. We don't calculate the enthalpy. So maybe there are some other contributions that actually can swing you across the convex hull. So those, I would say, are still in the midst. We have about another third which is between 10 and 50 milliV off of the convex hull. And here you can still save those because we have look only at the very small part of the magnetic phase diagram. So you might have other magnetic configuration that may be at the ground state and are more stable. But then we find a third which are significantly unstable. So 100 milliV plus above the convex hull. And those probably these are generally metastable compound that you can predict by looking at the global stability. I guess what I'm saying is if you accept this metastability criterion, then you have far more than only 130 or so. Now, the final thing we do is try to predict the thermodynamical quantity, which is the cooling temperature. And for doing this, as I mentioned, doing from DFT, you are very likely to get this wrong. And for Osler, it's even more problematic because experimentally it's known that the Osler are prone of site disorder. So you start to actually have to mess it up with your occupation of a various site. And this affect the cooling temperature. So what we decided to do was the following. We pulled off the experimental cooling temperature of all the Osler we know about it. And we tried to machine learn a regression which actually was a very simple regression to predict the cooling temperature of the one we didn't know. And it turned out that this regression works perfectly well with a 50 Kelvin or so accuracy. OK, so what did we find? So what are these 22? First of all, there are three classes. We found only cobalt 2 and 2 transition metal, manganese 2 and 2 transition metal, or a non-magnetic transition metal, manganese, and another non-magnetic transition metal. We didn't find anything else to be stable. So only these three classes. Now here what I try to do is to plot everything as a function of everything. You don't need to read this very careful. But I guess what I'm saying here from when you start to plot these things, you find out that something, you start to see pattern in your data. So here, for instance, I'm plotting the magnetic moment per formula unit as a function of the number of valence electron. And you see that the manganese one remained constant. So they seem to be independent on what is the chemical environment. So just as long as you have manganese in that particular occupation, you have the same moment. Now, if I look closely, essentially, I found three things. First of all, this cobalt 2 magnet. What we found is that the moment scale with the number of valence electron linearly. This is actually what the regression found. And this is more particular for this class that the CODI temperature also scale linearly with the number of valence electron. Black dots are experimentally non-compound. The red one are the new one we predicted. Now, this is not a great news in the sense that this regression just found the Zlatan-Powling curve. In fact, it's known that this class of magnet with cobalt 2 actually have a Zlatan-Powling curve. Now, the only thing I want to stress here, look, we found one of these cobalt 2 manganese titanium, which is a predicted CODI temperature of about 900 Kelvin because it's a very high valence electron count. So if this is real, that would be a pretty big deal. Because as I mentioned, we discovered magnet at very, very low speed. And we found one which is high performing with a very high CODI temperature. So remember this guy. Then we found the second class of this chemical composition. Now, here what the regression found, essentially, is that the Tc is dependent only, essentially, on the distance between the nearest neighbor manganese ion. And in fact, what you can find that are bound by this red curve. So they depend a little bit on a valence count as well. Now, this, in fact, is a known thing. So in the 60, people has developed this curve called the Castellis-Conamata curve, which essentially has an empirical curve that people drawn for manganese containing compound. And in fact, they depend only on the volume or the manganese-manganese distance. And what we found is that all our compounds are within those curves. Now, the interesting part here is that because of this formula unit, so you have only one manganese but formula unit, the typical manganese-manganese distance is large. So we have a very far edge of these curves that peak at about 3 amps long or so. So we predict that all those as a pretty lousy cooling temperature. Now, and this again, this compound having the cooling temperature, depending strongly on the manganese-manganese distance, are known. And in fact, what you can show is that for some of those, you see some of those here, the cooling temperature scale with the pressure. So if you change the manganese-manganese distance, you change the cooling temperature. Now, the final lot that we found is this manganese-2YZ. Now, this is we couldn't predict the cooling temperature. For one simple reason, that all the known one with this composition crystallized in an inverse or is less structure. All the one we found, sorry, these are the known one, all the one we found that crystallized in a regular or is less structure. So we have absolutely no data to do the regression on. So we couldn't predict the cooling temperature. What we know about those, that these are complex. So here, this is an example. This is my favorite one. It's something we spend a lot of time to characterize both theoretically and experimentally. This manganese-free gallium is a known fully compensated anti-ferromagnet where you have one sublattice pointing in one direction, one sublattice pointing in the other direction with some counting angle, and the two-angular momentum pointing in another direction as well. So you have a four-spin all pointing in different direction. So this is something you can get out of simply high throughput methods. So it's something that you really have to go with a heavy artillery. So this was done by combination of neutron, synchrotron, and DFT. Now, I think I skipped this in the interest of time. Just let me scroll through a couple of slides. And you're very welcome to ask about those. So this is all very good, right? Does this work? And there is only one way to prove it. Go to your experimental colleague and ask to grow some of them. And we spend so much time to try to assess metodynamic stability because I know that I've only limited number of shots in my experimental colleague. If I give them materials that cannot grow at some point they will lose traces and will not do it anymore, okay? So we went then with four compounds and we say, can you try to make those? So we pick up this cobalt to manganese titanium and then three of the manganese two containing material. So this is what you do. So you go in the lab and you make them. Now, so you make by arc melting. So this is me doing it. Now, of course, nobody believes that this is the way it happened. The way it happened is more like this. So I was watching somebody else doing it. Anyway, so if you go to the lab, this is the result. So cobalt to manganese titanium, this was made successfully. It crystallized essentially as predicted with almost an identical lattice constant. And as you see, this is the magnetization curve at 300 Kelvin and for Kelvin is essentially identical. No hysteresis. This is a cubic compound. So actually there is very, very little magnet at this line and isotopy. And this is actually magnetization at a partial temperature. The squid goes up to 700 Kelvin or so, but you extrapolate at 940 against 930 predicted. Now is luck, okay? The regression give you a 50 degree accuracy. But the point is you went from the periodic table to the lab and you can make those material. So the metal seems to work. There is a second case. This is manganese two platinum palladium. Now this turned out, as I said, I don't have to see for those, but at least was made. Now it turned out this was made with a twister. So it is a tetragonal distortion that we didn't have, but anyway. Now this turned out to be an intriguing one because if you look at the moment as a partial temperature, you find two peaks. So there's two phase transition here. All phase transition between very low moment configuration. So this is a complex magnet. We did quite a lot of analysis and they look like it's about two ground states which are all heavy compensated, but all anti-ferromagnetic. Now, so let me go back to the question just to finish. So I hope I persuade you that yes, you can really have your throughput and go to the lab eventually. This is a massive effort that goes with it, okay? However, I hope that as we will learn how to do these things, we'll learn to be more clever in selecting the prototype at the beginning and we will be more clever to utilize the data for other information. So the algorithm will be clever, but we need to be more clever than the algorithm. And finally, these are the people who did the job, people in my team and people in Stefano's team. Thank you. Thank you, Stefano. Thank you.