 Če več, ki nekaj je, zato je poživite s ovemi, pa daj sem začutil vizucati, ko ampak prič militarya in če je zelo jeovost. Zelo se omreja na visku modem, in je, da je vziv del kom Brock v enoj svih, je zelo jazno začutite, ampak je zelo sku poveča če je začutila. In so je zelo iz nič del v potvečnji 500 supergupitelj. vseč na komputingih spremljane. In zelo, da se tudi imelimo, da se zelo pričelimo, kako da se načinimo v tem radoče z vsečenim vsečenim vsečenim kapacitim. Zelo, da se pričelimo, da se vsečenim načinim vsečenim vsečenim je materijalne zelo, da se vsečenim vsečenim nekaj, da je vsečenik, in z vsečenim, In ničč nekaj ne bo, da je tako očetnje, ko je s vročenja bo, da je to še pocivjivo, do vrjučenju, pa se razlicenje. Moje pri impečno je, da b fish na vrče, da nje ga je tukaj zvok držet, ki nimi kot se prišli, katiljno, da pa se počeš. I ne vedno bomo, da zelo nere. tudi, ki je to do zelo tudi začitava izgovoril na toga hovi, je tudi o zelo tudi začitava znečenih materijali in zelo tudi začitava znečenih materijali. Kaj sem zelo daš zelo? So smo počkali v mašinu, zelo smo v nobelu, zelo smo izgovorili, kaj si je da je vso rožen? Vzelo so, da je vso došel genočnje, odbih bi tudi začitelje vzelo vzelo v mene, načine zagrije. Zdaj, zato se je odložila, da se, je to, prosim, počelti imenčni materjal, je Alagay v Novozelov, da imeli skorovni napoč in se začili vsega pravna vsega. Vzpe, da je pravno počelti valerija Nikolosi, da bi se vložila všeč vsega. Vspešno, da vedno lektro-kemikalne in trebala, za več vsega materja. Or, you can just grow it. So, we want to replicate this with a computational protocol. And so, what did we did? What did we do? We looked at databases of experimentally known compounds. So, we start from something that has been recorded and has been described, and most of the time has been described correctly. And then we try to figure out if that compound looks somehow layered. And if it looks layered, we characterize in full the three-dimensional parent and the possible two-dimensional child. And once we have done that, we characterize all the mechanical, electronic, magnetic properties of the two-dimensional material. And then we see if there is something interesting. And this is a very common protocol in high throughput calculations. And even in the particular field of two-dimensional materials was first done by Olle Ericsson and Risto Nieminen, this physical review ex, Richard Henning in Florida has been leading similar studies, Evan reading Stanford, and so on and so forth. So, let's say nothing new under the sun. So, in all of this, though, there is this fundamental concept that we want to be able to go reliably from input to output. And this is an example in which say we could ingest a structure from one of these experimental databases and we want to calculate the property and we want this to be done automatically without any human intervention and not only automatically but in a reproducible way and in a way where we can store all the results for further analysis, for further data mining, for further machine learning. So, actually to do this in a say systematic way required really the development of what I call an operating system for computational material science. This is an effort that we started with Boris Kozinski. When he was at Bosch, now he's a professor in Harvard and especially Giovanni Pizzi and an entire team of developers in Lausanne are pushing the development of this open source Python infrastructure that we call AIDA that tries to really cover what I consider all the key pillars in this field. That is the capability to remove from the user the need to follow up all the calculation. These are what we call the low level pillars of automation and data where the operating system takes care of the management of thousands of calculation every day on heterogeneous computing resources with asynchronous programming and couples those calculations to the database by automatically storing so without the possibility of say human error all the results in a database that is the appropriate structure for computational science and that preserves the provenance. So, we want to remove this from the users and have the users focus on this green pillar this high level workspace where they build workflows that are a little bit Lego bricks you start building a workflow that has a simple calculation and more and more complex calculations that then you might want to share and especially you might want to share all your data through pipelines to open repositories with standards that one needs to develop. So, I won't really describe AIDA there is plenty of material on the web tutorials, videos, virtual machines where you can download it but let me just give you the image that all these infrastructures sits either on your workstation in your group it can sit in a supercomputing center and you interact with the resources by actually preparing your calculation here and once the demon checks that all your calculations are appropriate it's in charge of submitting monitoring them and retrieving and the entire architecture that has been built to be agnostic to everything is agnostic to the code that you use to the property that you calculate to other mechanism like the transport to the remote resources even the backend databases can be swapped. And maybe if there is one image that I want to leave you with is the structure of the database well, the results of your workflow are stored as a directed acyklik graph so what you have is that everything that has been calculated in the course of a simple or complex workflow is stored with arrows that really relate and link the parents with the children so I call it a postmodern relation because a child can have many parents and you can traverse the graphs and ask complex queries a sort of simple example but this is a more interesting example from the exfoliation project in which exactly we ingest a material, in this case this vanadium oxon bromide and we do all the calculations that we need to do when they are all stored so say we look at the primitive cell we check all the position all the symmetries from that we prepare an input in this case for a quantum espresso calculation we relax the three-dimensional structures we find say if it could be two-dimensional if it is two-dimensional we start doing all the calculations including say figuring out if it's a metallic that is simple or if it's magnetic that is much more complex if it's ferromagnetic, antiferromagnetic ferromagnetic super cell calculations at the end when we have found the electronic structure we calculate the phonon it's mechanically stable and then basically we get to the end result and this was just for one system sorry to repeat this and we do this systematically so we can figure out all the materials if they are exfoliable what properties they are and then we can examine this band structure and phonon dispersions it's very relaxing actually to spend the time looking at this occasionally we do some science a cute example done by Tibosoje Marco Gibertini trying to understand actually phonon dispersions in two-dimensional material and one interesting piece of physics that emerges is that polar materials in two dimensions are not able to support the TOLO splitting that you have for optical phonons in three dimensions basically the longitudinal optical mode doesn't cost more energy than the transverse optic because you don't really have to fill up a three-dimensional crystal with electric field density and so what was splitting in the frequencies in two dimension becomes a splitting in the slopes so all of this it's stored in its raw form but it's also stored in a say human readable form that makes for great and bulky supplemental materials but then becomes useful for anyone that wants to go and browse this data in this let me actually sort of reiterate since we are talking about big data something that Gabor as mentioned yesterday and that was very enlightening to us this comes actually from a different project of creating this Gaussian approximation potential for ferromagnetic iron where we generate probably 100,000 ferromagnetic environments ab initio and originally we did this all very carefully and with what everyone would consider say a very careful sampling even for a metal like like iron well what it turns out is that because all these calculations that needs to feed a machine learning model and the calculation are in principle done with different supercells if you want the energy functional slightly different from calculations to calculation because we don't have absolute k-point sampling and so we really needed to go to this limit of absolute k-point sampling to actually be able to reproduce a correctly say DFT results like the thermal expansion and so that's what I think convince me that highly curated data is really what is needed for ferromagnetic iron. One of the challenges that we have happily solved in the past year is actually the ability to really run this thousands of calculations every day. This is an example in which we did 35,000 SCF calculations in two weeks in November in Pittsburgh and that really required rewriting the AIDA workflow engine with synchronous programming to be able to have all this individual calculation be continuously tracked and monitored and stored. So in addition to AIDA I think there is a huge effort that is done in collaboration with the supercomputing center to really rearrange the structure of services that are provided by supercomputing centers. We call this infrastructure as a service and this is say what we do typically at CSTS but we are starting to mirror our services at CINECA and as part of the max center of excellence that I mentioned later we are doing this with more supercomputing center and so the idea is that it's not that we need anymore just a CPU but we need file storage and in particular long term storage and we need the database services and all the services need to be integrated with a structure that in this case for us is really the AIDA infrastructure and so say if you are a computational scientist like everyone in this room you really work with the left side of this but this same infrastructure can actually be used to disseminate and document all your calculation and share them with the community at large so as part of this effort we have constructed a dissemination platform that we call Materials Cloud where it's open source where we provide educational material we provide tools some of the work that Michele Cioriotti showed yesterday on machine learning say chemical shifts or machine learning polarizabilites are available there and in particular we focus also on offering services like archival services for curated data sets that you can freely upload exactly as you do in the say archive preprint server and if your data sets happens to have been constructed with AIDA then it makes it very easy also to provide the raw data and the curated data and so we are working at all this infrastructure. Now if I go back instead to the science we were left with this idea that we wanted to look at materials databases and figure out if they were really three dimensional or they look layered like in this case you see what we have is that a material is layered in a dimension that is somehow unexpected it sort of spans a different unit cell it's interpenetrating there are zero dimensional manifold in between and so what we do is we construct first a connectivity between all the atoms using say van der Waals radii and then we analyze the dimensionality of this and if something looks promising we do all the calculations that we need now density functional theory for most of this material is good and in particular say van der Waals density functional theory we think that does a good job at reproducing the binding energy of this material these are 40 materials that have been studied with RPA by Thomas Bjorkman I would say compare fairly well with either the Langret Lundquist functionals or the revised Vydrov van Voris functional so with this this is our pipeline we analyzed half a million structure of which really after we cleaned up structures that had only say partial informations or partial occupancies of atoms we had say stoichiometric compounds that were 186,000 we used two databases the free code database and the proprietary ICSD and after looking at which structures were unique we had 108,000 candidates 5600 look layered and we did all the calculations that I discussed before that are summarized here in this say two dimensional table the descriptors that we use to figure out if our materials are layered and exfoliable or not on the Y axis in a semi logarithmic scale we have the binding energies very trivial and these other descriptors seem to work very well that is what we put here in the X axis is the change in interlayer distance in the three dimensional parent when we switch on and off so by removing the van der Waals interaction and monitoring if there is a change in the interlayer distance we figure out if this van der Waals interaction are important or not and the clustering of the data is very simple basically you see right away that we have more or less three different regions we have materials that have very large binding energies and we call those non exfoliable materials that have very low binding energies and dominated by van der Waals interactions and we call them easily exfoliable and then we have this green in between of materials that we think could be exfoliated and now we start to have experience of exfoliating materials there and that we call potential exfoliable so overall a portfolio of 1800 materials that could be explored and what is very reassuring is we recover everything that is known that is we find a graphine we find boron nitride we find black phosphorus we find transition metal decalcogenides and so on and so forth so now this is our say curated data set and is also our playground and the idea is well let's look at it and see if we find something interesting so first thing that we do is we classify it and that what is surprisingly it's not the class of say transition metal decalcogenides that is most represented there are many classes of materials there are materials that have never been considered like rare earth decalcogenides and the question here is which properties we should be looking at well something that one starts doing right away is band structure so looking at effective masses for holes and electrons those are very relevant for short channel effects where you are not really mobility limited but what you want is the right effective mass very asymmetric and so we collaborate with materializier that is an electrical engineer device simulator to look at IV characteristics of some of these materials with the goal that one wants to outperform a transition metal decalcogenides by a factor of 5 or 10 one of our specialities is transport and mobility instead where we calculate electron phonon interactions all over the brian zone and then we use this to solve a Boltzmann transport equation this was a nice example done together with Cholwan Park some of it also with Tibor Soyer and Francesco Mauri and Matteo Calandra of something that I really call multi scale and multi physics because what we are doing here is calculating band structures with GW electron phonon interactions lifetimes putting all of that in the Boltzmann equation and solving it and you see the results as by now we know we have seen examples of these electron phonon calculations in the talk of Tomeo Montserrat Chris Van de Valle you see as a function of temperature endoping the theoretical resistivity and the experimental resistivity look reassuringly comparable and so we have started to do this systematically now these are very expensive calculations and so Tibor has developed all the technologies to go and figure out together with Marco Gibertini the different valleys the different pockets as a function of doping and do all the extensive integration so that are needed just restricting it to that pocket in actually in a fat geometry so with open boundary conditions and using electrostatic doping and in doing this one learns actually something and this is an example that was our first validation before the high throughput in which we looked at different materials going up to black phosphorus and you see these are the theoretical mobility increasing and if you actually look at the valleys around the Fermi energy what you discover is that you decrease the number of valleys and they become more anisotropic because really what limits mobility is intervalle scattering so we have already again in the quest for the scriptor something very simple but very powerful that is really looking at the number at the shape of the valley their position and how to engineer this again it's a little bit like being in a candy store there are so many properties this is an example in which we looked at magnetic materials either magnetic insulator magnetic metals half semiconductors materials that have say a large gap on one spin channel and are semiconducting the other spin channel materials that have flat bands could be very interesting for say a single band hubart model for transparent conductors because obviously the interband dispersions are very small and so similar for plasmonic applications so this is something that we do with christian twizen we have done all these phonons so all of a sudden we can look at this phonon dispersion we find a number of materials that could have charge density waves both in the 2D and some of these in the 3D parent and maybe the thing that we sort of studied first and more extensively with Antimo Marazzo and Marco Gibertini was the topological character who moved all the rare earths and we studied the materials by tracking the Vanier centers according to the recipe of David and Alexei and we found a number of materials that were quantum spin-all insulator and actually the most promising of all of this was this one here with a point 15, point 16 electron volt band gap in density of functional theory and this is a material that was discovered in 2000 data actually close to in Minashez in the iron mines there it was a check expedition they call it a Giacutin gate they published the results on a crystallographic journal the results were imported in this crystallographic databases and so it came into a database and we did the calculations in this now the beauty of a Giacutin guide is that it's really a topological equivalent of graphene that is if you look at the monolayer it's actually an exagonal pattern of mercury atoms and if you look at the band structure well if you were to do a standard Ft calculation you would find the Dirac cone at the Fermi energy if you actually switch on spin orbit coupling that comes both from the mercury platinum, this is mercury platinum selenide you actually see that is an insulator and if you study a ribbon and this is actually doing the full mountain of GW calculation with spin orbit coupling you discover that this is a topological insulator with a huge band gap of half an electron volt but the story doesn't end here what I think is quite remarkable about this material is that this is the first example of the material where the topological character is given by the famous and celebrated Kain-Milly model and for this I just want to show you the band structure of the insulator calculated here at the DFT level with spin orbit coupling and you see what I can do now is a vanearized system and I can calculate the band structure just with the first nearest neighbor vanear model and it's only when I introduce the second nearest neighbor interaction exactly what is in the Kain-Milly model that I open up a gap and I get this blue band that really reproduce quite closely the green band that of course I can reproduce perfectly using all the long term interactions so this two vanear functions centered on the mercury do this blue-green job and it's really this term here that opens the gap and the standard second nearest neighbors and the implant sock don't really matter so experimentally start very enthusiastic we have been able to exfoliate the sample to grow the sample starting to do transport measurements in by now still multi-layers but also I would say the ARPES characterization looks quite promising this is Felix Bamberg and the diamond facility when we compare it with the calculation so let me conclude we say there is plenty of room at the top if we exfoliate we actually made the cover of national nanotechnology thanks to the blender skills of Giovanni Pizzi and we are looking at all these properties in this case so is it collective hysteria is it genuine innovation something in between I don't know yet very vast but all the interesting things have been probably found so we are sort of looking for the little gems let me conclude thanking the team this particular work was led by Nicola Muné but also with Marco Gibertini Davide Campo, Antimo Marazzi Giovanni Cepellotti and Giovanni Pizzi there is now a very strong group working on this infrastructure we are actually looking for seven new postdocs and engineers will post this very soon I call this our open science platform of course built on open source code starting from quantum espresso but we told this AIDA and materials cloud infrastructure funded especially by Marvel and Max and so conclusions I think I'm optimistic about this field I think it's genuinely going in the right direction but I think what it mostly need is the capability of generating data on demand wherever they are needed with this sort of workflows and in particular sort of accumulating curated data that can be used for machine learning thanks a lot for your attention