National Energy Research Scientific Computing Center (NERSC) NERSC 2022 Early Career Achievement Award Seminars, 13 Dec 2022

Previous Meeting

⏯

youtube image

►

From YouTube: Bin Ouyang: High entropy design principles for battery materials with fast diffusion

Description

High entropy design principles for battery materials with fast diffusion
Bin Ouyang, Florida State University

December 13, 2022
11:30-12:00 Pacific Time

High entropy materials promise a lot of opportunities in various structural and functional materials. However, research discussing whether it is useful for battery applications is just rising on the horizon. The first question to ask is, what can “high entropy” do to battery materials? Here in this presentation, we would like to provide a theoretical/computational answer to this question with confirmation from experiments. With great support from high throughput first principle calculations at the world-leading supercomputing platform --- NERSC, it has been found that high entropy provides a) more random atomic distribution that helps avoid the formation of detrimental chemical short-range order that hurts the Li percolation; b) a rich ensemble of local environments that can facilitate new percolation pathways for superionic conduction. The discovery a) has led to the discovery of low-cost battery cathodes that can charge within 10min to reach commercialized Li-ion battery capacity. The discovery b) has revealed a general mechanism that can turn bad ionic conductors into superionic conductors, which creates a new design paradigm for battery electrodes and solid- state electrolytes.

A

Yeah thanks a lot for uh joining uh my talk, so I'm, mostly like a computational material scientist and now I'm a first year assistant professor at the Florida State University I, in the Department of Chemistry and biochemistry so uh today. I just would like to share with you some of my computational understanding of how to make use of a special class of materials which is called high, entropy materials into design Next Generation batteries.

A

Okay, so uh let me start with a brief introduction, just in case uh some of you guys are not so familiar with what is going on uh recently in battery field. So if you take a look at all those commercialized batteries, uh particularly canceled materials in your cell phones, you know laptops and electric vehicles. uh It's kind of funny like there has been developed for uh several decades, but still using pretty much like three to four medals.

A

uh It's pretty much like lithium, carbon oxide and Elysium nicocobo, manganese oxide, which we'll call them nmc or in Tesla we're using NCAA material. So all of them actually go with only three to four medals and they take the same structure, which is called the layered oxide structure. When you have this alternation of lithium layer and transition metal layer, so it's kind of funny like there are so many people working on battery materials. But we realize that only in three to four medals can work as a active Castle material.

A

So we we joke it as something called the nmc curse and nmc stands for Nicole, manganese and Kobo, which are the three main metals that we can put into a battery and the reason behind that is because, when you remove lithium, when you use the battery and you're gonna leave an empty the same layer hopefully, and there will be a tendency that the transition metal from the metal that you're gonna migrate from the octahedral site going through this tetrahedron test tissue and goes to the leasing layer.

A

This kind of a migration is actually detrimental, like first of all, we're gonna block lithium diffusion in the lithium layer. Second, because transition metal, typically smaller it gonna shrink the height of lithium layer. In that case, leasing will diffuse much slower.

A

So naturally, a good metal should prevent this transition, metal, migration and a chemistry going to tell you how to do that. So it's pretty much related to uh you know the transition metal, electronic structures where you have octahedral coordination. You have two e orbitals about three G2G orbitals. On the other hand, I'm going to have tetrahedron kind of reverse it, so naturally we need a metal that favors octahager and these favors tetrahedral and the trick here is just to put three or six electrons to occupy all the t2g states.

A

And then we can use this chemical principle to select Metals here. Just give example of all these 3D transition, metals and obviously like titanium, magnetium chromium are bad choices, because when they get oxidized, they have less than three electrons and which they're going to migrate. Manganese, two plus and nine or three plus have the same issue that they have high spin states with all the D orbitals occupied and they're gonna go everywhere so magnitude person. Ion four plus, does not have this kind of issue, but unfortunately they are young has ions.

A

So if you make a compound on it, there will be a structural Distortion, which is also unfavorable as a castle material So. Eventually, what we have is manganese, four plus, which is actually redox inactive. So if we need a redox action, you only have Cobalt and Nico as the choices. So this is actually the chemical origin of this so-called lnc curse that most of the battery in this field use owning, copper and Nico, particularly for this layered cathode.

A

So this actually reads up a huge supply chain issues. If you ever pay attention to critical materials, people always getting nervous about, you know we we're still gonna running out of this supply of Kobe and Nico, particularly considering we have more and more electric vehicles nowadays. So this is some of these mineral issues that we have so far like 60 percent of the Cobalt actually coming from one single country in Africa, which is called the Democratic Democratic Republic of Congo.

A

There are all sorts of issues with Mining, cobos and also another interesting fact is like 98 of This Global Production is actually a byproduct of other minerals. So if we have a huge electric vehicle, if we really replace all the gasoline vehicles, I mean this Cobalt will very soon become inefficient. So if you pay attention to Industry like Tesla like they are starting to move in from Cobo based battery materials into Nico based battery materials, because Nico have a slightly better situations compiled with Kobo.

A

However, if you take a look, take a look at the price of different Metals, you will realize. Actually, nickel is not definitely not the ultimate choice, because Nico is almost the second most expensive model, just right behind Kobo. So naturally we need something else right. So we are material scientists we just want to free the periodic table and the secret behind it is disordering.

A

So the disordering we can introduce here is like we can mix, listen with transition metals by doing synthesis control, and in that case we lose this layer, the topology, and in that case we don't need to worry about transition, metal migration, there's no leasing layer for you to block, and in that case, lithium diffusion will only be related with local structures, which we call them.

A

Zero TM units and listen can diffused really fast according to DFT calculations, and that can enable fast diffusion in this disordered form of materials which we call them before the rock salt. So this is the beginning of our story, so now um I mean my Apostle advisor and also including me trying to push a limit to commercialize these materials, because we essentially can free the periodic table by using all the other metals which hopefully will be more accessible and cheaper, compiled with Kobo and Nico.

A

However, such materials have a lot of puzzles. You know like. uh Hopefully we can use manganese, just give a manganese as an example. But if you make these two materials, they look pretty much the same. You know they have the same amount of lithium, which means you have the same theoretic capacity, and also they have the same rate or Center, which is manganese Surplus, and only difference is that they have a different inert Metals. So we put inert model just to charge compensate in the chemical formula.

A

However, if you make them into batteries, very strange thing happens. You know the capacity which is x-axis of this materials differs by several times. We should read about you know. Fundamentally, these two materials are different. If you do something like X or D, to see the crystal structure, the bulk structure are actually almost the same. You know they have the same X or D diffraction patterns. But if you do something like electron deflection, they start to show very different subtle features. The diffusive scheduling features are totally different. Those features are something called chemical, social enroller.

A

So it's essentially a measurement of the power probabilities of certain species in this random lattices and it can be calculated very easily by calculating the probability of funding B polling with a divided by the concentration of B. So if you get zero, it's fully random, it's random limits, but if you get non-zero value, it means you have chemical short range order, so you can either implement this into a free transformation and the intensity you get from this Fourier transformation is basically takes. The Patterson function form it's essentially the analytic form of this diffusive scheduling features.

A

You see, experiment which here I show like a diffusive rings of features. So this is actually the key to our disorder materials. It has been observed actually a long time ago dating back to 1950s in metal alloys, but until 2018, when I started my research on these fields. You know we realized this social kind of confusing I mean you cannot simulate it in a typical way. So this is like a casual simulation that I showed you like. On the right hand, side is a simulation of electron deflection by just randomly placed cations into this disorder.

A

Rockstar lattices I mean, as you can compile here. This does not really look that similar with experiments and the most important things like you lose the Symmetry information. You see the ring only in experiment, but we will barely see any symmetry in the simulation, so this is actually the computational challenging. How do we simulate a closer images as with respect to experimental observations? So if you take a step backwards and take a look at what we are simulating here so in experiment, when we do this electron deflection so Mike we have a micrometer size probe.

A

So, ideally we have a larger jumble of micrometer size of atomic structures and typically when we do electronic structural calculations, we have this DFT level Theory, which is quite limited into the Armstrong scale, or up to a few nanometers. So, ideally, in order to simulate the selection diffraction, we need like billions of atoms which in the micrometer size and naturally we need efficient hamiltonian, for you know, estimating this configurational energy, so the physical model we use here is called the class expansion.

A

It's the basically mean field approximation of the disorder lattices by estimating, as a linear summation of the different clusters in this lattices. So it's essentially linear model, but it can be a really large linear model. So essentially, gonna brings you, like thousands and even ten thousand features if you're looking at a really complex system. So such kind of a nature of this model kind of limits, the state of art, research on only in binary and ternary, solid solution problems. So this is which which this is actually the limit we're trying to push here.

A

So we're trying to really push our limits to the dimensionality of this compositional space of materials. And this is basically a demonstration of this code infrastructure we're developing here. So I I don't want to go too much of these technique details, but just to give you a flavor of what has been down here. So it's basically a machine learning problem which is pretty much like a compressive sensing of these important features which all of them actually have some sort of physical intuitions.

A

So we develop a different kind of regularization and group based regularization methods to decorate the features and trying to reduce down these features into a reasonable amount of a number, and also we're trying to apply some kind of active learning techniques to select the right sample, which gives us fast convergency when we enumerate different orderings and we're also trying to automate it so that all the jobs like thousands of DFD calculations can be run smoothly on a good platform like nurse. So this is pretty much what we developed in this code infrastructures.

A

Let me just show you how useful this code can be. So let me just go back to this election diffraction. This is not so good refinement. This is what we got after. We implement this class expansion tools and trying to simulate the election diffraction. So now we not only get a closer imagery with respect to the experiment, but we are also recovering the right symmetries, simply because we now have the right statistics with a large enough atomic model.

A

So I just give a snapshot of part of this atomic model here and because of this model, we know the real arrangement of atoms in the disorder lattices anyway, this kind of atomic Arrangement. Now we can quantify our properties if you remember I mentioned that in this materials, lithium diffusion is only related with local structures, which we call the zero TM channel.

A

So this is a local structure with only lithium in the corner of this tetrahedral and what we want to do to make sure we have good lithium diffusion as the same initially that we need a lot of these local structures and we want them to percolate as a network. So this is essentially the analysis we need right. So we can generate a large Atomic models of these disorder structures and we can do percolation analysis to quantify how many lists- and we can percolate in this highway of lithium diffusion.

A

So this is one example of this computational generated percolation map, so the color here it basically tells you how many list them can percolate in this disordered structure and X and Y are basically the compositional variables that you can control. So the first thing we did is just to verify if this really makes sense right, so we design several groups of experimental samples here are sample four to one at different composition and we make them into batteries and we cycle them to measure the capacity, and we see that from prediction sample four to sample.

A

One should have a monotonically increase amount of extractable, listen and experiment. We indeed see that the initial capacity goes up monotonically, which agrees well with the simulation. On the other hand, if you go from the other composition Highline, so we from sample six sample four to sample six we're gonna minimize the extractable amount of lithium. Usually we do that for other purposes, and in this case we also get a confirmation from experiments that sample six indeed gave you the list. You know percolating amount of listening. So with this kind of verification.

A

Now we are confident about our model, so the next thing we want to do, hopefully is a good part right, so we can scale it up very easily. So here are four promising systems that we're looking at all made of Chip Metals, the nickel compound their manganese compound, the waste titanium and niobium as charge compensators, and all of these color Maps is kind of interesting because they have similar kind of evolution going on here.

A

So let me just take one of them and do me at the fixed lithium content, which is fixed x value with different y value, which is the foreign concentration as per final unit. We see that in all four all four systems. First of all, we have most of these populating amount of basin below random limits. Secondly, it first goes down and then goes up until it hits the random limit. So apparently, if you need good performance, what you need is to go to hypherination right.

A

We try this actually for quite a long time, which unfortunately, doesn't really work so well, because if you are working everyone, any kind of oxymorite doping fluorine into oxide is very difficult. However, it's kind of a by accident, we realize there's a short curve here. You know we are trying to approximate random limits, so in other words, what we need to do is actually to make a material that is randomized enough.

A

So the shortcut is essentially that we can put more Metals into the lattices and the idea that behind it is like when we have more Metals, we are diluting the dominated concentration of certain metals and, in that case, we're diluting the dominated power interactions. So parallel probabilities are actually the short range older origin. So in that case we can weaken this short range order, which seems to be detrimental to lithium diffusion. So, to remember this idea, we did a high school put face stability calculations of these Castle materials.

A

We screen the periodic table ending up with, like 7965 of these promising Castle materials and before I show you uh these computational data. Let me just quickly show you a proof of concept for experiments, so we pick up a few materials. Those are the concentration composition of these materials we selected. So the first thing we want to verify is that we indeed get a more random material right, so naturally we can do electron diffraction. So two generations matter give you very significant short range of other features for Transit metal.

A

You can barely see them for six transition. Metals. It almost disappear. There is only in diffract, uh picks of which is actually from the long range artery. So this is kind of a confirmed. High entropy means more random. However, we don't know about the performance right, so we make those battery materials in into uh into Castle then cycle them. So indeed for six transition metal. We get a good initial capacities and we later want to confirm.

A

This is indeed a lithium diffusion, because in in the in the field of battery, sometimes you get good capacity because of other reasons. So one way to test the lithium diffusion is to cycle this battery at different current rates, so we're basically charging or discharge this battery with different speed. So in that case, for two transition metal, what happens is life if you go with a high current rate, the capacity decays very quickly into 50 milliampere program.

A

But if you have the high entropy materials, you're gonna sustain about 180 milliampere per gram of capacity, which are kind of a highlight here, because if you're driving a Tesla, so this is is the capacity you are enjoying right now, but all materials have minimized amount of Kobo and we have all the other element to be much cheaper compared with Cobalt and, most importantly, this kind of a current rate means you charge less than 10 minutes to reach the same capacity as a commercialized battery in electric vehicle.

A

So in that case we are ending up with a solution. You know with more elements, but most of them are actually cheaper and more sustainable element compiled with the state of order elements and also with this kind of a high dimensional compositional spaces like a night element. Compositional space, we can still use our class expansion model to capture the energetics, as well as the evolution of chemical short range order and also the least simple collation. So this is how the way you know tailor the design of this Castle materials.

A

So again, let's just go back to the data, so this is basically a heat map of the model, compatibility of all these uh 7965 for entropy materials. We just compute it, and it's also kind of interesting that most of these high entropy materials have much lower three random limits compiled with lower entropy materials. So that basically wins about you know whatever composition you develop in general, they're gonna offer you reasonable, and sometimes even exceptional properties compiled with this low entropy counterparts. So this reads about that: such materials does not relies on any Heroes of metals.

A

It can basically work as long as you have a reasonable result, Center and some reasonable composition of elements. So this basically comes to the end of my first story. We kind of find a reason why we need to care about high entropy, Castle materials and the reason for that. To summarize it in one sentence is because High entropy kills those bad short-range order. In that case it facilitatively, simple collation and is so far from our Discovery, only work for this new generation of Castle materials, which is called the disorder rockzone.

A

However, as you're probably realize you know, battery is not just about castle right, so I'll have a list to use at least a matter as anodes, so there is no point to do high entropy in anode, and we also need the electrolyte and very hopefully for the state election right right, because it's just a safer and inflammable. So in that case the second question I'm trying to answer is like: can we really make also a high entropy large white? Can we get rid of these hero elements or critical models in electrolyte materials as well?

A

So this is actually really a hard question because we have a very different structure.

A

So we have varodial structures and materials as choice, so we need to find a generic principles for that, so before I jump into examples of material, let me tell you actually the pros and cons of disorder materials in terms of diffusion, so it's actually not always a good thing for going high entropy, so the energy landscape of diffusion of Aquila models in the disorder materials that can actually sometimes be bad for diffusion, because what you are seeing here is like randomized fluctuation of energy Landscapes.

A

So that means sometimes you listen or other Aqua metals can be trapped it in this local minimum and all the state are surrounding it going to be high energy, and it gonna be basically that listen. So this is definitely a bad situation and essentially, if you take a look at how connectivity is calculated, it basically appears the carrier concentration. This is actually a bad thing. However, there is a way to turn this around.

A

What we can do here is actually this order also create a diverse collection of local environments, so sometimes diffusion can be promoted because it's actually not the average property. As long as we can fund a low energy diffusion channels that are percolating, we can actually lower down the effective activation barrier and keep in mind the activation barrier actually goes exponentially, which can you know which can easily like suppress the compressive conductivity, which goes linearly with the number of accessible sites.

A

So this is basically a trade-off game and our winning point is like. We have a term that we can minimize, which goes exponentially so as inspired by this kind of a hypothesis. Then we just want to think about how can we Design This diffusion process in crystallitis? So before that?

A

Let me just show you: what is the composition of diffusion barrier like assuming that we have an atom diffusion from side, one to side, two, so actually activation barrier is actually coming from two sides, so the first one is actually the side energy differences of these two are states. Well, the second one is more like an intrinsic activation barrier, which dictates the energy state of the saddle. Point. Okay. So naturally, if you have other structure, so we have a well-defined side, one inside two right.

A

We have the same local environment, so hopefully you have one fixed constant of the side, energy difference. However, when you have a disorder materials, so the neighboring outcomes can be any situations right, so there will be a a variety of options, so in that case, you're going to have a distribution of the side, energy, and in that case it's possible that even in two distinct Wyclef positions, you're going to have overlap of energy state, so this is actually the game changer here. So let me just go one step forward here.

A

So in order structure, we can have fixed side, energy differences and if this side, energy difference is too large. Unfortunately, we have bad diffusion right, but in disorder structures we can as long as we can create this overlap. Those will be these three carriers that can diffuse with very low side energy differences among the neighboring sites and as long as those sites can percolate in a network, you can have high ionic connectivity. So this is basically a more detailed picture of of hypothesis.

A

In the very beginning, we can lower down the activation barrier by losing some of these carrier concentrations and then, let's just put that into a real material. Okay. So here we took one materials that we studied. We did a lot of high school screening in a nurse and this material is called nasikong and we have two crystalline sites, which is called the 6B side and 18e side, and it's originally coming from the sodium compound for sodium and batteries and naturally, in this material.

A

If it's older form, you have a fixed side, energy differences, let's take a sodium, the column, phosphate example. This is a very classic composition of this material and it's a terrible, ionic conductor. So if you do DFT calculations, you will realize there will be a huge side, energy differences of 0.85 EV. So if you need a good ionic conductor, this value should be lower than 0.4 0.3 EV. So this is a terrible conductor. However, if you create some Distortion, so here we use Distortion as the most important factors for this ordering.

A

The Distortion gonna distribute the side, energies and making the site energy distribution of these two sides overlapping with each other. So in that case, it's possible that you can have neighborly side having the same energy. So we did quantify this by calculating the percolating amount of lithium as diff at a different side, energy differences. So the x-axis tells you the maximum enabling site so, for instance like if you have 0.14, basically tells you the we are calculating the amount of lithium percolating with activation barrier less than 0.14 electronic votes.

A

So in that case, if you have undistalled structure, which is the older structure, nothing will be populating and there will be a abroad trench until you reach 0.85, which is to say that any differences right. So there will be basically a step function.

A

However, if you have a distorted structure which is actually the solid line you already gonna have 22 percent of the sodium percolating at a very low energy differences, which is 0.14, electron volt and some intermediate Point, like Point D. Here we already have 50 percent of sodium percolating with side energy differences, smaller or equal to 0.5, electron volts, which a lot of least a lot of sodium here is actually activated with low activation barrier.

A

So then we're trying to extend this to a more generic picture. We tried a bunch of other materials like lithium version, listen Garnet, which is a commercialized Electro, light Crystal structures. We see kind of the same phenomenon that Distortion gonna promote uh diffusions of listen regardless of structures. The only thing that difference is like the thresholds, the number of percolating amount of decimal difference, and in that case it's kind of confirmed. This is actually a generic effect that this order can promote the percolation of agromatos.

A

So now we kind of have uh more or less complicated, thoracic predictions, and then we move to the experimental side trying to verify it so and we synthesize three types of high entropy materials and the reason that we do high entropy here is because we can use entropy to stabilize large boundary Distortion. Otherwise, if you have a two distinct atoms, it will be very difficult to put them into the same crystal structure in the low entropy form. So this is a point of high entropy here, so we synthesize those three materials.

A

They have very reasonable, ionic conductivity. But what more interesting here is that all those low entropy counterpart here actually have odors of magnitude, low, ionic and conductivity compiled with the high entropy version, which means this is not. uh You know. This is essentially not a cocktail effect. It's not a linear combination of this low entropy counterparts, but actually some synergistic effect is happening which push up their ionic conductivity.

A

We also did some prepositions from xrd and neutral diffractions, just trying to get some evidences of this disordering tendency going on here. So uh what I I put on the left column is actually the high entropy materials and the numbers in the bracket tells you. You know how many occupancies we see in this crystalline site of these materials.

A

So if you will see both of them being non-zero means like it starts to get disordered, and we also kind of compiled this with the low entropy counterpart, which you know for all the cases lithium and sodium gonna stay in the lowest Energy site, which is also a sign of artery. So this kind of a conversation also confirm you know. This is essentially disordering that controls the ionic diffusion.

A

Okay. So, on base of this idea now you're trying to push limits to see how far we can go right, so we just try all the good sense in a good material, just make it high entropy to see like how how good the only connectivity can be. So we are still on the process.

A

You know to try and pushing the limit, but I have like to show you some of our results, which will be online uh December 23rd, and so this is one of the sample we made, which is a pretty high entropy with six transition metals, and we see a 3.3 milliseconds per centimeter of bulk conductivity, which is actually comparable with one of the best conductors, a particularly oxide conductor, which is a commercialized in the market already, and also we just made a bunch of these high entropy materials, and this we are happy to see that most of these materials can behave reasonably all those five compounds that I show here just look some random compositions.

A

They all have ionic conductivity larger than one milliseconds percent millimeters, which means High entropy can be actually a generic mechanism to boost ionic conductivity. So, in the end I'd like to acknowledge, you know a nurse which make all these things happen. We did a lot of high super screening calculations thanks to this world leading Computing uh infrastructures and I'd like to acknowledge uh the most of this. Work is actually done by my postal period, with otherwise from prophetical, theater and also uh Dr, zayan run and Dr.

A

Yansen are the experimentalist and working on synthesizing this compound and doing all those kind of organizations, and we would like to thank this uh funding agencies which supports the discovery of these high entrepreneurs in.

B

A

I also like to advertise a little bit. We are a new established research group since this August we already have like three members and two also going to join us very soon, and we're happy to see that Samsung decided to fund us on this High entropy bathroom materials immediately after I established, I'm, very excited, I I hope that we can find a new, uh interesting discoveries on on base of this topic. So that's pretty much about my presentation. I really appreciate your attentions I'm opening to questions.

B

um Yeah, thank you very much. We have time for maybe one quick question or so. If anybody has one.

B

um Oh, that's one less than one quick question about the materials, um so uh it was really interesting. The ideas you put forth is: is it? Are there any um differences in stability of the materials or does that have any implications for for safety.

A

Yeah yeah, that's that's a very excellent question. So ins in in terms of stability I mean in Battlefield. We we care about two types of stabilities. Well, the first one is the face: stability, it's like basically, whether we can make it so in this case, High entropy material have Advantage, because uh we typically synthesize this materials in high temperature. This is when entropy can stabilize those are multi-component materials, so they are very stable. Actually in a lot of situations, the second one is electrochemical, stabilities is basically different.

A

Component of a battery will react with each other or not so in this case, High entropy materials will not be different, so it's kind of depending on your Crystal structures and your Chemistry. So what happens in low entropy will also happen in high-end shopping materials. So that's basically my comment on on stability of such materials.

B

Okay, thank you. I also really enjoyed your slide that showed the workflow and the complicated workflow.

B

um Perhaps we'd like to talk to you more about that later, because we're thinking a lot about how to optimize various workflows was, were all those arrows and boxes and things performed at nurse or were they some of the off-site.

A

um I think partially they are professional nurse and uh we're actually planning on a matures project is eventually generating on Earth and we are planning to move that to a mutors project so that we can do High super generation of this uh class expansion models, so uh we're actually working on that. So we we are kind of developing code in the past few years and now we're trying to launch this automation online in nurse. Hopefully, okay,.

B

Okay, great thank.

A