Simulation of materials processing: towards design of new atomic and molecular layer deposition processes
This is a modal window.
The media could not be loaded, either because the server or network failed or because the format is not supported.
Formal Metadata
| Title |
| |
| Title of Series | ||
| Number of Parts | 10 | |
| Author | 0000-0002-5224-8580 (ORCID) | |
| License | CC Attribution 4.0 International: You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor. | |
| Identifiers | 10.5446/60712 (DOI) | |
| Publisher | 05jdrrw50 (ROR) | |
| Release Date | ||
| Language | ||
| Producer | 05jdrrw50 (ROR) | |
| Production Year | 2023 | |
| Production Place | Frankfurt am Main |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
| |
| Keywords |
00:00
SolutionFunctional groupMoleculeDensityMolekulardynamikRiver sourceMaterials scienceComputer animation
01:10
Stream bedChemistryStarvation responseBreed standardMaterials scienceChemical structureOrigin of replicationTuberculosisMetalSchwermetallSetzen <Verfahrenstechnik>MachinabilityZunderbeständigkeitErdrutschBreed standardChemistryPotato chipSiliconAddition reactionWalkingAtomic layer depositionFunctional groupOxideCheminformaticsInfrastructureComputer animation
04:41
Chemical structureOceanic trenchController (control theory)LegierenFiningsSiliconMetalZunderbeständigkeitMaterials science
06:11
Chemical structureAmalgam (chemistry)FormaldehydeSea levelMoleculePrecursor (chemistry)HalogenationSurface scienceAtomic layer depositionSubstrat <Chemie>Metabolic pathwayChemical reactionPhysical chemistryFluorideAtomic numberRadical (chemistry)OzoneMultiprotein complexWaterAluminium oxideCrystallographic defectCoordination numberStuffingHydroxylHalogenationIonenbindungLactitolMaterials scienceAtomic orbitalChemical structureSurface scienceFunctional groupCalculus (medicine)Sea levelAntigenTitannitridCell growthSunscreenGraphiteinlagerungsverbindungenTrimethylaluminiumTungstenThermoformingChemistryWalkingChemical vapor depositionIndustrial etchingPrecursor (chemistry)SpeciesActivation energyBinding energyMoleculeChemical reactionSubstrat <Chemie>FluorwasserstoffAtomic layer depositionOceanic trenchOxideProcess (computing)Stop codonPhysical vapor depositionCobaltoxideCatheterAreaFiningsHafniumAluminiumLegierenMethylgruppeYield (engineering)Computer animation
13:46
Surface scienceSubstrat <Chemie>Precursor (chemistry)Sea levelMetabolic pathwayAtomic layer depositionChemical reactionRutheniumCobaltFreies ElektronCell growthFormaldehydeHydrophobic effectCycloalkaneReaction mechanismRadical (chemistry)Chemical structureElektronentransferHydrogenSpeciesMetalAbbruchreaktionCobaltFood additiveCalculus (medicine)Human body temperatureMoleculeStickstoffatomHydrogenChemistrySurface scienceAtomic numberCopperPrecursor (chemistry)Chemical structureMaterials scienceReactivity (chemistry)Sea levelProcess (computing)Potato chipCatalytic converterChemical reactionZunderbeständigkeitBlue cheeseAmmoniaMetabolic pathwayElectronCross section (geometry)Argon oxygen decarburizationTitannitridGibbs free energyKupplungsreaktionOrange juiceAtomic layer depositionEssenz <Lebensmittel>Mixing (process engineering)Battery (electricity)Physical chemistryIslandPharmacyChemical vapor depositionComputer animation
21:22
WeaknessChemistryAtomic layer depositionSurface scienceLigandSpeciesCycloalkaneChocolateHalogenationAle MunicipalityNitrideZirkoniumdioxidMortality rateSpontaneous combustionMoleculeMaterials scienceVolatile organic compoundPrecursor (chemistry)Chemical reactionBulk modulusStop codonSubstrat <Chemie>FluorineZirkoniumCobaltoxideHafniumMetalFluorideStuffingPyridineStickstoffatomIndustrial etchingChemistryProcess (computing)SpeciesWaterAmmoniaGasErdrutschHydrofluoric acidBinding energyCyclopentadieneCobaltProtonationHydrogenSetzen <Verfahrenstechnik>Catalytic converterTitaniumHalogenationZirkoniumdioxidHydrochloric acidOctane ratingNitrideMachinabilityMixing (process engineering)Calculus (medicine)Aluminium oxideHydrogen peroxideRadical (chemistry)Materials scienceMoleculeLegierenAtomic numberSurface scienceOxideMolekulardynamikHafniumComposite materialTocopherolStop codonAtomic layer depositionCobaltoxideBlue cheeseHuman body temperatureMixtureCarbon (fiber)River deltaChemical structurePrecursor (chemistry)Computer animation
28:57
FluorineZirkoniumHafniumZirkoniumdioxidMortality rateCycloalkaneLigandCobaltoxideHydrogenAle MunicipalityHalogenationChemical structureKorngrenzeSubstrat <Chemie>Crystallographic defectOxideBlue cheeseOxidePhysical chemistryProtein domainXenon difluorideMaterials scienceSpontaneous combustionHybridisierung <Chemie>Cell growthChemistryPolymerMethaneDiolDimethylaminGlycerinSurface scienceTrimethylaminAdsorptionBarr bodyTransformation <Genetik>Reactivity (chemistry)Transition metalMoleculeFunctional groupPhenyl groupAction potentialAtomic layer depositionGel permeation chromatographyHydroxylStorage tankFunctional groupCobaltoxideBinding energyIndustrial etchingAluminiumHuman body temperatureStickstoffatomCoordination numberPolymerChemical reactionMaterials scienceLegierenEthylene glycolOctane ratingChemistryPorosityMoleculeFood additiveAromaticityGlycerinMachinabilityAluminium oxideCell growthOxideSurface scienceSpeciesOrganisches MolekülWattOxideGlykolsäureHybridisierung <Chemie>Chemical structureZirkoniumdioxidEnzymkinetikAlcoholAnimal trappingEthyleneAction potentialComputer animationDiagram
36:32
Freies ElektronRearrangement reactionElimination reactionProton-pump inhibitorAtomic layer depositionChemistryAtomMonoschichtCell growthSurface scienceDipol <1,3->Chemical reactionGel permeation chromatographyProcess (computing)ProtonationHuman body temperaturePrecursor (chemistry)VulcanizationIslandActivity (UML)Binding energySurface scienceVerfahrenstechnikCalculus (medicine)SpeciesActivation energyMachinabilityRearrangement reactionZinc oxideActive siteCell growthLigandOrganisches MolekülSet (abstract data type)Hope, ArkansasReaction rate constantBulk modulusMaterials scienceChemical reactionComputer animation
39:58
Hybridisierung <Chemie>Action potentialAtomic layer depositionCycloalkaneGel permeation chromatographyAle MunicipalitySurface scienceCoordination numberChemical structureMaterials scienceAtomOctane ratingMolekulardynamikAtomic numberFunctional groupComputer animation
41:40
MixtureOxideAle MunicipalityGezeitenküsteChemistryChemistrySurface scienceLevomethadonSpeciesIonenbindungPeriodateCobaltActive siteLecture/ConferenceMeeting/Interview
43:39
ChemistryHalogenationProton-pump inhibitorRearrangement reactionAtomic layer depositionDarmstadtiumFreies ElektronPenicillamineElimination reactionPolyvinylacetatHydrogenCalculus (medicine)AtomCovalent bondQuantum chemistrySurface scienceZincMeeting/InterviewLecture/Conference
45:24
KreatinphosphatCommon landMeeting/InterviewComputer animation
Transcript: English(auto-generated)
00:11
Great to be here in Rudisheim and meet some of the Bodstein people in person who I've been talking with for the last few months.
00:21
So I'm currently the head of group in Tyndall National Institute for Materials Modelling for Devices, where, as the name suggests, we do atomistic simulation of materials focused around ICT devices. And I'm the third Irish speaker this morning, so good contingent, except I'm from the sunny
00:43
side of the country down south. It always rains in Dublin. So if you're ever in Ireland, make your way down to Cork, where it's always sunny. So with that Cork versus Dublin out of the way, I'm sure a few people in the room
01:00
will know what I'm on about. Well, I'm going to talk about work we do in simulating processing of materials, mainly for electronic devices. So this is atomic and molecular layer deposition, and this is focused on first principles density functional theory based simulations, some of which we saw in the previous talks. But there's not a lot in the, let's say, machine learning informatics type space on
01:24
this particular chemistry. So, well, what we're ultimately trying to do is move from doing very large time consuming computationally expensive first principles surface chemistry simulations to doing those using these faster approaches, if nothing else, to screen out chemistries that won't
01:47
be useful. So in context, we do a lot of work with industry clients in the semiconductor space. So we've worked with the likes of Intel Applied Materials, who are all interested in new processes to deposit films of materials like metals or high-k dielectrics.
02:05
They understand that simulation is a good way to do it. It saves time, it saves person effort, it saves resources, et cetera. But even with that, these are still very, very time consuming and difficult simulations to do. So we'd ultimately like to make it possible for an engineer in one of these companies
02:25
to run a simulation of atomic layer deposition in the way they do now using software for classical processes like ion implantation and oxidation. But we're far from that yet. So these are the first steps to try and get there.
02:41
And I'm going to throw a few slides with some thoughts we've been having around what could be done. And clearly, there's a lot of people in this room who are further on than us, at least in the general space. It will be interesting to see what kind of insights people have across the next few days. So obviously, I don't do any of the work.
03:01
That's me. I just pay everyone. They do the work. Great team of students and postdocs. No, seriously. If you let me do it, nothing would happen. Lots of funding support and heavy use of computational infrastructure. This is the Irish National Supercomputing Service. So the motivation for the work we do is looking at the scaling of electronic devices.
03:21
You know Moore's Law, that the speed doubles every 18 months while the cost halves. And that's been driving this constant downscaling and increasing capability of our devices. So the traditional approach has been more and more. So this is keeping along the Moore's Law track, smaller nodes or smaller transistor
03:42
structures looking at new materials. So increased performance, but no real change in functionality. The transistor that's in your computer now is basically the same as one that was in the original Intel chip in the 1970s. Also becoming really interesting is more than Moore, which is also called heterogeneous integration.
04:01
So in addition to your standard silicon CMOS type devices, you have all these other things here, photonics, MEMS, on chip memory, biochips, et cetera. So Moore Moore's simple scaling, smaller and faster. More than Moore is a diversification, combining all these cool other technologies, not limited
04:25
to what you can do within regular CMOS. But it turns out that irrespective of which route you go along this graph, you still need to be able to make the devices, it doesn't matter what's going into them. Production of the devices is the key thing. So this is the recent Intel roadmap, now going towards the Angstrom scale.
04:44
So there's 20 Angstroms, right? And I mean, that is, that's amazing that they're already, they already know how to do 20 Angstrom scale devices. And you see, we're out here at 90 nanometers about 14 years ago or so.
05:01
We had high K metal gates and then we have all these things. But what's happened with this evolution is we've moved from what are called planar devices. So everything's flat and you're just throwing in material from on high, for example. But as we've moved down to the sub 20 nanometers structure, we're now making three dimensional
05:21
complex structures. They're also getting very, very small. So the thickness of a film is really important. You're down to the few nanometers. Also crucial is that you end up with these deep trenches. I'll show an example in a moment. So you make these trenches and the aspect ratio of those can be a thousand to one, which
05:41
is huge. So for all that, you need conformal film. So that will cover the entirety of your 3D complex structure with no holes or gaps, extremely high uniformity, fractions of a percent over a silicon wafer, which is 300 millimeters in diameter and very, very fine thickness control, in essence, layer by layer control
06:06
of thickness. Samsung have also put out their technology landscape. The green line here is kind of Moore's law, what you'd expect where we're falling off it as you might expect, but they're also looking at much more complex structures,
06:23
you know, two nanometer, one nanometer and bringing in exploratory devices. So all the major players are looking at moving from simple planar, like 2D like structures to much more complex three dimensional structures. And so the question for those of us in the processing world is how do we deposit films
06:43
of materials in these complex device architectures and remember the yield in a semiconductor device processing fab needs to be about a hundred percent for it to make money. So there's no room for any error. So you have to design for no defects that every single device works.
07:03
So what's used now and will only grow in importance over the next few decades is atomic level processing. So this is the umbrella term. We do it for deposition of materials and etch, so removing material. So deposition is the most well-known one here, atomic layer deposition, your computer,
07:22
your smartphone, your smartwatch, all have atomic layer deposition films in them. It's everywhere, it's crucial. The basic idea is you have a cyclic process involving two chemicals. So we start with a substrate here that's got some functionality on it. That's going to do some chemistry.
07:40
You then bring in molecule number one or precursor one. It's going to stick to the original surface in some way. And then the excess stuff will just be washed away by a purge. The key factor in this compared to say physical vapor deposition or chemical vapor deposition is once you bind the precursors to the available surface sites, everything stops.
08:04
It's self limiting. You can throw in as much precursor as you want after that and no more growth or no more chemistry takes place. The extra precursors are they're not going to react. So this gives you a self-limiting process. You then bring in the second reactant, which could be another molecule
08:23
in thermal ALD for making oxides. It's water or ozone. So there's your oxygen containing species. It reacts with the thing that's on the surface and it gives you back a layer of your target material and you purge away all the rubbish and then you cycle back around. In principle, each of those cycles deposits one layer of the target material,
08:46
meaning that just by playing with the number of cycles, you can control the thickness very, very finely. Obviously, in real life, you don't get one layer per cycle, but you might get half a layer per cycle, which is still useful enough.
09:00
And then, you know how to tune it. The second option is you can use plasma. Nitrogen, hydrogen, ammonia, oxygen creates these active radicals. These could be free atoms or chunks of a molecule that will then react at the surface from the first precursor.
09:21
And again, it'll bring you back to here. Your antigen plasma allows you to do things at lower temperature, because these species here in this cloud are all very reactive. So they don't need high temperature, unlike you do, say, say here. So you get ultra thin coatings on complex structures, high uniformity, and it's conformal. Etch involves very similar chemistry, but just does the opposite.
09:44
So what we do is we add a precursor to a substrate. You get this net, this new layer here. So we've modified the top of the substrate. This isn't volatile, so it won't go away. But when you add the second precursor, the magic happens so that the resulting molecule that forms here is volatile and goes away,
10:03
leaving you with a region that's been etched and you can keep going around and around and example and molecule that hydrofluoric acid, which obviously is really nasty from a practical and safety perspective, but does brilliant chemistry.
10:21
So it's our benchmark. And so we study, along with other colleagues, how you etch hafnium dioxide because it's widely used as a high K dielectric material. And but again, you have the same characteristics as you do for deposition, except now you talk about etching. And so this is the kind of thing you can get from ALD.
10:41
So this is a very, very deep trench in essentially a capacitor structure. And you can take different regions and you can look at the thickness of the coating. So, you know, your one nanometer difference here. That's fine. Doesn't really matter. Very similar here. It's only down at the very bottom that you get any kind of appreciable difference.
11:01
And but still, this is good enough for these to function as useful devices. If you did chemical vapor deposition or physical vapor deposition, you might get from here to here and then the thing would block. And all this region down here would never be coated. So this is the only way to coat these incredibly complex structures, which is great in further areas like health care, where you have devices
11:22
like stents and catheters that also have complex 3D structures. So we're seeing some growth in there as well. So lots of interest in doing first principles modeling here because it's chemistry at surfaces. And so the typical calculation that we would do is we're looking at aluminium oxide because that's the benchmark.
11:42
This is trimethyl aluminium, 3-methyl groups. You have hydroxylated surface, which has the surface OHs that are reactive. The trimethyl aluminium TMA comes in, binds to the surface and releases methane. And if you wait long enough, the surface is covered in these aluminium methyl species here.
12:01
You can calculate the energies of that. You bring in the water, you make ALO bonds here, ALOHs. So you recover the hydroxyls. You release methane. Again, you can calculate the energies involved and then you repeat again and again. So what we what we do in modeling is we normally look at one ALD cycle
12:21
under the assumption that what happens in that cycle will happen for all subsequent cycles. So we have look at proposed reactions, and then we do there's the functional theory level of the relevant energies and structures. So you will get something that looks like this here, this energy profile. So the y axis is the change in energy in electron volts
12:43
and going across the axis, the reaction coordinate. So this is an example of tungsten hexafluoride, which is a tungsten precursor on titanium nitride. And you then calculate essentially the interaction energy of the tungsten fluoride at the surface.
13:00
It's very, very small. I would have probably rejected that paper because that's not going to happen. But we'll just use it as an example. So I'll be kind to them. So you calculate these energy minima here as you move along. So we're eliminating fluorine along each step. OK, so six, four, three, one, and then no flurries.
13:24
And so we have these thermodynamic minima. And we also then have activation barriers. So that's the hill we have to get over in order for the reaction to proceed. So again, these are electron volts. So this is relatively moderate. OK, here's the highest energy barrier.
13:40
So this step here is the one that's got the highest energy cost in terms of the barrier. We don't have to make sure we run the the ALD process at a temperature that lets us get over these barriers. So this can give us an idea as to whether a particular reaction pathway will happen or not. However, these are this is just one potential pathway.
14:03
So this is a complex molecule with seven atoms in it on a titanium nitride surface, which itself is complex. And these guys have just taken one particular way to break up the precursor. But there's many, many ways to do it. But each of these calculations is probably three or four months
14:21
of PhD student effort. So it's pretty challenging. So just to briefly show you what we've been able to do, but also the limits on this. So we just finished a project on copper interconnect, essentially copper wires in a transistor is a cross section. This is the top bit that you'd see if you broke open
14:41
the the packaging on your chip. So nice big wires. That's fine. They transmit the signals as you go down, down, down, down to the transistor itself. The dimensions are in the nanometer scale. So if we zoom in to this region here, when we're putting in a piece of copper, it's actually a complex structure here. So you have a dielectric nonconducting material that stops
15:03
electrons and atoms short circuiting your device. You then put the copper in. But copper has a very nasty tendency at the nanometer scale level. It doesn't make conducting metal films. It makes little nonconducting islands.
15:21
These are great for catalysis, but they're absolutely awful for what we want to do here because we want a wire that conducts. So in order to plus copper likes to move around a lot. So in order to prevent all that, we have to take some of that valuable real estate and put in this blue barrier layer to stop the copper diffusing
15:43
through the dielectric here and creating short circuits. And we have to put in what's called a liner or a seed that helps promote two dimensional copper rather than those three dimensional islands. So if you have a very small volume to play with here because of your scaling of your device, then a very substantial
16:04
proportion of that volume is taken up by these two layers here. So there's a couple of things you can do. One is you can look at replacing the barrier liner materials, which Carolina did here in her Ph.D. and she found some really good options. But that's not the essence of this talk.
16:22
We are interested in alternatives to copper. So cobalt is one alternative that at these length scales, it shows a lower resistivity than copper. So it's a better option. It's not ideal because obviously the battery people like it and the mining of it is an issue. But it's a it's a good example of an alternative material.
16:44
Butenium is also possible, but it has its own supply issues, et cetera. So we went to model the atomic layer deposition using the plasma chemistry. So this is the first cycle again. This will be the metal containing species, the cobalt containing species.
17:01
And then we're going to model the atomistic chemistry of the plasma species, not the plasma itself, because that's physics and we're mere chemists. But the plasma makes various species radicals, for example, hydrogen atoms, nitrogen atoms, so we can work with those. So that's what we're going to going to put in here.
17:20
And we're going to try and look at these again. This is another of these energy profiles showing, you know, stable structures here and then these barriers here. So we'd calculate those things. So the first thing you have to do is sorry, you have to build a surface to work so very briefly in the plasma
17:42
deposition process, as you're growing the metal cobalt. So we take the cobalt guys here. After the plasma step, there will be nitrogen hydrogen species sitting on the surface, which are then available for the cobalt containing molecule to react with. So we need to work out what are stable ways to terminate cobalt
18:03
with nitrogen hydrogen species, NH, NH2. And basically what these graphs show is stability lines. This is an ab initio thermodynamics with free energies here. You've got temperature in Kelvin up here. And then these spaghetti lines are basically different ways to terminate the surface.
18:24
So we have different numbers of NH species, plus different numbers of NH2 species all on the same surface. Calculate the energies of those. Convert that into this graph here. And basically you pick a temperature if we pick 500 Kelvin
18:40
and you walk up and you see which of the lines you cross. So in this case, we're just about crossing the blue line. So 500 Kelvin, we would say that that's what the surface is terminated with. But then we were able to throw in experimental numbers from actual AOD experiments. So using ammonia and this cobalt precursor, 550 Kelvin.
19:05
This guy here, this red dashed line here is the most stable. So it has five NH species on the surface. And then you do similar for this other cobalt termination. So we're we're interested in how the the surface orientation 001 versus 100 impacts on the chemistry.
19:20
And it does seem to play a role. And that's useful because you can grow polycrystalline and materials, each of which may have a different reactivity. So we end up with these kinds of structures here. So looking down on a plan view, the orange balls are the cobalt species. Blue is nitrogen and white is hydrogen. And we have a distribution of NH species on this surface here.
19:44
And then on this other surface, we have a mix of NH, these guys and NH2s. They have this rabbit like appearance. So at the typical AOD conditions, that's the surface that you have. That's the surface that will then do the chemistry. OK, so then we we run the chemistry.
20:01
This is the cobalt containing molecule. Here's the cobalt atom. Here's the organic bits and the NH terminations. And we calculate a hydrogen migrating from the surface goes to one of these rings, the cyclopentadienyl. And that ring, now that it's happy, it's fully aromatic, it goes away.
20:21
The cobalt molecule in the precursor is sticking to the surface. So we've got our cobalt down. That's what we want. So it looks like this. So this is great. We calculate the barrier, the hill for that reaction. It's only 0.56 eV, so pretty moderate. So that's encouraging. To use this as a precursor, we can then build up complexity
20:41
and add two of these molecules, because in reality, you're not bringing down one molecule at a time, unlike in the calculations. So you run through the hydrogen transfer, make the CPH and you bind. And then the same can happen here. So we can go through this whole process. You know, but these five second movies are, you know, five months of calculations.
21:03
So they're far from trivial to do. But you can get out those barriers again and the stability of these things. Again, for us, these are quite moderate. So so that's useful. So we don't have a surface with the cobalt containing precursor on it. That's stage one. Stage two is doing the plasma.
21:23
So the typical radicals you expect from a nitrogen, hydrogen or ammonia plasma are NH, NH, NH2 just sitting in the gas phase. So the kind of model you have here, it's surprising or difficult to find a model of plasma chemistry. So please pretend that's a surface.
21:40
You have all these guys here, like these are just hanging around and they have some energy with them because they've been generated in the plasma. So they will bounce down onto the surface and react. So that's what we do, right? So we added H atoms. We added a bunch of hydrogens because that's the simplest radical. We did molecular dynamics in in DFT at 600 Kelvin.
22:03
So we only do relatively short simulations because they'd still be running if we tried to do longer simulations. And the white atoms are the nitrogen. So you see this mess of things. But if we look over here, we've made some CPH molecules.
22:20
So spontaneously, the hydrogen atoms bound to a carbon in this organic bit and made the organic, which then goes away, leaving cobalt at the surface. So you can run through all those types of calculations with all these radicals here and see what can be eliminated.
22:40
What was interesting was and unexpected was you can insert a nitrogen into that five-membered ring. So it gives you this thing here. This is pyridine. You know, we thought, OK, maybe there's something wrong there. We did the calculation again from a new configuration, but we keep getting pyridine.
23:01
So we've predicted pyridine as a product during the plasma ALD of cobalt in this particular setup. So our colleagues are looking to see if indeed that's produced, whereas you'd normally say will be the protons binding to the organic bits and releasing the cyclopentadienyl anion.
23:21
And then we can then as we add NHNH2, they will then start to bind to the cobalt that was on the surface. So you can get these kinds of things here. So what that's doing then is that's giving us back the active NHNH2 we need in the next cycle. And then these remaining guys on the surface,
23:41
they get eliminated as ammonia by the hydrogen species. So the hydrogen comes down, starts to protonate here. You get ammonia and that flies away with a very low desorption energy. So that eliminates the nitrogen from here. Now, you do still get some nitrogen incorporated, but we know in cobalt
24:03
you can get about 10 percent nitrogen and it still behaves as cobalt metal. So that's not too bad. So with that work, then we've done the first full atomic layer deposition cycle of plasma of the metal. So you build up your surface model here
24:21
that's going to then react with the metal containing precursor. And you get various structures for the metal containing precursor. Then you bring in the plasma generated species and you get all this stuff happening here where we've made ammonia. We've made pyridine. We've made CPH, deposited the cobalt, eliminated the organic
24:45
over some NH and NH2 species. And then you can keep cycling. So this was was done by G over during his postdoc. But just this single slide is a huge amount of DFT calculations. It's probably two, three million CPU hours used just to generate the results here.
25:06
Not particularly useful if your industry partner wants to find a plasma chemistry process in six months. So we can do it and we can get all the detail, et cetera. But it takes a long time. So hence the interest in transferring from from this to,
25:24
you know, machine learning type approaches like we saw in the previous talks. So briefly to look at the etching process is another example. And so I mentioned you can use hydrofluoric acid to modify the top layer of your substrate here. This is not volatile, so it sticks. And then you bring in some other containing molecule here.
25:42
Here we use that tickle and that gives you this titanium fluoride, which is indeed volatile and goes away. And each cycle of this etches a certain thickness. OK, so what's been done on this is looking at other high zirconia alumina, metal nitrides are of interest.
26:01
And my colleague Suresh in Synopsys has done the first calculations on atomic layer etching of alumina. So we wanted to study crystalline hafnium zirconia to predict at what temperature. This is sorry, this thermal atomic layer etching will take place
26:20
and look at the chemistry and can you calculate a theoretical etch rate and then look at amorphous materials because they're much more interesting, but much more difficult to do. So you can compare two processes very briefly here. You can have a spontaneous etch in which the material is just etched over and over and over until you hit some sort of stop point.
26:40
And that's what you normally do in most semiconductor processes using mixtures of hydrochloric acid and hydrogen peroxide or nitric acid, et cetera. Really nasty stuff here. We're interested in is the self-limiting where you in each cycle, you only eliminate a small amount of material.
27:00
So we have these models of hafnium and zirconia. This is looking down on the top surface, so you can see the red oxygens. And we do a model. So we look at two models here. This is the spontaneous etching, where in essence, you're going to release the metal fluoride and water as gases.
27:22
This material and then the self-limiting converts the metal oxide surface into some sort of fluorinated surface. So we built these models here. The fluorines are now these blue kind of balls. And you can see this is fully fluorinated. This is partially fluorinated.
27:41
And this is another mix of fluorinated and oxygenated because we don't know for sure which composition you might have. So we wanted to see how important was the ratio of oxygen to fluorine in terms of the chemistry. So we can calculate the delta e's. We can also do those at 500 Kelvin
28:02
for the formation of these different products. So here and here, these are the fully etched, the spontaneous etching where you just form the gas phase products. And then these are the different partially fluorinated surfaces. That don't fully etch away.
28:21
And if you look at the delta e's and delta g's, you'll see that generally these partially fluorinated surfaces are always more stable thermodynamically than the fully etching. So we essentially convert that into these graphs here. For the hafnia and zirconia,
28:41
and I draw your attention to the red and blue lines. So as you walk along the x axis, you're increasing the temperature. And the blue line is the spontaneous etching where you just eat away all the material. The red line is our self-limiting where the surface layer is converted. So as you walk along the temperature axis here,
29:03
the red line is always lower in energy. So that's preferred. The self-limiting is preferred. And at about 650 Kelvin, they cross. And above that, the spontaneous etch is preferred. So you would say for hafnia, as long as you keep the temperature below say 600 Kelvin, you will get the self-limiting controlled etch.
29:24
For zirconia, we can get a similar temperature, which is about 500 Kelvin. So keeping below those temperatures allows you to control the etch. The experimental data says that for zirconia, about 470 and for hafnia about 570 Kelvin
29:42
gives you self-limiting controlled etching. So that's OK. So we can use then this simple idea to predict for a given surface and potential reactant where you'd have self-limiting or simple etching taking place, or even if it's possible.
30:02
So if you swap HF for HCL, which we thought might also be useful. It turns out that these two lines basically all sit above the zero. So it's not favorable to etch or convert with HCL. And that ties in with experiment where nothing happens. So HF is still the best.
30:20
But you can use this analysis, which is relatively quick, to assess how a given reactant might drive the etching chemistry. We can then briefly explore the stability as a function of the coverage. And basically, if everything if all the HF's adsorbed and converted
30:41
the HF02 to HF4 on the surface, you'd follow the dashed line. But as you see, at some point, it starts to level off for both of these species. So you never convert all the oxygen. But we can get a maximum coverage and we can work out the amount of material removed in each cycle
31:01
or the thickness removed. Apologies. So about 0.6 angstroms removed per cycle on both the experimental data. The hafnia is pretty good. Coincidence, most likely. And the zirconia is also in the right place. So if this was, say, minus two angstroms per cycle, we'd be way off. So the difference between this and our 0.6 is OK.
31:24
So that's all right. And then we can we can follow how the reaction proceeds, calculate the interaction energies and what species form, et cetera, et cetera. And then if we move on to amorphous, briefly amorphous materials, right? They've no long range order, no crystallinity, hard to do in simulations. But we managed to build some structures.
31:42
This is your crystalline surface. Here's an amorphous surface that we made essentially heat the thing up to high temperature and then quickly quench it so that you trap it in this amorphous structure. The typical characterization we do, like radial distribution functions, coordination numbers are all consistent with us being amorphous.
32:00
And you then have a lower coordination numbers of hafnium, which seems obvious, but it is actually quite important. And then actually experimentally, our colleagues, Steve George's group in the US have compared crystalline and amorphous materials with a range of etch chemistries. And what you find is the amorphous material always has always has a higher etch rate.
32:22
So 0.36, 0.02 here, here, no etch at all. And then half an angstrom per cycle. So amorphous materials always etch faster. And what we found was if you calculate our thermodynamic graphs here for crystalline and amorphous, the temperature up to which self-limiting amorphous is more favorable is much higher.
32:43
And the overall interaction of HF at this amorphous surface is more favorable than it is on the crystalline. So the amorphous material has a more favorable etch. So the last bit of chemistry then is looking at we call hybrid polymer oxide or organic inorganic materials.
33:03
Example here is the benchmark value cone. So it has an aluminium oxide like layer here, and then it has this organic bit. And the nice thing about these hybrid materials is you can combine the advantages of the soft, squishy organic chemistry and where we know how to make interesting molecules
33:21
with targeted properties, but they don't like high temperature. So if you put them into semiconductor processing, they melt and you your machine is down for six months to be cleaned by putting the inorganic bit and you stabilize the organic bit. And now you can be processed. So that's useful. But you can also then take something like this and you can burn off
33:42
the organics to make a porous metal oxide film, just great for gas storage or catalysis. So these are really interesting. This sorry. Yeah, nearly there. This is the benchmark glycerol and ethylene glycol. Difference is glycerol has two OHs, these red balls.
34:00
So I think glycol is two OHs, glycerol has three of them. And what I do is just quickly show you glycerol grow thicker films. Didn't know why from the experiment. So we calculated just to briefly explain. I'll bring up the numbers right. So you can stand them up right like this. So one of the OHs is used to bind aluminium.
34:23
The other one's available. Or it can lie flat or both of the oxygens in the hydroxyl are used to bind. That's in the ethylene glycol. In the glycerol, you have one, two, three oxygens. Here's the upright and you still have three. But now this hydroxyl is available.
34:40
And what we find is the the line flat or double reaction structure is more stable than the upright by this amount here. So this is the extra gain and stability in the red when you go from the upright to the flat. So the flat's more stable. But what it turns out that the thickness
35:03
when you have glycerol, even when it's flat, is is larger than with the ethylene glycol. The hydroxyl here is more reactive to the next aluminium. The aluminium can bind to this ethylene glycol, but it tends to bind down here so it doesn't protrude as much. And therefore, the film growth is is not as thick.
35:24
So so we explained that the last thing we did here briefly was we looked at alternatives. So we have aromatic molecules which are rigid. And we have different ways to bind them to oxygen or nitrogen. And what you do is is if you calculate the flat
35:40
compared to the upright, the flat is actually less stable. So these rigid organic aromatics will stay upright, giving you nice thick films and experimental growth per cycle is equal to the length of the precursor here. So you're adding it in an upright configuration. So these are good options. So you end up with these kinds of experiments here
36:00
with the ethylene glycol glycerol, growth rate will eventually fall off. But with the aromatic, you can get a pretty good increase and then it kind of plateaus away, but they work well. You can functionalize the aromatic bit, and it doesn't change the chemistry of the deposition, allowing you a huge space to tune.
36:21
OK, so if I can have one more minute, I'd just like to briefly talk about where we're trying to go and in terms of beyond these first principles, DFT. So our colleagues have used kinetic Monte Carlo. This is for zinc oxide growth. And what that does is you generate this lattice here. So this is the surface layer of zinc oxide with various species on it.
36:42
You add the zinc oxide precursor and it sticks here and you can build up then different models of of these precursors on the surface. And then what will happen is a proton will migrate from here to here, and then it will eliminate one of the ligands X, giving you XH going away.
37:05
So those are all driven by the various barriers that are in place. So, for example, the proton migration across the surface has a barrier in this range, depending on what sites are involved. So you calculate those from the DFT and you put them into the lattice KMC and you can calculate the growth of a film, blah, blah.
37:22
You can so you can then actually run the thing dynamically at various temperatures and look at the change in mass as you go through each cycle and the higher the temperature you go, the higher the change in mass, as you would expect. So we can predict change in mass per cycle and we can predict a growth per cycle number as well.
37:44
That's the experiment in gray. That's the calculations. And you can see there's a bit of a deviation understanding that deviation is important. But there's things missing in these models that will be there in the real experiment. So to accelerate this, basically, we're interested
38:00
in developing machine learning models to get those activation barriers. An activation barrier calculation is weeks and weeks and weeks of, you know, thousands upon thousands of hours CPU time. And it might not give you a useful answer. The calculation can explode, et cetera. One of the problems is we're lacking widely available information compared to bulk materials. So the materials project is full of bulk materials, all kinds of them.
38:24
Brilliant. But there's no real comparison for surface chemistries. So that's a big challenge. And some people are trying to do this, which is nice. So you can use machine learning to predict the this nucleation of of a material
38:41
and by comparing this to experimental microscopy data to look at the size of islands involved, you can actually extract then what the diffusion barrier of species is and what the binding energy is. And so that's an interesting approach, but
39:02
it requires the experimental data to to get these images. So we don't want that. And the closest we have is doing gas phase organic reactions where you can predict activation energies for some of these SN2 like processes here. But still, that's a very specific set. So we'd hope that this would allow us
39:24
to find things that aren't considered in the DFT because we can only do a finite number of simulations, but we may find unrealized reactions that can help. The other difficulty is just how accurate is the activation barrier? Because the rate constant in the KMC is exponentially dependent on the activation barrier. So how sensitive are the macroscopic quantities
39:42
you predict to the errors in the activation energy? So lots of interesting questions and challenges for us to realize to make calculation of surface chemistry into something that a process engineer can actually use. So that should be it. Yeah. So thanks a million for your time and patience. And it's wonderful to have had this opportunity to present to everybody.
40:03
Could you remind me if you have any quick questions? This is very nice work. Regarding the construction of the amorphous material.
40:25
And this was done for materials up to which number of atoms? We have let me just go back. Yeah. So here this is the model. So we have we had a 72 atom bulk, which we did in lamps with molecular dynamics to generate a number of samples.
40:43
And then we expanded that. So this is 288 atoms. And then we we put that into the DFT cleaver surface and relax that and ran a couple of examples of those to have a small ensemble. But we found no significant deviations between the
41:02
the the structures that come out of each of those. So this is just an example structure that you get out. So typical heat and crunch, it's big enough to give data like rate of distribution functions and coordination number distributions that are consistent with experiment,
41:22
because the problem here is you've got to think, what's the smallest model that will still give me the correct description? Like we've done 4000 atom models here, but they take six months to run. So it's not very practical. So about 288 atoms here, and it's it seems reasonable.
41:43
Thank you. Yes, I'm sorry. Thank you, Michael. Very interesting presentation. I have more of a technical question for the chemistry on the surface in the presence of plasma.
42:00
What sort of techniques do you use? I mean, is it QMML? Is it like some special force? No, it's it's we've done this for a few things now, and it looks quite reasonable. So it's basically just sorry. I always need my images, right? So you it's just a vast calculation, periodic supercell. Yeah.
42:21
And you basically you have your cobalt surface. And then we just randomly distribute, say this example had, I think, nine hydrogen atoms. I can't do charged ions, obviously, because it's periodic. Right. So we we we but we know from our experimental friends that these are the dominant species present.
42:41
So that accounts for most of the interesting chemistry. And so you just put these in randomly at some distance above the surface. It's usually about 10 angstrom. So that's kind of not biased by being near a particular site. And then you you you run an ab initio MD at 600 Kelvin. Yeah. And you can obviously set velocities maybe to point down, right,
43:03
just to give it some energy. And you run that for, you know, 2200 femtoseconds, because otherwise you'd never get anything done. But we'd really like to see, you know, how you can build on this to get closer to the actual species that are present. So there's nothing unusual in doing this calculation, actually. Yeah.
43:26
No, it's yeah, because we're making and breaking bonds here. Right. But if you could fit in ML potential, you could you could do interesting things. And I'm not so confident. My question is more about that.
43:40
Are you trying to do quite some timing because of the hydrogen atoms? Yeah, yeah. So. Yes, no, it is, I agree. So it's not in VASP at the moment, at least in the version that we're using. But I've seen like Michaelides has done a huge amount of work
44:02
on this for hydrogen on surfaces, and it's really important for particularly weak interactions. So I will argue that, well, we actually have quite a there's a covalent bond forming here. So once the hydrogen gets near an atom and binds, it's going to stay there. So for now, that's that's enough.
44:22
What I think would be really interesting, particularly if the hydrogen is going to maybe hop around on the surface or for these just very quickly, it could be important this is where animations are a problem, right? Just for these KMC calculations.
44:41
Right. Right. Where you're modeling, you know, this hydroxyl, right? This so we've put down a certain amount of these zinc precursors, but they haven't completely covered the surface. So in your zero Kelvin DFT calculation or even your ab initio MD, that's likely to just be stuck there. Right.
45:01
Because it's a rare event in an MD. But obviously, the quantum hydrogen could be quite different. So I think it would be really interesting then to explore that for this kind of surface chemistry here. But at this point, we're not there. But, yeah, no. Great, great point.
45:21
Thank you. Great.