Ab initio modelling of liquids with ionic compounds
This is a modal window.
Das Video konnte nicht geladen werden, da entweder ein Server- oder Netzwerkfehler auftrat oder das Format nicht unterstützt wird.
Formale Metadaten
| Titel |
| |
| Serientitel | ||
| Anzahl der Teile | 23 | |
| Autor | ||
| Lizenz | CC-Namensnennung 3.0 Deutschland: Sie dürfen das Werk bzw. den Inhalt zu jedem legalen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen. | |
| Identifikatoren | 10.5446/61422 (DOI) | |
| Herausgeber | ||
| Erscheinungsjahr | ||
| Sprache |
Inhaltliche Metadaten
| Fachgebiet | ||
| Genre | ||
| Abstract |
|
Leibniz MMS Days 202310 / 23
00:00
ComputersimulationFlüssiger ZustandFlächentheorieBeobachtungsstudieFokalpunktAtomarität <Informatik>DruckverlaufDampfSchmelze <Betrieb>PhasenumwandlungIdeal <Mathematik>Betrag <Mathematik>BruchrechnungKoeffizientEindeutigkeitLogiksyntheseIndexberechnungÜberlagerung <Mathematik>EbeneGreen-FunktionUmsetzung <Informatik>Produkt <Mathematik>BildschirmmaskeEmulationSimulationNabel <Mathematik>PolareTheoretische PhysikPhysikalische TheoriePhysikalisches SystemFunktion <Mathematik>TeilbarkeitScherbeanspruchungWärmeleitfähigkeitDifferenteBeobachtungsstudieComputersimulationOffene MengeKategorie <Mathematik>SimulationDampfdruckAutomatische IndexierungUniformer RaumEindeutigkeitFlächentheorieZentrische StreckungIdeal <Mathematik>Gleitendes MittelKoeffizientEinbettung <Mathematik>Zusammengesetzte VerteilungFunktionalMaterialisation <Physik>SelbstrepräsentationDiffusionskoeffizientVektorpotenzialAutomatische HandlungsplanungDatenbankUmsetzung <Informatik>OrdinalzahlForcingDatenfeldPhysikalische TheorieSoftwareentwicklerNichtlineares GleichungssystemSoftwareQuaderDatenstrukturLogiksyntheseBildschirmmaskePhasenumwandlungZahlenbereichWahrscheinlichkeitsverteilungSoundverarbeitungPunktKorrelationsfunktionMereologieResultantet-TestEnthalpieStandardabweichungCASE <Informatik>Algorithmische LerntheorieMixed RealityGruppenoperationFlüssiger ZustandWärmeleitfähigkeitAtomarität <Informatik>StrömungsrichtungGreen-FunktionFokalpunktPartikelsystemParametersystemInteraktives FernsehenIntelligentes NetzTermBruchrechnungMetropolitan area networkBus <Informatik>Mechanismus-Design-TheorieBildverstehenRohdatenBildgebendes VerfahrenDruckverlaufARM <Computerarchitektur>TransmissionskoeffizientSystemaufrufEinsComputersicherheitDiskrete-Elemente-MethodeWeb SiteGüte der AnpassungAggregatzustandComputeranimation
06:58
SimulationLineare RegressionDatenbankKraftOrdinalzahlMaschinelles LernenSelbstrepräsentationProgrammierumgebungBewegungsunschärfeEigenwertproblemVektorraumDatenstrukturGraphProdukt <Mathematik>Message-PassingComputersimulationVektorpotenzialNeuronales NetzKartesisches ProduktLineare AbbildungBasisfunktionAbstandPolynomFaltungsoperatorWahrscheinlichkeitsverteilungMatrizenrechnungRelation <Informatik>ZahlenbereichOpen SourceWellenfunktionInvarianteEnergiedichteMaßstabPhysikalisches SystemFunktion <Mathematik>StellenringPermutationKreisbewegungTranslation <Mathematik>ExtrapolationPhysikalische TheorieTrajektorie <Kinematik>RechenbuchSoftwareVektorpotenzialDatenstrukturMAPMatrizenrechnungAnfangswertproblemMereologieCASE <Informatik>AbstandSelbstrepräsentationTermHybridrechnerMathematikNotepad-ComputerWasserdampftafelFunktionalSoundverarbeitungSchnitt <Mathematik>Elektronisches ForumOrdinalzahlKugelPhysikalisches SystemMetrisches SystemKategorie <Mathematik>VektorraumEnergiedichteForcingOrtsoperatorSystemprogrammierungMessage-PassingProgrammierumgebungSimulationProgrammierungOrdnung <Mathematik>Minkowski-MetrikRechter WinkelZahlenbereichStellenringFlächeninhaltGrößenordnungKonstruktor <Informatik>MultiplikationsoperatorEinsBinärdatenÄhnlichkeitsgeometrieHalbleiterspeicherStammdatenTranslation <Mathematik>RechenschieberFlächentheorieLineare RegressionPotenzielle EnergieFitnessfunktionZentrische StreckungComputersimulationPhysikalische TheorieTrajektorie <Kinematik>FaltungsoperatorDatenfeldStetige FunktionPermutationKreisbewegungInverser LimesGrenzschichtablösungParametersystemNeuroinformatikExtrapolationNeuronales NetzOpen SourceInvarianteComputeranimation
13:51
ComputersimulationPhysikalische TheorieEnergiedichteKraftPhysikalisches SystemDatenstrukturDichtefunktionalTrajektorie <Kinematik>ExtrapolationSimulationRechenbuchZahlenbereichOrdinalzahlTotal <Mathematik>InformationSystemprogrammierungDickeGraphikprozessorPaarvergleichFunktion <Mathematik>WahrscheinlichkeitsverteilungDynamisches SystemKoeffizientQuadratzahlDisplacement MappingLineare AbbildungLineare RegressionAutokorrelationsfunktionWärmeleitfähigkeitStrom <Mathematik>Fluss <Mathematik>Meta-TagFlüssiger ZustandMaschinelles LernenMaßstabLOLA <Programm>AnalysisSkalierbarkeitRechnernetzNeuronales NetzFokalpunktMereologieMultiplikationsoperatorGravitationsgesetzDatenstrukturSimulationEnergiedichteDiffusionskoeffizientLineare RegressionTeraelektronenvoltbereichZustandssummeDickeData MiningEinfache GenauigkeitParametrische ErregungCASE <Informatik>ÄhnlichkeitsgeometriePunktPhysikalisches SystemOrdinalzahlComputersimulationWellenpaketStandardabweichungFlüssiger ZustandZahlenbereichKategorie <Mathematik>FeuchteleitungZusammengesetzte VerteilungLaufzeitfehlerTrajektorie <Kinematik>RechenbuchDichtefunktionalTotal <Mathematik>ResultanteStreuungSkalarproduktTouchscreenForcingVektorpotenzialFunktion <Mathematik>DifferenteWärmeleitfähigkeitDisplacement MappingSichtenkonzeptFehlermeldungDatenfeldKorrelationsfunktionProzess <Informatik>Komplexes SystemSystemprogrammierungDatenanalyseSoftwareQuadratzahlArithmetisches MittelAlgorithmische LerntheorieInformationBeobachtungsstudieAggregatzustandGrenzschichtablösungLeistung <Physik>MaschinenschreibenIndexberechnungZweiRegulator <Mathematik>MomentenproblemSchlussregelMenütechnikNeuronales NetzInterface <Schaltung>Demoszene <Programmierung>QuellcodeRelativitätstheorieEinsWeb SiteBasisfunktionMixed RealityCAN-BusGewicht <Ausgleichsrechnung>Objekt <Kategorie>Computeranimation
Transkript: Englisch(automatisch erzeugt)
00:00
I want to talk about some modeling studies we have done recently and we investigate different kinds of models to model overall complex liquids and at the end we observe that this machine learning approach seems to be most efficient and outcomes with other approaches what we
00:24
see. The focus of our studies are de-boitactic solvents and what are these compounds and in principle these are mixtures of two compounds for example a cheap organic salt like choline chloride which is produced on a megaton scale and this you can mix with some organic compounds
00:42
like urea, glycol, melonic acid or other organic compounds and then you get an overall cheap mixture which is liquid at room temperature and the advantage of these liquids is they are overall cheap and they are eco-friendly and they have a low vapour pressure.
01:01
And what's now a deep eutectic solvents and there's a current debate but in principle you have a decreased matting point compared to ideal behaviour and in principle what you see when you have this equation here you have the, I think is this something like a pointer,
01:23
okay but what you see in principle you can calculate the properties of the pure compound for example the melting enthalpy is known from the literature and then what is measured in an experiment and when you use this equation then you can calculate the mole fraction of
01:42
the compound and when you see this activity coefficient is wrong then you have an ideal behaviour and when it's higher or lower then you have non-ideal behaviour and in what you see in principle the most important part is the first part of this equation, where you have the part with the melting enthalpy and the second part is not overall so important
02:04
for non-ideal behaviour but when you see now you have this ideal behaviour and you have a melting point with signifying below then you have to say you have a deep eutectic solvent. Okay where you can apply these solvents for example you can make a unique gold nanoparticles you see this nice star there and these are made of gold and this whole surface is covered
02:24
by high index planes and this means you have gold atoms with a low number of neighbours and therefore you have very reactive gold nanoparticles. Then you can use it also as a solvent and catalyst where the solvent forms in a separate phase which is also nice for
02:42
synthesis you can use it to generate highly stretchable and non-viralitetile ionic gales as you should see here in an example and what you can also do you can use it also for a frontal polymerization and where you obtain at low temperature a full conversation and
03:00
then you can use it also to get thick coatings on materials which you can also use subsequently for laser structuring and other properties. And what you can learn from molecular dynamics simulations when you study these liquids, this is from a first study where you have used global scale charges and what you see here red are the experimental activity coefficients
03:26
and on the x axis you see the number of hydrogen bond donor atoms which are close to the anion and you see a good correlation of the anion embedding in the hydrogen bond neighbours and there you can say okay the non-ideal mixing behaviour is correlated to the embedding
03:43
of the anion into the hydrogen bond network and you can also then use the Kirkwood-Wolfe theory and you can also use the Kirkwood-Wolfe theory to calculate this from the molecular
04:05
dynamics simulation. So in principle here you have the radial pair distribution function and this you can use to get when you integrate it to get this value and with this you can get this relative of this activity coefficient with respect to the particle density and then
04:23
you get also see it, you see the good correlation compared to the embedding of the anion. So you can use the structure to attribute some properties on the molecular scale to some
04:41
experimental properties. And what the problem was of this previous force field, this was made by global scale charges so you need also some, it's not like a black box approach in principle to be on the safe side, you always need a first principle molecular reference to check that the structure
05:02
is good and not too bad at the end because you have these ionic compounds and there the structure might not be perfect and therefore we thought okay the best would be to develop polarizable force field because this considers some effects and this was done by Omer Chai Estebo, my PhD student and we started with the CHAR model and we tried to improve it
05:24
but what we observe that we do not get any good results if we do not add an additional parameter in our force fields which decreases the interaction of the anion with this buteroxyl group of the cation.
05:42
And for this we need again a first principle molecular reference and when we have done this then you see we have our previous force field, this is this light colour in green and light colour in red and after we have added this additional damping function then we see we have a good agreement between our first principle molecular reference and our
06:00
force field. So and now we can check for example experimental properties, for example we have the diffusion coefficient of the cation of urea, then we have the conductivity and also we can calculate the viscosity and what we see we are also good, very good compared to the experimental reference.
06:21
So this means as soon as we get a structure of the force field which is close to the first principle molecular reference we see we can also reproduce very well the experimental properties. However, the major drawback is we need always a first principle molecular reference method, we try to use semi-empirical approaches and in this case we observe overall a very strong
06:41
deviation but here so such a good agreement as you see here is not possible with semi-empirical approaches we tested several and therefore we thought okay maybe it might be better to use machine learning in the atomic potentials. So the idea is the end principle you construct a reference database then you use some representation
07:05
for the atomic environment which we see later what you can use and then you can make a regression fit and then for the potential energy surface and then you can use this to run the simulations and this can be faster. And the challenge for us was when you look at literature then mostly you have over simple
07:23
systems like water and other systems but we change now to systems with ionic molecule compounds so we needed of course much more reference data what we expected then in case of water and this is therefore our question was is it still good to use these machine
07:41
learning based potentials to model our systems. Okay and at first we tested the approach of Christoph Schütt from Berlin and he used message passing a narrow network potentials and in principle you have following you have
08:02
there an atom and this atom you can represent with a vector these are some properties and then when you pipe this proper this vector or this into a more network then you get an
08:21
energy for example for this atom. However you have also the neighbor atoms now and then you can say okay you have your potential which affects these vector properties depending on the position in space. So and in case of this approach of Christoph Schütt he used a continuous filter convolutional
08:40
layers to update this so in principle what he did take he just take distance between the atoms then expand this by an exponential function so that he generates for many value about 300 and then use a convolutional network to update this vector of the single atom and
09:00
then you can based on the values then you can predict the energy of the atoms. However what we see for our systems there was no significant benefit to use this approach because our system needed overall very large systems and for this we used the program schnetpack.
09:23
So therefore we check the second possibility this is our descriptor based neural network possessions and this is just you construct a matrix or an area of numbers to represent the environment of an atom we see on the next slide here we used an open source package
09:41
deep MD and what we see this is fast and memory efficient compared to schnetpack and is about two orders of magnitude faster. So in principle how it works so you have your atom and up to a certain yeah cut off distance all atoms are considered and based on the yeah as you see here it's with scale linear
10:05
resize so if you have changed twice the number of atoms then you need the computational time request twice the times so it's skates very perfectly and at the end you add up the energy of each atom so and now how this is constructed you construct a local environment matrix so
10:25
you'd see you just like in the other pose of crystal should you just take the distances between the atoms then you use additional awaiting function which assures that this function which you use for the neural network potential decreases to zero so you have the
10:44
initial part which is the inner region then you have this which smooths it to zero and then up outside the cut off all the values are in principle zero and when you use this then you see it's invariant by rotation and translation but not by permutation and therefore
11:02
they use a matrix construction so this is our principle our first matrix where you have here these s value then here this is in principle the x value so this is just the distance on the x axis divided by this r and and also you have the same for y y and z and and yeah
11:28
and then you use a neural network to generate from this single value about 100 values which are this row and and you do it also for the other things and then you get this large matrix
11:44
and then you use the full matrix is used for this part of the matrix equation and the first part which is only about four here this is a dimensional four about four and use it for this one and then you multiply this then you get at the end the matrix which is invariant
12:00
under permutation and this matrix you can transform it into a vector and then use it to fit the the neural network potential okay so and yeah and then we can use this model to to rate it and then it's of course the question how we can collect the reference
12:24
data most efficiently and because what we would not like to do we would not like to run it for spins and molecular dynamic simulation because then would be no benefit and you know of course the challenge is we have a limited extrapolation capabilities of machine learning potentials and the solution is we generate a trajectory at a lower level of trajectories
12:44
like semiperical approaches and we see that the structure is not very good compared to first principle reference but at least it's not completely far off and and then we use selected structures and to calculate with a higher level of theory and then at the next
13:04
step we make adaptive sampling to refine the model and so this means in principle how it works we run a simulation with our generated force field or the machine learning based force field and other several steps let's say about 1,000 steps we use for other models
13:26
to calculate also the energy for this structure and these other models they are in principle not really are some models they are just the same model but they have used different initial values to optimize the parameters of neural network potentials so and when you see that
13:44
these models matches always like looks here on in the button then it's fine and when you see you are in a region where they differ then you know okay you must take some additional structures to train the model and and then we run these simulations and some numbers
14:03
so we need and initially so about 80,000 structures from semi empirical calculations which we were for which we collect the energy and forces for system of about 305 atoms so it's that's not really a big problem is Cp2k and then we need about 10,000 additional structures
14:22
by sampling to refine the model so in total we need about a 30,000 DFT calculation which an overall small system and for example for 100 picoseconds you need about 200,000 calculations so it's significant more efficient than these first principle dynamic simulation and you
14:42
can also run all this calculation principle in parallel so now then you can check the accuracy and of the predicted energies and forces and on the left side there you see the average energy is set to zero and you see that we have over a good agreement and
15:01
the mean average error is about 4.5.48 milli electron volt so it's over small deviation and on the left side you see the forces and you see we have much more values because the forces can be carried for all atoms and they also used to optimize the force field and you see we have a good agreement however you must be careful because when you run these
15:25
simulations you cannot be sure that you are on the same yeah that you all the simulation is too stable because it might be that the simulation runs into a structure which was not included into part into your parametrization and then you will get off of the trajectory
15:44
because these errors accumulate and therefore you must be always to be on the safe side you must check other several steps always with these other models that the energy is still good and then you are on the safe side but when you have seen this whole simulation
16:02
this is the case then you can use it to make this kind of structure and now we can for example compare the structure and here yeah yeah we see what we can do so if you can now run simulation is 2500 atoms about and the length of our simulation is one nanosecond
16:22
so it's two million steps and we need only about two hours with only 10 hours with two GBU so we can obtain trajectories of 2.5 nanoseconds in a day and now we can compare this of course to our reference structure is used for training and you see it fits perfectly so you see a
16:43
single trajectory is sufficient for reliable in this case of structure but now when you want to go to dynamical properties like diffusion coefficients and then you see you need more simulations with diffusion coefficient can be calculated by the mean square displacement
17:03
which is here you make a linear regression and then you can data mine diffusion coefficient and when you take only one value then you see it this value slightly scatters you see here the screen dots which show how this scatter and the red values how this scatter and black and but when you take about five values then you are on the safe side and you can then
17:21
really compare these simulations and this is of course then yeah you will not run so many for spins in molecular demystulation because they are too expensive but with these approaches we use it's not a problem because you have this in in a few hours. Then you can compare this to experiment and what we see diffusion coefficients are smaller
17:45
but the activation energy is the same however what you what we have here is only one experimental value for example when we look for a NASA value there can be measured more easily like the conductivity then you see here three different kinds of experimental values in the literature
18:03
and our simulations are in principle between so you see that we get also results close to experimental values and then we can also use them for example to data mine activation energy for this process and others. Then of course you can also get more detailed views on structures for example like hydrogen
18:22
bond dynamics where we can use all the correlation functions to data mine the lifetime of the hydrogen bonds and then we can also do this at different temperatures and based on Arrhenius behavior we can then data mine also the activation energy for those processes and in this case
18:41
you see for example that the anion and the oxygen atom of Aurea they have a similar activation barrier at these hydrogen atoms so they are very similar and therefore you see that's also this good embedding and mixing and in case of the the oxygen here is see this has a very loosely bond to this weaker bond and therefore it's moved more off-center more
19:05
likely off-center and it prefers to interact with this nitrogen atom of the cation. Okay so in summary we have seen that machine learning potentials, atomic potentials allow more efficient and available investigation of liquids than a single first principle molecular
19:25
dynamic simulation run we see that or we know that the training of a complex system like these mixtures of ionic and molecular systems need about two or three weeks including 30,000 single point signal translocation of small system for training and then we can run a
19:42
large number of simulations and what's also known that the simulation runtime skates linear with the number of atoms and in the next step we want to study the hydrogen diffusion amorphous gas permeation barriers with metadynamics simulation. Okay I thank Umed, he has done these studies with the machine learning potentials, my cooperation
20:03
partners LUG for funding and et cetera Tresen for the computational time and I would like also to invite you to Leibniz MMS Winter School which is in the middle of October and this will be a joint summer school from the MMS network with the Center of Scalable Data Analytics
20:22
and Artificial Intelligence in Dresden Leipzig, it will be at the IOM in Leipzig and it will have two parts, the first part will be more for people who start with machine learning and it's about data processing, data analysis and machine learning and the second part will be more for more advanced users and more information on the summer school will follow soon and
20:44
yeah thank you for your attention.