Time integration methods for finite element discretizations in weather forecasting

Video thumbnail (Frame 0) Video thumbnail (Frame 488) Video thumbnail (Frame 1005) Video thumbnail (Frame 1529) Video thumbnail (Frame 6119) Video thumbnail (Frame 9025) Video thumbnail (Frame 16604) Video thumbnail (Frame 17542) Video thumbnail (Frame 18015) Video thumbnail (Frame 27678) Video thumbnail (Frame 29415) Video thumbnail (Frame 30163) Video thumbnail (Frame 31043) Video thumbnail (Frame 32748)
Video in TIB AV-Portal: Time integration methods for finite element discretizations in weather forecasting

Formal Metadata

Title
Time integration methods for finite element discretizations in weather forecasting
Title of Series
Author
License
CC Attribution 3.0 Germany:
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
Publisher
Release Date
2019
Language
English

Content Metadata

Subject Area
Abstract
There is a new interest in finite element methods for solving the equations in numerical weather forecasting. In contrast to finite difference and finite volume methods explicit time integration methods are hampered by non-diagonal mass matrices in front of the derivatives. We will compare different mixed finite and discontinuous Galerkin methods for the two-dimensional linear Boussinesq approximation in the context of split-explicit time integration schemes. Especially different lumping procedures are investigated which replaces non-diagonal mass matrices by simple diagonal block-diagonal matrices. These methods are compared with energy conserving implicit Runge-Kutta methods for a non-hydrostatic gravity wave example.
Goodness of fit Greatest element Decision theory Moment (mathematics) Element (mathematics)
Prediction Exponentialgleichung Multiplication sign Disintegration Student's t-test Collision Infinity Element (mathematics) Finite element method Mach's principle
Standard deviation Presentation of a group State of matter Decision theory Multiplication sign Direction (geometry) Numerical analysis Mass Icosahedron Sphere Infinity Fraktalgeometrie Duality (mathematics) Goodness of fit Mathematics Calculus of variations Equations of motion Different (Kate Ryan album) Gitterverfeinerung Modulform Glattheit <Mathematik> Position operator Stability theory Physical system Area Predictability Moment (mathematics) Sampling (statistics) Affine space Element (mathematics) Numerical analysis Mathematics Prediction Volume Finite element method Spacetime
INTEGRAL Multiplication sign Outlier Direction (geometry) Mereology Mathematics Coefficient of determination Tensor Velocity Pressure Physical system Area Polynomial Logical constant Complex (psychology) Element (mathematics) Product (business) Category of being Velocity Frequency Spacetime Metre Functional (mathematics) Connectivity (graph theory) Student's t-test Rule of inference Continuous function Tensorprodukt Element (mathematics) Independent set (graph theory) Spacetime Nichtlineares Gleichungssystem Distribution (mathematics) Scaling (geometry) Linear equation Military base Projective plane Basis <Mathematik> Line (geometry) Continuous function Cartesian coordinate system Affine space Exponentialgleichung Function (mathematics) Autoregressive conditional heteroskedasticity Einbettung <Mathematik> Vertical direction Pressure Finite element method
Point (geometry) Scaling (geometry) Exponentialgleichung Different (Kate Ryan album) Multiplication sign Model theory Differential equation Nichtlineares Gleichungssystem Mereology Pressure Physical system
Point (geometry) Functional (mathematics) Momentum Diagonal Multiplication sign Mass Inverse element Infinity Sparse matrix Mereology Element (mathematics) Mathematics Independent set (graph theory) Many-sorted logic Term (mathematics) Operator (mathematics) Exponentialgleichung Matrix (mathematics) Physical system Stability theory Differential (mechanical device) Eigenvalues and eigenvectors Moment (mathematics) Physical law Model theory Computability Mathematical analysis Algebraic structure Inverse element Maxima and minima Sparse matrix Voting Exponentialgleichung Nichtlineares Gleichungssystem Right angle Iteration Diagonal Block (periodic table) Metric system Mathematician Pressure Matrix (mathematics) Directed graph Finite element method
Group action Direction (geometry) Multiplication sign Decision theory Price index Time domain Cartesian product 2 (number) Time domain Mach's principle Wave Velocity Gravitation Set theory Spacetime Stability theory
Point (geometry) Wave Term (mathematics) Green's function Matrix (mathematics) Mach's principle
Supremum Different (Kate Ryan album) Forcing (mathematics) Inverse element Nichtlineares Gleichungssystem Term (mathematics) Sparse matrix Finite element method
my philosophy there attempts to go from find bottoms to a planet element this good decision because it's at the moment and that the topic and
my last a at the at all research institute I want a moment a bit about that might
be to do them have something also label and that the last vestige together with the bachelor student and chefs these it's a bachelor student worried that has only 1st steps and I want to explain a little bit what we have implemented and what are our experience though I want to give a short
introduction the my what I loved about the finite elements will be done at the end that we only applied to to a linear problem and then my especially in the less time equation methods and I want to explain what other problems with this time the collision muscles if we all Kong-funded
finite elements of this to decisions some American sample and that's it so at the moment the
Hatfield listed some idea of what happens in the moment and I think the mathematics from the former about the prediction of reaches in some sense a really new a font us and then you can look in in different a key different direction in this in this research 1 as someone is that you are going from sound latitude and longitude words to unselected puts OK I'm smarter could see or the has made it was already 20 years ago or something like that or so years ago but I think in the last 10 years together with that we are changing from to fully composable or non had the systems that this unspotted goods also get the new the new interests and we got to do this quits in in different directions so we have different type of threats and we are at the same time we want we want put refinement and they want to have state also on disk with find things on handful different goods we want to have some type of unified mass arts and we want to have a stable methods and methods which a preserve as low as much as possible a form the continuous more and we also go at MIT this quits to more sophisticated space at the position methods so what is it we are looking on the top and for find the differences and on designs fractals which we are going to something which people in other areas called a method method of all combat that of affinity and and us all we are looking for another type of of presentation of the equations of motions like in and or number formulation and going then from this disk from this continuous formulation to a discrete land and from that that we get something about it this position and that of the pellet only go in space we do on the the same thing in time and there are now at moment also appeal not of sings about new retirement at this conversation methods which of follow from this Hamiltonian and formulation and that this year that for instance there is a special workshop about this so called so-called integrate in in the in that at the end of the year so it's also and you see that at all the time I think in in numerical weather prediction there are now people are sitting which to deep deep mathematics in some sense so what
now I them is no at the mathematics though what we've done with those of the bachelor student is that we want to solve this linearized equations there which is a the is a boson next type equation and you have this two-dimensional you have tool to velocity components we have a pressure component and we have a buoyancy component and we have in the in the system as a sweet timescale so it's a advection timescale we have this is a time scale and we have this this Acosta timescale all the time scales which we also have to follow a constant cumbersome and that was a people I think the last of this year in quarterly Olmützer society we had people from the gung-ho project start presented sings about finite elements for this application and we following the bachelor worked through 2 we implement that and then to see how all tied equation as that's that can be useful so you already I a spirited equation in 2 parts so we have this no time scale the think about maybe 20 meter per 2nd and the thing about this fast timescale sound speed of 260 FIL segments and this 1 is somebody in in between and ice but already the equation here in this to rise version in this 2 parts which we also have later on I call this this rule part and this the 1st part the all finite elements so what we've done is we want those off this simple equation honor when the copies
input so we started simple with the student with simple one-dimensional finite elements and all 210 support poked elements and implement that does 41 so and could not so good and key tells them out with a bachelor's then Bikel because you have not that knowledge about all that as I'm usually also that have appeared in 4 K we started with to aetioporphyrin had elements so we was on Monday on the 1 side we the was a continuous forget elements which also advise all all put in on the polymer found to be pleased the office mom and we denote it by policy on and we have this continuous elements which also was pollen functions which you also of we all plus 1 and been the by teach you all and then we take spaces we have to choose taking as basis for the for the good for the too about a component 3 have to take a space for for the buoyancy and they have to take 1 for the pressure and that what people 1st proposed a little bit is dude tested takes sweetie of this fund for the the affinity elements bases have to take which was disk this continuous and area feel about topic elements never take a space which is which is continuous and you have and and on dementia case you have linear elements and then that you have to take for the buoyancy that for the for the velocity and for the pressure you have to take at this continuous space and this 3 spaces then the therefore if it has this property that you can take the great enough that you could coming here and take the dog built and on Monday the wage and again the end up some they're here but now we have to go to Judea and 2 would be we have to choose the buoyancy the well a city and pressure and now we take for the buoyancy be take this this space he as a tensor product but for the for the for the velocity components we have a a tensor product which is linear a full for you in this direction the tagger the take we take continuous space and on the other line we take this continuous space which 1 although less for the it's it's wise for us are inappropriate for P again we take a tensor product of 2 this continuous spaces and and this raises if you take it again we will have this type of embedding bubble so it can starting with the buoyancy in you go compute this type of a call you end up in in this in this neck buffeted amend space and then you make that elections and end up in this space so now that that's a small more no OK good so what we have we have a we make only experiments with the low although I'm that's functions we have piecewise constant and that's function piecewise-linear and that's functions and they have to work was of it the world is german yeah and peace last but not think functions you have to implement at all and so I had we have love admit how to do that in some efficient and that's all too but easy if you go to find other elements and here you see how is like this fine and elements and look like and and and if you know if you want to go to find out and amends to have to go from this to to a weak formulation and we have to choose all the I'll trial functions under the velocity components and the pressure and the B from this and that's French so that I have showed and again be take test functions which are often from from the same and that's functions the multiplied with this test functions make here for the pressure to term some of the Pasha integrations and offer that we are looking for outliers functions infinite element space which of fuel for all test functions this this equation and if you again compute all so I want not to explain in detail how we compute the advection tome especially for the for the pressure or he'll for the it's it's this continuous audio phobia so we have to do something special at the end we then we end
up but the modeling our system so you have a system of all the different differential equations and I have have written again here old got lecture part if you if you cut that this part the new see seawall do equations all each equation you can solve by itself and their feel the pressure Tamiflu the boy tell you it's the acoustic Germany have been African a boy and 2 2 so and we have to solve best that equation in time so now
comes in we my favorite methods which I use in in the final point context that all it experts that matter so for this simple models I want to solve this fully explicit models but with different time steps that because of this different time scales so it's
it's not true so we are going back to
this this system here so I have and have so part and the fastball and we have we have developed is
sometimes of a combined of of a special type of work with them as art which we use for the slow parts of seemed that this fast part this year then what we do is we in each stage we solve again as the French iteration in all in all cases it's all or iterations of now but he this part this is frozen so it's fixed and only this part which is the FastPath is this this really the right hand side function which we can compute in each point and we can solve this in a system with this fixed source film and then at the end the we get a new speech right we can compute again this slow pop can form again this type of work all right inside soft this we know I creation and read feel maybe 3 stages with a fixed time stamp and then we take smaller prime steps to solve this system but the the special muscle and if you are looking in our for this for this fast part here then you you you will see a avid a little bit simplified if you take only momentum and and pressure that this fast part has this symplectic structure so you see if you would it depends on p and p depends on on you and then you can of what's what mathematicians call and symplectic all I think you as a small arrow or so we solve fills the 1st creation with with a full what all along and then use that you should be and plus 1 then dissolved a 2nd iteration but at the Beck model this is this is the stable mass art and there's also something which is this is a centralized symplectic are law which is also called in the you met little the literature something like of of time much as a dumping so we make here again before what all of that and the place you at this this for what us that also with some with some lecture on the peak the the bits the head and to stabilize the whole procedure that is also not fully understand it at the moment so now we are going to this to apply this with explicit methods and we to slide and we have to bring the the year the mess May-Britt switches on front what you have seen it if you had a welcome at the time that elements have to bring this to the right hand side so you'll for each computation of HBO's of terms you have to soften some sense an assistant who can make before that it wouldn't good they composition but if you go to Sweetie'' it's it's a major problem now what people do in the finite element of literature they did do some some slumping the the the data the rule-based this mate by diner got all blocked by the bricks which are called though that you have now the filled options you'll you replace pose of them but this the lump matrix or the other 1 you only use a because that I compute on the on the large times that he I take the exact messmate matrix that I do on a small that so I take you didn't metrics rare have sort of a simple linear system so it's diagonal Proc no problem and SSL a cell the idea so what I do know I sup upright shot something taken in the media I sup like does he did that he'll from a for the from the wonderful for Malaysian and I added again so now this is something with which I called my slow term I have to compute it only when this is my fast left CompuAdd often and we tried that fills and we see OK this was of 10 people OK did a lot with us but experts at methods by this idea doesn't work so that was unstable the watermarked then I make some eigen value analysis and see that that at least if you see that the that this operator has some eigenvalues which are sitting on the not on the left hand side and and this makes the problems and now the idea walls and not to take it this simple last lump breaks but to go to a matrix made say beyond looking for something which isn't in those metrics of of of the mass matrix quite has a has a sparse structure something like an approximate inverse and that we work here that the at this this term here and and again though you don't change the equation the only change what is the slope or the voters the fast and for this involves matrix we tested 2 ideas the first one is we taken and be taken and in in those metrics which is the the minimizer of this of this will bring a small and in Ingalls has cut sparsity peplum so you could put in for the food is is so the so the minimum this Matsuo or the idea is that we compute the inverse and then it came to take only a sparse pattern of this in the matrix A and the sparsity deep pet is you should be the same as as the sparsity pattern of of 10 or something like that so and then this idea we made some experiments so we take the a
copy in domain so 150-kilometre to do to the east and the west and thank can move us in the IoT and sweet thousand segments and we start with the atmosphere requests that ran out the the velocity and this this budget but the business well and buoyancy has has this type of a special spot show and this 1 you put the patient that you have a give us some of some some of some movement that so you're all EU Parliament us and we take a space of this decision so It's index I action and also in this set direction it's it's like the the metal so in the developed a direction it's my course and we take a slow times that of 20 seconds and they have to take something like a 90 and small time steps below the large times that to fulfill this this stability to Terence for this for the and this is the this
is the this is the buoyancy at the at at th the the stopping pardon and then you after the 3 thousand seconds if you also
include some some of the some horizontal movement here you get this type of picture and we are now looking somewhere here in the middle of this point see you in the case where we have no movement to see what happens with this different must lumping
procedures so I have blocked it here now sweet tired of of of lumping procedure so I have 1st the 1 of the the this to this in this blue is we make no lumping what we call the exact solution some sense then be made a full lumping so we have 4 boast terms we have landing matrix and then we make we in there we have only the slumping on on the fast and there said all things will work so and then I'll there should be a little bit
of and now we of going the maybe that's the last 1 now what we now do is that go to this the the spores this this idea that these th sup thanked foster and in the case that we take the indoors which is off boss then you see for us if we take we do it with our this this station we at his some problems and then if you make also this this small southern decision in fast for the Caucasus that belies the sprit experts and then you see that the exact solution and that what we get here all on this user you see no no differences will this force experiments and that was something like outcome of this bechamel rock and what I want to send the and now I am also have already a finite element code into these useful for the fully compostable equations but notice that that so that's what I want to say sentence
and
Feedback