A Jacobian-free Multigrid Preconditioner for Discontinuous Galerkin Methods applied to Numerical Weather Prediction

Philipp Birken, Andreas Dedner, Robert Klöfkorn · 2025

Read A Jacobian-free Multigrid Preconditioner for Discontinuous Galerkin Methods applied to Numerical Weather Prediction by Philipp Birken, Andreas Dedner, Robert Klöfkorn (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-24504.txt.

Introduction

Philipp Birken, Andreas Dedner, and Robert Klöfkorn build a fast solver for weather flow models in a 2025 paper. Their flow method, a high-order grid scheme, suits storm-scale forecasts well. Such flows are stiff, with thin layers and slow sound speed ratios, so each time step needs a hard solve. The solver must be fast, lean on memory, and run well in parallel on big machines.

Their solver wraps a Newton loop that skips the costly matrix build at each step. A multi-scale smoother, grown from a low-order grid tucked inside the fine one, preps each solve. Maps between the grids keep mass true while they move data back and forth. Light time-step smoothers run in parallel with low memory. Tests cover both dry and thick flow cases in two dimensions.

Catalog record

CREATORPhilipp Birken, Andreas Dedner, Robert Klöfkorn
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-24504

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Jacobian-free Multigrid Preconditioner for Discontinuous Galerkin Methods applied to Numerical Weather Prediction Philipp Birken, Andreas Dedner, Robert Klöfkorn Abstract Discontinuous Galerkin (DG) methods are promising high order discretizations for unsteady compressible flows. Here, we focus on Numerical Weather Prediction (NWP). These flows are characterized by a fine resolution in z z -direction and low Mach numbers, making the system stiff. Thus, implicit time integration is required and for this a fast, highly parallel, low-memory iterative solver for the resulting algebraic systems. As a basic framework, we use inexact Jacobian-Free Newton-GMRES with a preconditioner. For low order finite volume discretizations, multigrid methods have been successfully applied to steady and unsteady fluid flows. However, for high order DG methods, such solvers are currently lacking. This motivates our research to construct a Jacobian-free precondtioner for high order DG discretizations. The preconditioner is based on a multigrid method constructed for a low order finite volume discretization defined on a subgrid of the DG mesh. We design a computationally efficient and mass conservative mapping between the grids. As smoothers, explicit Runge-Kutta pseudo time iterations are used, which can be implemented in parallel in a Jacobian-free low-memory manner. We consider DG Methods for the Euler equations and for viscous flow equations in 2D, both with gravity, in a well balanced formulation. Numerical experiments in the software framework DUNE-FEM on atmospheric flow problems show the benefit of this approach. keywords DG, FV, implicit, multigrid, preconditioner, Jacobian-free, numerical weather prediction, DUNE, DUNE-FEM 1 Introduction Discontinuous Galerkin methods has emerged as one of the most promising discretizations to replace second order finite volume methods for simulations of turbulent, time dependent, compressible fluid flow. • It is of high order, a property of critical importance for efficient Large-Eddy-Simulation (LES) of turbulent flows. • It can handle the complex geometries necessary in real world applications. • It has computational advantages on modern computer architectures, and lends itself well to parallelization In this article, we make use of Discontinuous Galerkin methods (DG) on quadrilateral grids. However, the techniques developed here can be used with other DG formulations as well, for example, Discontinuous Spectral Element methods (DG-SEM) where use is made of a tensor product extension from a one dimensional formulation or other general DG methods on simplicial grids. Many applications require implicit time stepping to avoid excessive time step restrictions. This is the case when the mesh size varies considerably within the domain, or when the phenomena of interest to occur over long time scales compared to the fastest dynamics. The former occurs in particular when high aspect ratio grids are employed, as is the case for wall bounded viscous flows, but also for atmospheric flows where the gravitational force requires finer resolution in the vertical direction. The latter is also a problem in numerical weather prediction (NWP). In particular, the fastest processes are sound waves travelling at the speed of sound, whereas the phenomena of interest travel at the flow velocity. For low Mach flows, as is often the case in NWP, this is a serious issue. In this article, we will consider fully implicit time integration, which leads to one large global system of equations. Complementary approaches that we will not follow in this text, are implicit-explicit (IMEX) methods, and horizontal explicit, vertical implicit (HEVI) ones that have seen a lot of application in NWP Bao et al. (2015) ; Giraldo et al. (2024) . These make use of a geometric or term wise splitting of the equations and use different time integration methods for the different parts. The technique suggested here can be employed subsequently in those methods as well. Implicit time stepping requires solving large sparse linear systems of equations, which in turn requires efficient solvers to be competitive computationally. As pointed out in Bastian et al. (2019) , to achieve the peak floating point performance on modern processors, it is necessary to reuse data loaded into the CPU, i.e. the arithmetic intensity of the algorithms must be sufficiently high. Algorithms based on frequent matrix-vector products with assembled matrices are unlikely to achieve sufficient arithmetic intensity, and it is therefore necessary to consider matrix-free or Jacobian-free algorithms Kempf et al. (2020) ; Kronbichler and Kormann (2019) and preconditioners suitable for such solvers. The lack of efficient and grid independent solvers for DG discretizations of compressible time dependent flow problems limits the applicability of high order discretizations in many practical applications. Previous attempts are for example Bassi et al. (2009) ; Birken et al. (2013) ; Carr et al. (2016) ; Franco et al. (2023) ; Wang et al. (2022) ; Puppo et al. (2024) . Robust and efficient preconditioning techniques for Jacobian-free Newton-Krylov (JFNK) solvers are therefore an active area of research. Multigrid methods have been shown to scale optimally in many settings, in particular for elliptic problems, but also for finite volume discretizations of compressible flows Caughey and Jameson (2001) ; Birken et al. (2019) ; Birken (2021) . For the multigrid method to be Jacobian-free, the smoother must be Jacobian-free, restricting the possible choices. Recently, a new Jacobian-free multigrid preconditioner for use in JFNK methods was suggested in Birken et al. (2019) and then extended in Kasimir, J. and Versbach, L. M. and Birken, P. and Gassner, G. J. and Klöfkorn, R. (2021) ; Birken et al. (2024) . The approach uses a geometric multigrid method defined on a sub cell finite volume discretization (also known as low-order-refined (LOR) preconditioning). It is therefore similar to p -multigrid, but allows to make use of existing smoothers for FV methods, such as Birken (2012) ; Birken et al. (2013) . This motivates us to build on that foundation. We follow the framework from Blom et al. (2016) , which makes use of high order time integration, and a JFNK solver with a smart choice of tolerances. We extend the preconditioner named above to the discretized 2D viscous compressible flow equations with a source term for gravity, as is common for atmospheric flows. Thereby, we make use of the technique from Schuster et al. (2014) to obtain a well balanced scheme. The paper is organized as follows. In Sections 2.1 and 2.2 we briefly recall the main building blocks of the DG discretization and implicit time stepping. In Sections 2.3 and 2.4.3 we introduce the proposed solver and preconditioner for DG discretizations. We then investigate in Section 3 the effectiveness of this approach for various atmospheric test cases. 2 Governing equations, Discretization and Solvers We consider a general class of time dependent nonlinear advection-diffusion-reaction problems for a vector valued function \mathbf{U}\colon(0,T)\times\Omega\to\mathbb{R}^{r} with r\in{\mathbb{N}}^{+} components of the form \displaystyle\partial_{t}\mathbf{U}=\mathcal{L}(\mathbf{U}) \displaystyle:= \displaystyle-\nabla\cdot\big(F_{\rm c}(\mathbf{U})+F_{\rm v}(\mathbf{U},\nabla\mathbf{U})\big)\ \ \mbox{ in }(0,T]\times\Omega (1) in \Omega\subset\mathbb{R}^{d} , d=1,2,3 . Suitable initial and boundary conditions have to be added. F_{\rm c} describes the convective flux, F_{\rm v} the viscous flux, and S a source term. Note that all the coefficients in the partial differential equation are allowed to depend explicitly on the spatial variable x and on time t but to simplify the presentation we suppress this dependency in our notation. For the discretization, we use a method of lines approach based on first discretizing the differential operator in space using a DG approximation, and then

Read the complete text