Towards Automated Algebraic Multigrid Preconditioner Design Using Genetic Programming for Large-Scale Laser Beam Welding Simulations Conference: ; ; CCS: Computing methodologies Genetic programming CCS: Computing methodologies Massively parallel and high-performance simulations CCS: Mathematics of c
Dinesh Parthasarathy, Tommaso Bevilacqua, Martin Lanser, Axel Klawonn, Harald Köstler · 2024
Read Towards Automated Algebraic Multigrid Preconditioner Design Using Genetic Programming for Large-Scale Laser Beam Welding Simulations Conference: ; ; CCS: Computing methodologies Genetic programming CCS: Computing methodologies Massively parallel and high-performance simulations CCS: Mathematics of c by Dinesh Parthasarathy, Tommaso Bevilacqua, Martin Lanser, Axel Klawonn, Harald Köstler (2024) 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-2412-08186.txt.
Introduction
Designing multigrid solvers by evolution is the idea from Dinesh Parthasarathy and co-authors in a 2024 paper. Multigrid methods can be near optimal for large simulations, but they need many choices, and those choices strongly affect speed. Instead of learning parts with machine learning, the team uses genetic programming to build good solver cycles from known parts.
The test case is laser beam welding, modeled with coupled thermo-elastic equations over time. The system is nonlinear and hard to solve, so Newton's method is used with an algebraic multigrid preconditioner from hypre, run through PETSc. Flexible cycles are added to lift solver speed further. Readers see how automated search can take over a tuning job that once took expert trial and error.
Catalog record
| CREATOR | Dinesh Parthasarathy, Tommaso Bevilacqua, Martin Lanser, Axel Klawonn, Harald Köstler |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-08186
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Towards Automated Algebraic Multigrid Preconditioner Design Using Genetic Programming for Large-Scale Laser Beam Welding Simulations Conference: ; ; CCS: Computing methodologies Genetic programming CCS: Computing methodologies Massively parallel and high-performance simulations CCS: Mathematics of computing Solvers Dinesh Parthasarathy, Tommaso Bevilacqua, Martin Lanser, Axel Klawonn, Harald Köstler Abstract . Multigrid methods are asymptotically optimal algorithms ideal for large-scale simulations. But, they require making numerous algorithmic choices that significantly influence their efficiency. Unlike recent approaches that learn optimal multigrid components using machine learning techniques, we adopt a complementary strategy here, employing evolutionary algorithms to construct efficient multigrid cycles from available individual components. This technology is applied to finite element simulations of the laser beam welding process. The thermo-elastic behavior is described by a coupled system of time-dependent thermo-elasticity equations, leading to nonlinear and ill-conditioned systems. The nonlinearity is addressed using Newton’s method, and iterative solvers are accelerated with an AMG ( AMG ) preconditioner using hypre BoomerAMG interfaced via PETSc. This is applied as a monolithic solver for the coupled equations. To further enhance solver efficiency, flexible AMG cycles are introduced, extending traditional cycle types with level-specific smoothing sequences and non-recursive cycling patterns. These are automatically generated using genetic programming, guided by a context-free grammar containing AMG rules. Numerical experiments demonstrate the potential of these approaches to improve solver performance in large-scale laser beam welding simulations. 1. Introduction Laser beam welding is a non-contact method for joining materials that has gained significant importance as industrial production becomes more automated. Its appeal lies in the short cycle times and minimal heat-affected zones it offers. However, the rapid cooling rates characteristic for this process can result in a residual melt oversaturated with specific alloy elements, potentially causing solidification cracks. Gaining a quantitative understanding of how these fractures develop and their relationship to process parameters is crucial for refining and enhancing welding processes. This specific focus is central to the efforts of the DFG research group 5134 ‘‘Solidification Cracks During Laser Beam Welding -- High Performance Computing for High Performance Processing’’ 3 3 3 https://www.for5134.science/en/ . Another central part of the research group is to use the potential of modern supercomputers and high performance simulations to get further insights into the process of laser beam welding. A crucial part thereof is the optimization of parallel efficient iterative solvers for large thermo-mechanical finite element simulations. Especially, the latter aspect is the focus of this article. Here, the thermo-mechanical behavior of laser beam welding is described by a coupled set of time-dependent thermo-elasticity equations, as outlined in ( Simo and Miehe 1992 ) . Due to the complexity of the problem, the linear systems resulting from finite element discretization in space, backward Euler discretization in time, and linearization with Newton’s method tend to be ill-conditioned. Besides some work on developing and implementing parallel monolithic domain decomposition preconditioners for thermo-mechanical finite element problems ( Bevilacqua et al. 2024 ) , we are also interested in finding efficient AMG ( AMG ) preconditioners for such problems. As a result, in this work, we explore an automated approach for the design of AMG methods. More specifically, the scope of this work is to find optimal configurations for hypre BoomerAMG ( Henson and Yang 2002 ; Falgout et al. 2006 ) as a preconditioner and accelerate the convergence of the GMRES (Generalized Minimal Residual) method; in our numerical simulation we use our software package FE2TI ( Klawonn et al. 2020 ) based on PETSc ( Balay et al. 2019 ) with an interface to FEAP ( Taylor 2014 ) . Here, we apply BoomerAMG directly to the coupled problems arising from the linearization. Unlike approaches that exploit special structures, such as – defining block-triangular preconditioners and applying AMG on separate blocks – the goal here is to find an optimal BoomerAMG configuration that delivers the best performance when used as a monolithic preconditioner for coupled thermo-elasticity problems. Multigrid methods have several algorithmic components, such as, smoothers, cycle types, intergrid operators, and coarsening schemes, all of which determine their efficiency. These choices are not only complex but also problem-dependent. Leveraging recent advances in artificial intelligence (AI), efforts have been made to find optimal multigrid components, such as neural smoothers ( Huang et al. 2023 ; Zhang et al. 2024 ) , intergrid operators ( Greenfeld et al. 2019 ; Luz et al. 2020 ; Katrutsa et al. 2020 ; Katrutsa et al. 2017 ) , and coarsening schemes ( Taghibakhshi et al. 2021 ) . But unlike previous efforts, we take a complementary approach towards constructing efficient multigrid cycles from a set of available multigrid components. Traditional multigrid methods employ recursive cycle types such as V-, W-, and F-cycles. However, we consider an extended search space of multigrid cycles—referred to as flexible cycles—with level-specific smoothing sequences and arbitrary non-recursive cycling patterns across the grid hierarchy. The aim is to exploit this flexibility and design more efficient multigrid methods than the standard types. Due to the prohibitively large search space, manually designing a flexible cycle is impractical, so it is formulated as a program synthesis task. Previous work by Schmitt et al. generated such flexible geometric multigrid programs using grammar-guided genetic programming (G3P) ( Schmitt et al. 2021 ; Schmitt and Köstler 2022 ) . This approach has been extended to generate efficient flexible algebraic multigrid (AMG) methods ( Parthasarathy et al. 2024 ) . We intend to integrate the latter into the laser beam welding simulation software for automated AMG preconditioner design. As a preliminary step, this paper explores the potential benefits of such flexible AMG preconditioners, generated offline, decoupled from the simulation, using G3P. The remainder of the paper is organized as follows: In Section 2 we present the thermo-elasticity equations and the problem of laser beam welding, in Section 3 we introduce the discretization in space and time, in Section 4 we discuss key concepts of GP ( GP ), in Section 5 we describe our approach towards automated AMG design, and finally in Section 6 we show some numerical results. Figure 1 . Triangulated surface representing the geometry of the melting pool ( Bakir et al. 2018 ) and an example of a discrete representation of it using hexahedral elements. 2. Laser Beam welding problem We start presenting the general framework of our problem. To focus primarily on evaluating the performance of the preconditioners, we streamline the model problem by simplifying the laser beam’s action. Consequently, we use experimental data to determine the geometry of the melting pool, that is, the region where the material is melted by the laser. Within this region, we impose the melting temperature of the material across all degrees of freedom (DOFs) by applying Dirichlet boundary conditions. It is important to note that both the laser beam and the associated melting pool geometry, including the relevant Dirichlet boundaries, change position over time, moving with a problem-specific speed. The interested reader can find a more realistic laser heat source model which has been introduced to achieve more realistic simulations in ( Hartwig et al. 2024 ; Hartwig et al. 2023a ; Hartwig et al. 2023b ) . For context, Fig. 3 offers a general outline of a typical model geometry undergoing laser beam welding, while