Numerical Solution of Mixed-Dimensional PDEs Using a Neural Preconditioner

Nunzio Dimola, Nicola Rares Franco, Paolo Zunino · 2025

Read Numerical Solution of Mixed-Dimensional PDEs Using a Neural Preconditioner by Nunzio Dimola, Nicola Rares Franco, Paolo Zunino (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-2505-08491.txt.

Introduction

Nunzio Dimola, Nicola Rares Franco, and Paolo Zunino speed up hard equation solving with a neural aid in a 2025 paper. Their equations mix domains of distinct sizes, like bulk rock plus thin cracks. Such mixed problems are slow and prone to poor conditioning. Each new crack shape can force a fresh, costly run.

The team trains a neural preconditioner that helps solvers converge fast. It learns from low-detail problems and works on finer grids. It also adapts to new shapes of the thin parts without retraining. Tests show faster solving than classic methods in set cases. The authors note limits but see wide use for coupled physics problems.

Catalog record

CREATORNunzio Dimola, Nicola Rares Franco, Paolo Zunino
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE97 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-08491

Browse the full library · About the Tower of Babel

Full text

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

Numerical Solution of Mixed-Dimensional PDEs Using a Neural Preconditioner Nunzio Dimola, Nicola Rares Franco, Paolo Zunino Abstract Mixed-dimensional partial differential equations (PDEs) are characterized by coupled operators defined on domains of varying dimensions and pose significant computational challenges due to their inherent ill-conditioning. Moreover, the computational workload increases considerably when attempting to accurately capture the behavior of the system under significant variations or uncertainties in the low-dimensional structures such as fractures, fibers, or vascular networks, due to the inevitable necessity of running multiple simulations. In this work, we present a novel preconditioning strategy that leverages neural networks and unsupervised operator learning to design an efficient preconditioner specifically tailored to a class of 3D-1D mixed-dimensional PDEs. The proposed approach is capable of generalizing to varying shapes of the 1D manifold without retraining, making it robust to changes in the 1D graph topology. Moreover, thanks to convolutional neural networks, the neural preconditioner can adapt over a range of increasing mesh resolutions of the discrete problem, enabling us to train it on low resolution problems and deploy it on higher resolutions. Numerical experiments validate the effectiveness of the preconditioner in accelerating convergence in iterative solvers, demonstrating its appeal and limitations over traditional methods. This study lays the groundwork for applying neural network-based preconditioning techniques to a broader range of coupled multi-physics systems. Keywords mixed-dimensional PDEs, finite element approximation, scientific machine learning, preconditioning 1 Introduction Research on mixed-dimensional partial differential equations (PDEs) is advancing to address challenges in complex multiscale applications in fields such as geophysics, biomechanics, and neuroscience. These disciplines often involve thin structures with high aspect ratio, including structural mechanics and geomechanics [ 30 , 24 , 14 ] , microcirculation and perfusion modeling [ 40 , 34 ] , axons in neural networks [ 7 ] , and environmental flows [ 18 ] . Modeling these interactions with traditional 3D meshes would require massive computational resources, making direct simulations impractical. Mixed-dimensional PDEs, which enable lower-dimensional representations to be embedded within higher-dimensional domains, present a promising alternative. However, the computational effort associated with these models rises significantly when trying to capture the behavior of the system for varying configurations or interactions within the low-dimensional structures, the latter representing, e.g., fracture networks in geological formations, fiber distributions in structural materials, or vascular and neural networks in biological systems. Similarly, multiple simulations are needed when the structure of the low-dimensional problem is uncertain or unknown. In all such cases, despite the simplifications provided by mixed-dimensional PDEs, the overall computational demand remains high due to the cumulative cost of repeated individual simulations. This growing need underscores the importance of developing efficient solvers and preconditioners capable of handling the repeated solutions essential to representing the statistical variability of intricate networks in real-world applications. Today, the discretization of mixed-dimensional PDEs is established, providing computationally efficient schemes suitable for applications with intricate networks [ 5 , 16 , 22 ] , with rigorous error estimates ensuring accuracy. The well-established literature on block and operator preconditioning techniques, see, e.g., [ 28 ] , has been instrumental in the development of computational solvers for such problems. [ 21 , 23 ] have paved the way for the efficient resolution of discrete mixed-dimensional problems. Along this direction, we also mention the work of [ 8 ] on algebraic multigrid methods for metric-perturbed coupled problems, which has significantly advanced the understanding of block preconditioning techniques in mixed-dimensional PDEs. Parallel to this, other researchers have also explored approximate algorithms that can enhance scalability for large systems: see, e.g. [ 12 ] . The goal of this study is to develop a novel preconditioning strategy that addresses the twofold challenges inherent in mixed-dimensional PDEs: the ill-conditioned behavior of these problems and the need for adaptability across diverse geometric configurations of the low-dimensional subproblems. We aim to create a preconditioner that not only mitigates the ill-conditioning of the system, but also maintains this improved conditioning across a variety of configurations for 1D or other low-dimensional structures. To accomplish this, the preconditioner is designed as a nonlinear operator rather than a conventional matrix, enabling it to process the entire solution manifold effectively. Our method takes advantage of the latest advances in the approximation capabilities of neural networks and learning approaches related to machine learning [ 20 , 6 , 26 ] . These advancements have sparked active research in the creation of novel solvers and preconditioners tailored for different types of problem [ 15 , 19 , 17 , 29 , 2 ] , this being a non-exhaustive list of examples supporting the potential of neural network-based preconditioners as a viable alternative for solving large, sparse linear systems across various scientific and engineering domains. Building upon these studies, we develop a nonlinear preconditioner designed to efficiently manage the numerical challenges associated with the complex structure of mixed-dimensional equations. By incorporating a shape descriptor into the learning framework, the preconditioner can generalize across different configurations of the 1D graph or low-dimensional structures without the need for retraining. In order to demonstrate the efficacy of the proposed approach, we focus our attention on a mixed-dimensional 3D-1D model problem, which serves as a representative template for a broader class of coupled systems with similar characteristics. The results show that the learned preconditioner significantly accelerates the convergence of the iterative Generalized Minimum Residual (GMRES) solver, even as the complexity of the problem increases due to changes in the 1D graph’s topology or the size of the 3D domain. Unlike traditional preconditioning techniques, which typically require problem-specific tuning or reconfiguration when the underlying system changes, our method maintains consistent performance across diverse problem settings and mesh resolutions . Moreover, the proposed preconditioning approach can be easily generalized, making it a versatile and robust tool for a wide array of mixed-dimensional problems beyond the specific examples addressed in this study. This manuscript is organized as follows. First, we introduce the mathematical formulation of the 3D-1D mixed-dimensional PDE and highlight the challenges associated with solving such systems. Then, we present the proposed preconditioner learning framework, discussing both the theoretical foundations and the practical implementation of the nonlinear preconditioner. Next, we apply the preconditioner to the 3D-1D problem, illustrating its construction and providing a detailed analysis of its efficiency and scalability. Finally, we validate the approach through extensive numerical experiments, comparing its performance against conventional preconditioning methods, and conclude with a discussion of the implications and potential future directions for extending this work. 2 A mixed-dimensional problem We consider here a mixed-dimensional elliptic problem as a simple template for a wider class of coupled problems with similar characteristics. It will be our reference for comparisons, analyses, and considerations on preconditioner performances. Originally, we define the problem in a three-dimensional (3D) slender domain embedded in an external one, where we set partial differential equations coupled by suitable interface conditions. The mixed-dimensional model is obtained by a dimensional reduction strategy that represents the slender interior domain as a one-dimensional (1D) manifold, namely a metric graph. More precisely, we address a coupled problem defined by an exterior domain \Omega\subset\mathbb{R}^{3} and an interior domain \Sigma\subset\Omega . We assume \Sigma to be a generalized cylinder with a centerline \Lambda , the latter being a 1D domain with arc length

Read the complete text