A Domain Decomposition Deep Neural Network Method with Multi-Activation Functions for Solving Elliptic and Parabolic Interface Problems

Qijia Zhai · 2025

Read A Domain Decomposition Deep Neural Network Method with Multi-Activation Functions for Solving Elliptic and Parabolic Interface Problems by Qijia Zhai (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-2512-18178.txt.

Introduction

Qijia Zhai of Sichuan University develops a deep learning method for interface problems, equations whose coefficients jump sharply across a boundary between two regions. His 2025 paper covers elliptic and parabolic interface problems in two to ten dimensions. The approach splits the domain and trains two independent neural networks, one per subdomain. The networks are joined through interface conditions placed inside the loss function.

The key device is a multi activation mechanism inside each network. It blends several activation functions, such as tanh and Gaussian type functions, with weights that respond to the interface. That blending helps learning near the boundary, where the coupling demands are hardest. Zhai proves conditional error bounds linking solution accuracy to the trained loss and quadrature errors. Numerical tests across interface geometries from two to ten dimensions confirm the method works accurately. Such solvers matter wherever materials meet, from heat flow to electromagnetics.

Catalog record

CREATORQijia Zhai
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE117 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-18178

Browse the full library · About the Tower of Babel

Full text

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

A Domain Decomposition Deep Neural Network Method with Multi-Activation Functions for Solving Elliptic and Parabolic Interface Problems Qijia Zhai Abstract We present a domain decomposition-based deep learning method for solving elliptic and parabolic interface problems with discontinuous coefficients in two to ten dimensions. Our Multi-Activation Function (MAF) approach employs two independent neural networks, one for each subdomain, coupled through interface conditions in the loss function. The key innovation is a multi-activation mechanism within each subdomain network that adaptively blends multiple activation functions (e.g., tanh \tanh and Gaussian-type) with interface-aware weighting, enhancing learning efficiency near interfaces where coupling constraints are most demanding. We prove conditional error bounds relating solution accuracy to trained loss values and quadrature errors. Numerical experiments on elliptic and parabolic interface problems with various interface geometries (2D–10D) validate the effectiveness and accuracy of the proposed method. 1 Introduction In recent years, great attention has been directed towards a new category of computational techniques known as deep neural networks (DNNs), especially physics-informed neural networks (PINNs) [ 40 , 13 ] . These methods utilize feed-forward neural networks to approximate unknown field variables by minimizing loss functions derived from the residuals of governing field equations, initial and boundary conditions, as well as observed data. The loss functions are evaluated at randomly selected collocation points distributed throughout the computational domain where the field variables are to be determined. By incorporating the governing equations directly into the loss functions, PINNs effectively overcome a major limitation of traditional data-driven modeling approaches: the requirement for large volumes of high-quality training data, which is often difficult and costly to obtain. Since their development, PINNs have found application across a wide range of engineering problems, including solid mechanics [ 23 , 24 , 4 ] , fluid mechanics [ 5 , 36 , 51 ] , flow in porous media [ 2 , 19 , 53 ] , and heat transfer [ 6 ] . Elliptic interface problems form a class of partial differential equations (PDEs) that are distinguished by the presence of discontinuities in both the solution and its derivatives across interfaces. These problems are particularly common in a wide range of scientific and engineering fields, such as fluid dynamics [ 9 ] , material science [ 47 ] , electromagnetics [ 11 ] , and bio-mechanics [ 15 ] , where different physical properties are present in adjacent subdomains. The interfaces between these subdomains often correspond to sharp changes in material properties, leading to complex mathematical formulations that are challenging to solve accurately and efficiently. In addition to elliptic interface problems, parabolic interface problems also arise in numerous real-world applications, for instance in the modeling of heat conduction in composite materials [ 1 ] or diffusion-driven phenomena with discontinuities [ 20 ] . Like their elliptic counterparts, parabolic interface problems involve sharp changes in material parameters across interfaces; however, these problems include an additional time dimension. This temporal component further complicates the numerical treatment, as one must accurately capture both the spatial discontinuities and the evolving solution behavior over time. Handling these coupled spatial-temporal complexities often demands refined meshes or special numerical techniques that can adapt to both the moving solution fronts and the interface geometry, significantly increasing computational cost and implementation challenges [ 8 , 3 ] . Traditional mesh-based methods, such as the finite element method (FEM) [ 12 , 10 , 38 , 22 , 46 , 26 , 8 , 34 , 54 ] , finite difference method (FDM) [ 33 , 28 , 18 , 17 , 7 , 3 ] , and finite volume method (FVM) [ 39 , 35 , 44 , 41 ] , have been widely used to solve elliptic interface problems. These methods typically rely on mesh generation, where the domain is discretized into smaller subdomains (elements or grids). However, accurately capturing the behavior of solutions near interfaces, especially when the geometry of the interface is irregular or high-dimensional, poses significant difficulties. In particular, these conventional methods require fine meshes around the interfaces to resolve the discontinuities, dramatically increasing the computational cost. Moreover, mesh generation can be highly non-trivial, particularly in complex geometries, often leading to issues related to stability and convergence. Unlike mesh-based methods, deep learning approaches, especially PINNs, can operate in a mesh-free manner, which significantly alleviates the difficulties associated with mesh generation and refinement [ 13 , 40 ] . Additionally, neural networks possess inherent flexibility, allowing them to approximate complex solution spaces, even in high-dimensional domains. Their ability to learn from data also makes them well-suited for problems where analytical solutions are difficult to obtain or where experimental data is available for training. Furthermore, the parallelization capabilities of deep learning models, coupled with advances in computational hardware, such as GPUs and TPUs, contribute to their efficiency in solving large-scale and high-dimensional problems. The application of deep learning to elliptic interface problems, however, is still in its early stages, and several challenges remain to be addressed [ 13 ] . These include the design of neural network architectures that can effectively capture discontinuities, the development of efficient training strategies that ensure convergence, and the integration of physical constraints into the learning process to ensure that the solutions remain physically meaningful. Nevertheless, there have been many representative works that have successfully used deep learning methods to solve elliptic interface problems [ 50 , 43 , 42 , 25 , 27 , 45 , 30 , 48 , 32 , 31 , 49 , 16 , 52 ] . We now review several representative neural network-based methods for solving PDEs with interfaces or domain decomposition. Extended Physics-Informed Neural Networks (XPINN) [ 29 ] is a domain decomposition framework developed primarily for parallel computing and accelerating computations, rather than specifically for interface problems with discontinuous coefficients. While XPINNs divide the computational domain into subdomains handled by separate neural networks with continuity conditions at boundaries, they do not explicitly address discontinuities in solution gradients or material coefficients. Several specialized methods have been developed for interface problems with discontinuous coefficients. Multi-domain PINNs (M-PINN) [ 50 ] employ separate networks for each subdomain with standard activation functions, coupled through interface loss terms. Interface PINNs (I-PINN) [ 43 ] incorporate interface conditions directly into both network architecture and loss formulation. Adaptive Interface-PINNs (AdaI-PINN) [ 42 ] further enhance I-PINNs with adaptive sampling strategies near the interface. Mesh-free methods (MFM) [ 25 ] use scattered points with radial basis functions, offering flexibility for irregular domains but with limited application to parabolic problems. Discontinuity Capturing Shallow Neural Networks (DCSNN) [ 27 , 45 ] transform problems to higher-dimensional spaces to capture jump discontinuities with shallow architectures, though they focus primarily on elliptic problems. In this paper, we propose a Multi-Activation Function (MAF) method for solving elliptic and parabolic interface problems. Our method employs two independent neural networks, one for each subdomain \Omega_{1} and \Omega_{2} , coupled through interface conditions in the loss function. The key innovation lies in the multi-activation mechanism within each subdomain network: instead of using a single fixed activation function, we adaptively combine multiple activation functions (e.g., \tanh and Gaussian-type) with interface-aware weighting. This design allows each network to efficiently capture sharp variations required to satisfy interface

Read the complete text