Graph ODEs and Beyond: A Comprehensive Survey on Integrating Differential Equations with Graph Neural Networks
Zewen Liu, Xiaoda Wang, Bohan Wang, Zijie Huang, Carl Yang, Wei Jin · 2025
Read Graph ODEs and Beyond: A Comprehensive Survey on Integrating Differential Equations with Graph Neural Networks by Zewen Liu, Xiaoda Wang, Bohan Wang, Zijie Huang, Carl Yang, Wei Jin (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-2503-23167.txt.
Introduction
Zewen Liu and co-authors survey the marriage of graph neural networks and differential equations in a 2025 paper. Networks learn on graph data; equations describe how things change continuously in time and space. Together they promise models that respect both structure and dynamics.
The survey sorts methods by principle and shows where each fits. Uses span molecules, traffic forecasting, and epidemic spread, all systems that are networks in motion. Open challenges and future directions close the review. For researchers entering this fast-growing borderland, the paper serves as both map and compass.
Catalog record
| CREATOR | Zewen Liu, Xiaoda Wang, Bohan Wang, Zijie Huang, Carl Yang, Wei Jin |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-23167
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Graph ODEs and Beyond: A Comprehensive Survey on Integrating Differential Equations with Graph Neural Networks Zewen Liu, Xiaoda Wang, Bohan Wang, Zijie Huang, Carl Yang, Wei Jin Abstract . Graph Neural Networks (GNNs) and differential equations (DEs) are two rapidly advancing areas of research that have shown remarkable synergy in recent years. GNNs have emerged as powerful tools for learning on graph-structured data, while differential equations provide a principled framework for modeling continuous dynamics across time and space. The intersection of these fields has led to innovative approaches that leverage the strengths of both, enabling applications in physics-informed learning, spatiotemporal modeling, and scientific computing. This survey aims to provide a comprehensive overview of the burgeoning research at the intersection of GNNs and DEs. We will categorize existing methods, discuss their underlying principles, and highlight their applications across domains such as molecular modeling, traffic prediction, and epidemic spreading. Furthermore, we identify open challenges and outline future research directions to advance this interdisciplinary field. A comprehensive paper list is provided at https://github.com/Emory-Melody/Awesome-Graph-NDEs . This survey serves as a resource for researchers and practitioners seeking to understand and contribute to the fusion of GNNs and DEs. Conference: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2; August 3–7, 2025; Toronto, ON, Canada Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’25), August 3–7, 2025, Toronto, ON, Canada DOI: 10.1145/3711896.3736559 ISBN: 979-8-4007-1454-2/2025/08 † † cc-license: by 1. Introduction Understanding and predicting complex behaviors in natural and engineered systems is a fundamental challenge across scientific and industrial domains. Many real-world phenomena exhibit dynamic evolution over time, governed by intricate interdependencies between variables. Examples include climate patterns shaped by atmospheric and oceanic interactions ( 1 ) , population dynamics influenced by birth and migration rates ( 2 ) , financial markets driven by investor behavior and economic indicators ( 3 ) , disease progression driven by biological factors ( 4 , 5 ) , and the spread of infectious diseases determined by transmission dynamics and intervention strategies ( 6 , 7 ) . Capturing these temporal changes and underlying mechanisms requires mathematical models that not only describe system behavior but also provide predictive insights. To effectively model dynamical systems, Differential Equations (DEs), such as Ordinary Differential Equations (ODEs) ( 8 ) , Partial Differential Equations (PDEs) ( 9 ) , and Stochastic Differential Equations (SDEs) ( 10 ) , relate one or more unknown functions to their derivatives, thus describing how outputs vary given changing variables. At their core, DEs consist of three essential components: (1) state variables that describe the system’s condition, (2) derivatives that model and capture the rate of change, and (3) parameters that influence the dynamics under given initial and boundary conditions. These elements work together to provide a structured approach to understanding how systems evolve over time. Despite the crucial role of DEs in modeling complex phenomena, various challenges arise from real-world applications. Notably, many systems exhibit intricate, high-dimensional dynamics that are difficult to capture using purely knowledge-driven DE formulations ( 11 ) , as deriving accurate governing equations often requires human expert involvement. Moreover, computational efficiency remains a major obstacle, especially for high-dimensional and nonlinear PDEs, since traditional numerical solvers must manage an enormous number of equations corresponding to the system’s graph structure, often rendering these approaches prohibitively expensive ( 12 , 13 , 14 ) . In response to these challenges, neural differential equations (NDEs), such as Neural ODEs ( 15 ) , have emerged as a data-driven alternative that learns the underlying dynamics directly from data, bypassing the need for explicit formulation of governing rules. This innovative approach enables the modeling of systems for which traditional equations may be intractable or unknown. Nevertheless, while NDEs excel at capturing temporal evolution, it remains challenging to model spatial dynamics, such as epidemic spread in social networks ( 16 ) or transportation flows in urban networks ( 17 ) , where discrete interactions complicate continuous-state representations. This limitation urgently calls for methods that can effectively integrate temporal dynamics with spatial context. To handle the above issues, recent research has leveraged Graph Neural Networks (GNNs) ( 18 , 19 , 20 , 21 ) , powerful tools for learning relational data, to build graph-based NDEs and model the complex interactions between variables. Early explorations integrate the graph learning capabilities of GNNs within the continuous-time framework of NDEs and propose Graph neural ODEs ( 22 , 23 , 24 ) , which offer a versatile and powerful approach to modeling complex systems that evolve over both space and time. This integration not only enables the capture of dynamic temporal behavior but also leverages the rich spatial relationships encoded in graph structures. Beyond Graph Neural ODEs, the broader class of Graph Neural Differential Equations (Graph NDEs), including Graph Neural PDEs ( 25 ) and Graph Neural SDEs ( 26 ) , bridges the gap between NDEs and GNNs. Contributions. In this work, we aim to present a comprehensive and latest review of methods that combine graph neural networks with differential equations, addressing the gap by summarizing key tasks, methodologies, and applications in this evolving field. Our contributions can be summarized as follows: (a) We offer the first comprehensive review of Graph NDEs that model continuous spatial and temporal dynamics. (b) We introduce a structured taxonomy of Graph NDEs in Section 3 and conduct an in-depth review of research integrating GNNs with different classes of differential equations, including ODEs, PDEs, and SDEs, as detailed in Section 4 . (c) We explore the diverse applications of Graph NDEs in Section 5 , highlighting their impact across various real-world scenarios. (d) We identify emerging trends, key challenges, and promising future research directions in Section 6 , aiming to inspire further exploration in this interdisciplinary field. Connections to existing surveys. While previous surveys have explored Graph NDEs, they often lack comprehensiveness in methodology and categorization, limiting their ability to fully bridge GNNs and NDEs. Many focus on specific applications of neural differential equations ( 27 , 28 , 29 ) , overlooking spatial dynamics. Others examine the integration of GNNs with differential equations ( 30 , 31 ) but remain narrow in scope regarding DE types and categorization. In contrast, our survey compiles a broad range of recent studies, offering a detailed review of methodologies, challenges, and applications. Additionally, we present a well-structured taxonomy as well as valuable insights for future research. 2. Background 2.1. Learning on Graphs In this paper, we define a graph as \mathcal{G}=(\mathcal{V},\mathcal{E}) , where |\mathcal{V}|=N represents the number of nodes, and \mathcal{E}\subseteq\mathcal{V}\times\mathcal{V} represents the set of edges connecting nodes. The features of all nodes is represented as \mathbf{X}=\{\mathbf{x}_{1},\mathbf{x}_{2},...\mathbf{x}_{N}\}\in\mathbb{R}^{N\times D} , where D denotes the feature dimension. The adjacency matrix of \mathcal{G} is denoted as \mathbf{A} , where \mathbf{A}_{ij}=1 if the edge e_{ij}\in\mathcal{E} and \mathbf{A}_{ij}=0 if e_{ij}\notin\mathcal{E} . GNNs provide a flexible framework to learn graph representations. A common paradigm is message passing, where each node v updates its representation \mathbf{h}_{v} based on aggregating messages from its neighbors \mathcal{N}(v) . A GNN with L layers can be described as: (1) \mathbf{h}_{v}^{(l+1)}=f_{\phi}\Bigg(\mathbf{h}_{v}^{(l)},\bigoplus_{u\in\mathcal{N}(v)}f_{\theta}\bigl(\mathbf{h}_{v}^{(l)},\mathbf{h}_{u}^{(l)},\mathbf{e}_{uv}\bigr)\Bigg),\forall l\in[L], where f_{\theta} and f_{\phi} are learnable functions parameterized by \theta